{"id":89305,"date":"2026-08-21T13:50:05","date_gmt":"2026-08-21T10:50:05","guid":{"rendered":"https:\/\/twodots.gr\/?p=89305"},"modified":"2026-08-21T13:50:11","modified_gmt":"2026-08-21T10:50:11","slug":"moe-modela-stin-akri-cache-bandwidth-bottleneck","status":"publish","type":"post","link":"https:\/\/twodots.gr\/bg\/moe-modela-stin-akri-cache-bandwidth-bottleneck\/","title":{"rendered":"MoE \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1 \u03c3\u03c4\u03b7\u03bd \u03ac\u03ba\u03c1\u03b7: \u03b3\u03b9\u03b1\u03c4\u03af \u03b7 cache \u03b4\u03b5\u03bd \u03bb\u03cd\u03bd\u03b5\u03b9 \u03bc\u03cc\u03bd\u03b7 \u03c4\u03b7\u03c2 \u03c4\u03bf bandwidth bottleneck"},"content":{"rendered":"<div class=\"td-article-lede\">\n<p><strong>\u03a4\u03b1 MoE \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1 \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03bd\u03b1 \u00ab\u03c7\u03c9\u03c1\u03ad\u03c3\u03bf\u03c5\u03bd\u00bb \u03c3\u03b5 consumer hardware \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03c0\u03c1\u03bf\u03c3\u03c6\u03ad\u03c1\u03bf\u03c5\u03bd \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf interactive \u03c1\u03c5\u03b8\u03bc\u03cc, \u03b5\u03c0\u03b5\u03b9\u03b4\u03ae \u03b7 \u03b1\u03c1\u03b1\u03b9\u03ae \u03b5\u03bd\u03b5\u03c1\u03b3\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 \u03bc\u03b5\u03b9\u03ce\u03bd\u03b5\u03b9 \u03c4\u03bf\u03bd \u03c5\u03c0\u03bf\u03bb\u03bf\u03b3\u03b9\u03c3\u03bc\u03cc \u03b1\u03bb\u03bb\u03ac \u03b4\u03b5\u03bd \u03b5\u03be\u03b1\u03c6\u03b1\u03bd\u03af\u03b6\u03b5\u03b9 \u03c4\u03b7 \u03bc\u03b5\u03c4\u03b1\u03c6\u03bf\u03c1\u03ac \u03b2\u03b1\u03c1\u03ce\u03bd.<\/strong> \u038c\u03c4\u03b1\u03bd \u03bf\u03b9 \u03b5\u03bd\u03b5\u03c1\u03b3\u03bf\u03af experts \u03b2\u03c1\u03af\u03c3\u03ba\u03bf\u03bd\u03c4\u03b1\u03b9 \u03ba\u03c5\u03c1\u03af\u03c9\u03c2 \u03c3\u03b5 SSD, \u03c4\u03bf decode \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03c4\u03b1 bytes \u03c0\u03bf\u03c5 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b4\u03b9\u03b1\u03b2\u03b1\u03c3\u03c4\u03bf\u03cd\u03bd \u03b3\u03b9\u03b1 \u03ba\u03ac\u03b8\u03b5 token. \u0397 cache \u03b2\u03bf\u03b7\u03b8\u03ac \u03bf\u03c5\u03c3\u03b9\u03b1\u03c3\u03c4\u03b9\u03ba\u03ac, \u03b1\u03bb\u03bb\u03ac \u03c3\u03c4\u03bf \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03b4\u03b5\u03bd \u03b1\u03c1\u03ba\u03b5\u03af \u03bc\u03cc\u03bd\u03b7 \u03c4\u03b7\u03c2 \u03b3\u03b9\u03b1 \u03bd\u03b1 \u03b3\u03b5\u03c6\u03c5\u03c1\u03ce\u03c3\u03b5\u03b9 \u03c4\u03bf \u03ba\u03b5\u03bd\u03cc \u03ad\u03c9\u03c2 \u03c4\u03b1 10 token \u03b1\u03bd\u03ac \u03b4\u03b5\u03c5\u03c4\u03b5\u03c1\u03cc\u03bb\u03b5\u03c0\u03c4\u03bf.<\/p>\n<p>\u0397 \u03c0\u03c1\u03bf\u03b5\u03b3\u03b3\u03b5\u03b3\u03c1\u03b1\u03bc\u03bc\u03ad\u03bd\u03b7 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7 <em>Cacheable by Design?<\/em> \u03c4\u03bf\u03c5 Shriniwas Ramesh Suram \u03c3\u03c5\u03bd\u03b4\u03c5\u03ac\u03b6\u03b5\u03b9 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ad\u03c2 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b5\u03b9\u03c2 \u03c3\u03b5 Qwen3, traces \u03c4\u03bf\u03c5 router, \u03c0\u03c1\u03bf\u03c3\u03bf\u03bc\u03bf\u03b9\u03ce\u03c3\u03b5\u03b9\u03c2 cache \u03ba\u03b1\u03b9 \u03b5\u03ba\u03c0\u03b1\u03af\u03b4\u03b5\u03c5\u03c3\u03b7 \u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03c9\u03bd MoE \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03c9\u03bd. \u03a4\u03bf \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf \u03c3\u03c5\u03bc\u03c0\u03ad\u03c1\u03b1\u03c3\u03bc\u03b1 \u03b3\u03b9\u03b1 \u03bc\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03af\u03c1\u03b7\u03c3\u03b7 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03cc\u03c4\u03b9 \u03c4\u03bf edge AI \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03b4\u03cd\u03bd\u03b1\u03c4\u03bf. \u0395\u03af\u03bd\u03b1\u03b9 \u03cc\u03c4\u03b9 model size, \u03b5\u03bd\u03b5\u03c1\u03b3\u03ad\u03c2 \u03c0\u03b1\u03c1\u03ac\u03bc\u03b5\u03c4\u03c1\u03bf\u03b9, memory tiers, routing profile, latency \u03ba\u03b1\u03b9 \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03bf\u03cd\u03bd\u03c4\u03b1\u03b9 \u03c9\u03c2 \u03ad\u03bd\u03b1 \u03b5\u03bd\u03b9\u03b1\u03af\u03bf \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1.<\/p>\n<\/div>\n<div class=\"td-article-toc\">\n<div class=\"td-toc-title\">\u03a0\u03b5\u03c1\u03b9\u03b5\u03c7\u03cc\u03bc\u03b5\u03bd\u03b1<\/div>\n<ul>\n<li><a href=\"#moe-sparsity-bandwidth-wall\">\u03a0\u03ce\u03c2 \u03b7 \u03b1\u03c1\u03b1\u03b9\u03ae \u03b5\u03bd\u03b5\u03c1\u03b3\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 \u03c3\u03c5\u03bd\u03b1\u03bd\u03c4\u03ac \u03c4\u03bf bandwidth wall<\/a><\/li>\n<li><a href=\"#peirama-consumer-hardware\">\u03a4\u03b9 \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b5 \u03c4\u03bf \u03c0\u03b5\u03af\u03c1\u03b1\u03bc\u03b1 \u03c3\u03b5 consumer hardware<\/a><\/li>\n<li><a href=\"#batching-paging-thrash\">\u0393\u03b9\u03b1\u03c4\u03af \u03c4\u03bf batching \u03ba\u03b1\u03c4\u03ad\u03c1\u03c1\u03b5\u03c5\u03c3\u03b5 \u03b1\u03bd\u03c4\u03af \u03bd\u03b1 \u03b5\u03c0\u03b9\u03c4\u03b1\u03c7\u03cd\u03bd\u03b5\u03b9<\/a><\/li>\n<li><a href=\"#llama-moe-trace-routing\">\u03a4\u03b9 \u03b1\u03c0\u03bf\u03ba\u03ac\u03bb\u03c5\u03c8\u03b5 \u03c4\u03bf llama-moe-trace \u03b3\u03b9\u03b1 \u03c4\u03bf routing<\/a><\/li>\n<li><a href=\"#lru-cache-path-mapped-serving\">\u03a0\u03cc\u03c3\u03bf \u03b2\u03bf\u03b7\u03b8\u03bf\u03cd\u03bd \u03b7 LRU cache \u03ba\u03b1\u03b9 \u03c4\u03bf Path-Mapped Serving<\/a><\/li>\n<li><a href=\"#cache-aware-training-quality\">\u0393\u03b9\u03b1\u03c4\u03af \u03b7 cache-aware \u03b5\u03ba\u03c0\u03b1\u03af\u03b4\u03b5\u03c5\u03c3\u03b7 \u03b4\u03b5\u03bd \u03c0\u03ad\u03c1\u03b1\u03c3\u03b5 \u03c4\u03bf \u03cc\u03c1\u03b9\u03bf \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2<\/a><\/li>\n<li><a href=\"#training-rerouting-syndyasmos\">\u039f \u03c3\u03c5\u03bd\u03b4\u03c5\u03b1\u03c3\u03bc\u03cc\u03c2 training \u03ba\u03b1\u03b9 rerouting<\/a><\/li>\n<li><a href=\"#epicheiriseis-edge-ai\">\u03a4\u03b9 \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 \u03b3\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03ae\u03c3\u03b5\u03b9\u03c2 \u03ba\u03b1\u03b9 \u03bf\u03bc\u03ac\u03b4\u03b5\u03c2 AI<\/a><\/li>\n<li><a href=\"#elegchoi-pilot-moe\">\u0395\u03c0\u03c4\u03ac \u03ad\u03bb\u03b5\u03b3\u03c7\u03bf\u03b9 \u03c0\u03c1\u03b9\u03bd \u03b1\u03c0\u03cc pilot MoE \u03c3\u03c4\u03b7\u03bd \u03ac\u03ba\u03c1\u03b7<\/a><\/li>\n<li><a href=\"#periorismoi-meletis\">\u03a0\u03b5\u03c1\u03b9\u03bf\u03c1\u03b9\u03c3\u03bc\u03bf\u03af \u03c0\u03bf\u03c5 \u03b4\u03b5\u03bd \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c7\u03b1\u03b8\u03bf\u03cd\u03bd<\/a><\/li>\n<li><a href=\"#cache-meros-systimatos\">\u0397 cache \u03b5\u03af\u03bd\u03b1\u03b9 \u03bc\u03ad\u03c1\u03bf\u03c2 \u03c4\u03b7\u03c2 \u03bb\u03cd\u03c3\u03b7\u03c2, \u03cc\u03c7\u03b9 \u03cc\u03bb\u03b7 \u03b7 \u03bb\u03cd\u03c3\u03b7<\/a><\/li>\n<\/ul>\n<\/div>\n<h2 id=\"moe-sparsity-bandwidth-wall\">\u03a0\u03ce\u03c2 \u03b7 \u03b1\u03c1\u03b1\u03b9\u03ae \u03b5\u03bd\u03b5\u03c1\u03b3\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 \u03c3\u03c5\u03bd\u03b1\u03bd\u03c4\u03ac \u03c4\u03bf bandwidth wall<\/h2>\n<p>\u03a3\u03b5 \u03ad\u03bd\u03b1 Mixture-of-Experts \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf, \u03ba\u03ac\u03b8\u03b5 layer \u03c0\u03b5\u03c1\u03b9\u03ad\u03c7\u03b5\u03b9 \u03c0\u03bf\u03bb\u03bb\u03bf\u03cd\u03c2 \u03b5\u03be\u03b5\u03b9\u03b4\u03b9\u03ba\u03b5\u03c5\u03bc\u03ad\u03bd\u03bf\u03c5\u03c2 feed-forward experts \u03ba\u03b1\u03b9 \u03ad\u03bd\u03b1\u03c2 router \u03b5\u03c0\u03b9\u03bb\u03ad\u03b3\u03b5\u03b9 \u03bc\u03b9\u03ba\u03c1\u03cc \u03c5\u03c0\u03bf\u03c3\u03cd\u03bd\u03bf\u03bb\u03bf \u03b3\u03b9\u03b1 \u03ba\u03ac\u03b8\u03b5 token. \u0397 \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03ae \u03c7\u03c9\u03c1\u03b7\u03c4\u03b9\u03ba\u03cc\u03c4\u03b7\u03c4\u03b1 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03ad\u03c4\u03c3\u03b9 \u03bd\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03bf\u03bb\u03cd \u03bc\u03b5\u03b3\u03ac\u03bb\u03b7, \u03b5\u03bd\u03ce \u03c4\u03bf \u03b5\u03bd\u03b5\u03c1\u03b3\u03cc \u03c5\u03c0\u03bf\u03bb\u03bf\u03b3\u03b9\u03c3\u03c4\u03b9\u03ba\u03cc \u03c6\u03bf\u03c1\u03c4\u03af\u03bf \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf. \u03a4\u03bf Qwen3-235B-A22B, \u03b3\u03b9\u03b1 \u03c0\u03b1\u03c1\u03ac\u03b4\u03b5\u03b9\u03b3\u03bc\u03b1, \u03b1\u03c0\u03bf\u03b8\u03b7\u03ba\u03b5\u03cd\u03b5\u03b9 235 \u03b4\u03b9\u03c3. \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03c4\u03c1\u03bf\u03c5\u03c2 \u03ba\u03b1\u03b9 \u03b5\u03bd\u03b5\u03c1\u03b3\u03bf\u03c0\u03bf\u03b9\u03b5\u03af \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 22 \u03b4\u03b9\u03c3. \u03b1\u03bd\u03ac token.<\/p>\n<p>\u0391\u03c5\u03c4\u03cc \u03c4\u03bf sparsity \u03ba\u03ac\u03bd\u03b5\u03b9 \u03ad\u03bd\u03b1 \u03bc\u03b5\u03b3\u03ac\u03bb\u03bf open-weight \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03b8\u03b5\u03c9\u03c1\u03b7\u03c4\u03b9\u03ba\u03ac \u03b5\u03ba\u03c4\u03b5\u03bb\u03ad\u03c3\u03b9\u03bc\u03bf \u03ad\u03be\u03c9 \u03b1\u03c0\u03cc data center, \u03b1\u03bb\u03bb\u03ac \u03b4\u03b5\u03bd \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03cc\u03bb\u03b1 \u03c4\u03b1 \u03b2\u03ac\u03c1\u03b7 \u03c7\u03c9\u03c1\u03bf\u03cd\u03bd \u03c3\u03c4\u03b7 \u03b3\u03c1\u03ae\u03b3\u03bf\u03c1\u03b7 \u03bc\u03bd\u03ae\u03bc\u03b7. \u0391\u03bd \u03bf\u03b9 experts \u03c0\u03bf\u03c5 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c4\u03bf \u03b5\u03c0\u03cc\u03bc\u03b5\u03bd\u03bf token \u03b2\u03c1\u03af\u03c3\u03ba\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c3\u03b5 RAM \u03ae SSD, \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03bc\u03b5\u03c4\u03b1\u03c6\u03b5\u03c1\u03b8\u03bf\u03cd\u03bd \u03c0\u03c1\u03b9\u03bd \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b7\u03b8\u03bf\u03cd\u03bd. \u03a3\u03c4\u03bf autoregressive decode \u03b7 \u03c0\u03c1\u03ac\u03be\u03b7 \u03b5\u03c0\u03b1\u03bd\u03b1\u03bb\u03b1\u03bc\u03b2\u03ac\u03bd\u03b5\u03c4\u03b1\u03b9 token \u03c0\u03c1\u03bf\u03c2 token, \u03bf\u03c0\u03cc\u03c4\u03b5 \u03b7 \u03b1\u03c0\u03cc\u03b4\u03bf\u03c3\u03b7 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03ba\u03b1\u03b8\u03bf\u03c1\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03bf \u03b1\u03c0\u03cc \u03c4\u03bf memory bandwidth \u03c0\u03b1\u03c1\u03ac \u03b1\u03c0\u03cc \u03c4\u03b1 \u03b4\u03b9\u03b1\u03b8\u03ad\u03c3\u03b9\u03bc\u03b1 FLOPs.<\/p>\n<p>\u0397 \u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03b7 \u03b4\u03b9\u03ac\u03ba\u03c1\u03b9\u03c3\u03b7 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03bd\u03ac\u03bc\u03b5\u03c3\u03b1 \u03c3\u03c4\u03bf <em>model fit<\/em> \u03ba\u03b1\u03b9 \u03c3\u03c4\u03bf <em>usable throughput<\/em>. \u03a4\u03bf \u03c0\u03c1\u03ce\u03c4\u03bf \u03b1\u03c0\u03b1\u03bd\u03c4\u03ac \u03b1\u03bd \u03c4\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03c6\u03bf\u03c1\u03c4\u03ce\u03bd\u03b5\u03c4\u03b1\u03b9 \u03ba\u03b1\u03b9 \u03c0\u03b1\u03c1\u03ac\u03b3\u03b5\u03b9 \u03ad\u03be\u03bf\u03b4\u03bf. \u03a4\u03bf \u03b4\u03b5\u03cd\u03c4\u03b5\u03c1\u03bf \u03b1\u03bd \u03b7 \u03ad\u03be\u03bf\u03b4\u03bf\u03c2 \u03c6\u03c4\u03ac\u03bd\u03b5\u03b9 \u03b1\u03c1\u03ba\u03b5\u03c4\u03ac \u03b3\u03c1\u03ae\u03b3\u03bf\u03c1\u03b1 \u03b3\u03b9\u03b1 chatbot, copilot, \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae \u03c0\u03b5\u03c1\u03b9\u03b5\u03c7\u03bf\u03bc\u03ad\u03bd\u03bf\u03c5 \u03ae \u03ac\u03bb\u03bb\u03b7 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ae \u03c1\u03bf\u03ae. \u0397 \u03af\u03b4\u03b9\u03b1 \u03bb\u03bf\u03b3\u03b9\u03ba\u03ae \u03b5\u03bc\u03c6\u03b1\u03bd\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03ba\u03b1\u03b9 \u03c3\u03c4\u03b7\u03bd <a href=\"https:\/\/twodots.gr\/moonep-exisorropisi-ai-experts-moe\/\">\u03b5\u03be\u03b9\u03c3\u03bf\u03c1\u03c1\u03cc\u03c0\u03b7\u03c3\u03b7 experts \u03c3\u03c4\u03b1 MoE \u03c3\u03c5\u03c3\u03c4\u03ae\u03bc\u03b1\u03c4\u03b1<\/a>: \u03c4\u03bf routing, \u03b7 \u03bc\u03bd\u03ae\u03bc\u03b7 \u03ba\u03b1\u03b9 \u03c4\u03bf serving engine \u03b5\u03af\u03bd\u03b1\u03b9 \u03b5\u03be\u03af\u03c3\u03bf\u03c5 \u03c3\u03b7\u03bc\u03b1\u03bd\u03c4\u03b9\u03ba\u03ac \u03bc\u03b5 \u03c4\u03bf \u03cc\u03bd\u03bf\u03bc\u03b1 \u03c4\u03bf\u03c5 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf\u03c5.<\/p>\n<h2 id=\"peirama-consumer-hardware\">\u03a4\u03b9 \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b5 \u03c4\u03bf \u03c0\u03b5\u03af\u03c1\u03b1\u03bc\u03b1 \u03c3\u03b5 consumer hardware<\/h2>\n<p>\u0397 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7 \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b5 desktop \u03bc\u03b5 Intel i9-12900, RTX 3070 8 GB, 32 GB DDR4 RAM \u03ba\u03b1\u03b9 WD SN530 NVMe. \u03a4\u03bf Qwen3-235B-A22B \u03c3\u03b5 Q4_K_M GGUF \u03ba\u03b1\u03c4\u03b1\u03bb\u03ac\u03bc\u03b2\u03b1\u03bd\u03b5 134 GB. \u03a0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 40 GB \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03c3\u03b1\u03bd \u03bd\u03b1 \u03c4\u03bf\u03c0\u03bf\u03b8\u03b5\u03c4\u03b7\u03b8\u03bf\u03cd\u03bd \u03c3\u03c4\u03b1 \u03b3\u03c1\u03b7\u03b3\u03bf\u03c1\u03cc\u03c4\u03b5\u03c1\u03b1 \u03b5\u03c0\u03af\u03c0\u03b5\u03b4\u03b1 VRAM \u03ba\u03b1\u03b9 RAM, \u03b5\u03bd\u03ce \u03c4\u03bf \u03c5\u03c0\u03cc\u03bb\u03bf\u03b9\u03c0\u03bf \u03ad\u03bc\u03b5\u03bd\u03b5 \u03c3\u03c4\u03bf\u03bd SSD.<\/p>\n<p>\u039a\u03ac\u03b8\u03b5 token \u03b5\u03bd\u03b5\u03c1\u03b3\u03bf\u03c0\u03bf\u03b9\u03bf\u03cd\u03c3\u03b5 \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 22 \u03b4\u03b9\u03c3. \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03c4\u03c1\u03bf\u03c5\u03c2, \u03c0\u03bf\u03c5 \u03b1\u03bd\u03c4\u03b9\u03c3\u03c4\u03bf\u03b9\u03c7\u03bf\u03cd\u03c3\u03b1\u03bd \u03c3\u03b5 \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 12 GB \u03b1\u03bd\u03b1\u03b3\u03bd\u03ce\u03c3\u03b5\u03c9\u03bd \u03c4\u03b5\u03c4\u03c1\u03b1\u03bc\u03c0\u03b9\u03c4\u03c9\u03bd expert weights. \u03a4\u03bf warm single-stream decode \u03bc\u03b5\u03c4\u03c1\u03ae\u03b8\u03b7\u03ba\u03b5 \u03c3\u03c4\u03b1 0,441 token \u03b1\u03bd\u03ac \u03b4\u03b5\u03c5\u03c4\u03b5\u03c1\u03cc\u03bb\u03b5\u03c0\u03c4\u03bf. \u039f \u03c1\u03c5\u03b8\u03bc\u03cc\u03c2 \u03c3\u03c5\u03bc\u03c6\u03c9\u03bd\u03bf\u03cd\u03c3\u03b5 \u03bc\u03b5 \u03c4\u03bf \u03b4\u03b9\u03b1\u03b8\u03ad\u03c3\u03b9\u03bc\u03bf I\/O \u03ba\u03b1\u03b9 \u03bc\u03b5 \u03c4\u03bf \u03b1\u03c0\u03bb\u03cc \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf <em>throughput = bandwidth \/ bytes \u03b1\u03bd\u03ac token<\/em>.<\/p>\n<div class=\"td-chart td-chart--metrics\">\n<div class=\"td-chart-head\">\n<p class=\"td-chart-title\">\u03a4\u03bf \u03bc\u03b5\u03c4\u03c1\u03b7\u03bc\u03ad\u03bd\u03bf bandwidth wall \u03c4\u03bf\u03c5 \u03c0\u03b5\u03b9\u03c1\u03ac\u03bc\u03b1\u03c4\u03bf\u03c2<\/p>\n<p class=\"td-chart-subtitle\">\u039f\u03b9 \u03c4\u03b9\u03bc\u03ad\u03c2 \u03b1\u03c6\u03bf\u03c1\u03bf\u03cd\u03bd \u03c4\u03bf \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf desktop, \u03c4\u03bf Qwen3-235B-A22B Q4_K_M \u03ba\u03b1\u03b9 \u03c4\u03b7 \u03b4\u03b9\u03b1\u03bc\u03cc\u03c1\u03c6\u03c9\u03c3\u03b7 \u03c4\u03b7\u03c2 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7\u03c2\u00b7 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03b3\u03b5\u03bd\u03b9\u03ba\u03cc benchmark \u03ba\u03ac\u03b8\u03b5 MoE deployment.<\/p>\n<\/div>\n<div class=\"td-chart-body\">\n<div class=\"td-metric-grid\">\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">448 GB\/s<\/span><span class=\"td-metric-label\">VRAM bandwidth<\/span><span class=\"td-metric-note\">\u03a4\u03bf \u03c4\u03b1\u03c7\u03cd\u03c4\u03b5\u03c1\u03bf tier \u03c4\u03c9\u03bd 8 GB<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">50 GB\/s<\/span><span class=\"td-metric-label\">RAM bandwidth<\/span><span class=\"td-metric-note\">\u03a4\u03bf \u03b5\u03bd\u03b4\u03b9\u03ac\u03bc\u03b5\u03c3\u03bf tier \u03c4\u03c9\u03bd 32 GB<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">2,4 GB\/s<\/span><span class=\"td-metric-label\">NVMe bandwidth<\/span><span class=\"td-metric-note\">\u03a4\u03bf tier \u03cc\u03c0\u03bf\u03c5 \u03ad\u03bc\u03b5\u03bd\u03b5 \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 \u03c4\u03bf 70% \u03c4\u03bf\u03c5 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf\u03c5<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">0,441 tok\/s<\/span><span class=\"td-metric-label\">Warm decode<\/span><span class=\"td-metric-note\">Single stream \u03c3\u03c4\u03bf \u03bc\u03b5\u03c4\u03c1\u03b7\u03bc\u03ad\u03bd\u03bf setup<\/span><\/div>\n<\/div>\n<\/div>\n<\/div>\n<p>\u0393\u03b9\u03b1 \u03c3\u03c4\u03cc\u03c7\u03bf 10 token \u03b1\u03bd\u03ac \u03b4\u03b5\u03c5\u03c4\u03b5\u03c1\u03cc\u03bb\u03b5\u03c0\u03c4\u03bf \u03bc\u03b5 12 GB expert reads \u03b1\u03bd\u03ac token, \u03bf \u03c5\u03c0\u03bf\u03bb\u03bf\u03b3\u03b9\u03c3\u03bc\u03cc\u03c2 \u03b1\u03c0\u03b1\u03b9\u03c4\u03b5\u03af \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 120 GB\/s \u03c3\u03c5\u03bd\u03b5\u03c7\u03bf\u03cd\u03c2 \u03bc\u03b5\u03c4\u03b1\u03c6\u03bf\u03c1\u03ac\u03c2 \u03c0\u03c1\u03bf\u03c2 \u03c4\u03b1 \u03b2\u03ac\u03c1\u03b7. \u039f SSD \u03ad\u03b4\u03b9\u03bd\u03b5 2,4 GB\/s: \u03ad\u03bd\u03b1 \u03ba\u03b5\u03bd\u03cc 50\u00d7. \u0397 \u03b5\u03ba\u03c4\u03af\u03bc\u03b7\u03c3\u03b7 \u03b4\u03b5\u03bd \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03ba\u03bd\u03cd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03ba\u03ac\u03b8\u03b5 \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03b8\u03b1 \u03ad\u03c7\u03b5\u03b9 \u03af\u03b4\u03b9\u03b1 \u03c4\u03b1\u03c7\u03cd\u03c4\u03b7\u03c4\u03b1\u00b7 \u03b4\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03b3\u03b9\u03b1\u03c4\u03af \u03bc\u03b9\u03b1 \u03bc\u03b9\u03ba\u03c1\u03ae \u03c1\u03cd\u03b8\u03bc\u03b9\u03c3\u03b7 \u03bb\u03bf\u03b3\u03b9\u03c3\u03bc\u03b9\u03ba\u03bf\u03cd \u03b4\u03b5\u03bd \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b1\u03ba\u03c5\u03c1\u03ce\u03c3\u03b5\u03b9 \u03c4\u03bf \u03c6\u03c5\u03c3\u03b9\u03ba\u03cc \u03cc\u03c1\u03b9\u03bf \u03b1\u03c5\u03c4\u03bf\u03cd \u03c4\u03bf\u03c5 hardware.<\/p>\n<h2 id=\"batching-paging-thrash\">\u0393\u03b9\u03b1\u03c4\u03af \u03c4\u03bf batching \u03ba\u03b1\u03c4\u03ad\u03c1\u03c1\u03b5\u03c5\u03c3\u03b5 \u03b1\u03bd\u03c4\u03af \u03bd\u03b1 \u03b5\u03c0\u03b9\u03c4\u03b1\u03c7\u03cd\u03bd\u03b5\u03b9<\/h2>\n<p>\u03a4\u03bf batching \u03b8\u03b1 \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03c3\u03b5 \u03b8\u03b5\u03c9\u03c1\u03b7\u03c4\u03b9\u03ba\u03ac \u03bd\u03b1 \u03bc\u03bf\u03b9\u03c1\u03ac\u03c3\u03b5\u03b9 \u03bc\u03af\u03b1 \u03b1\u03bd\u03ac\u03b3\u03bd\u03c9\u03c3\u03b7 experts \u03c3\u03b5 \u03c0\u03bf\u03bb\u03bb\u03ad\u03c2 \u03c0\u03b1\u03c1\u03ac\u03bb\u03bb\u03b7\u03bb\u03b5\u03c2 \u03c1\u03bf\u03ad\u03c2. \u03a4\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03c0\u03c1\u03cc\u03b2\u03bb\u03b5\u03c8\u03b7\u03c2 \u03c4\u03b7\u03c2 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7\u03c2 \u03c4\u03b1\u03af\u03c1\u03b9\u03b1\u03be\u03b5 \u03bc\u03b5 \u03c4\u03b9\u03c2 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b5\u03b9\u03c2 \u03b3\u03b9\u03b1 batch 1 \u03ba\u03b1\u03b9 8, \u03cc\u03c0\u03bf\u03c5 \u03b7 aggregate \u03b1\u03c0\u03cc\u03b4\u03bf\u03c3\u03b7 \u03ad\u03c6\u03c4\u03b1\u03c3\u03b5 \u03b1\u03bd\u03c4\u03af\u03c3\u03c4\u03bf\u03b9\u03c7\u03b1 0,128 \u03ba\u03b1\u03b9 0,189 token \u03b1\u03bd\u03ac \u03b4\u03b5\u03c5\u03c4\u03b5\u03c1\u03cc\u03bb\u03b5\u03c0\u03c4\u03bf. \u03a3\u03c4\u03bf batch 32, \u03cc\u03bc\u03c9\u03c2, \u03b1\u03bd\u03c4\u03af \u03b3\u03b9\u03b1 \u03c4\u03b7\u03bd \u03c0\u03c1\u03bf\u03b2\u03bb\u03b5\u03c0\u03cc\u03bc\u03b5\u03bd\u03b7 \u03ac\u03bd\u03bf\u03b4\u03bf \u03c3\u03c4\u03b1 0,35 token \u03b1\u03bd\u03ac \u03b4\u03b5\u03c5\u03c4\u03b5\u03c1\u03cc\u03bb\u03b5\u03c0\u03c4\u03bf, \u03b7 \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b7 \u03ad\u03c0\u03b5\u03c3\u03b5 \u03c3\u03c4\u03b1 0,087.<\/p>\n<p>\u0397 \u03b1\u03b9\u03c4\u03af\u03b1 \u03b1\u03c0\u03bf\u03b4\u03cc\u03b8\u03b7\u03ba\u03b5 \u03c3\u03b5 paging thrash. \u039f\u03b9 32 \u03c4\u03b1\u03c5\u03c4\u03cc\u03c7\u03c1\u03bf\u03bd\u03b5\u03c2 \u03c1\u03bf\u03ad\u03c2 \u03b4\u03b7\u03bc\u03b9\u03bf\u03cd\u03c1\u03b3\u03b7\u03c3\u03b1\u03bd \u03ba\u03b1\u03c4\u03b1\u03b9\u03b3\u03b9\u03c3\u03bc\u03cc page faults \u03ba\u03b1\u03b9 \u03c4\u03bf \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b9\u03ba\u03cc \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03ad\u03c7\u03b1\u03c3\u03b5 \u03c4\u03b7\u03bd \u03b1\u03c0\u03bf\u03b4\u03bf\u03c4\u03b9\u03ba\u03ae \u03c3\u03b5\u03b9\u03c1\u03b9\u03b1\u03ba\u03ae \u03c0\u03c1\u03cc\u03c3\u03b2\u03b1\u03c3\u03b7 \u03c3\u03c4\u03bf\u03bd \u03b4\u03af\u03c3\u03ba\u03bf. \u03a4\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03b4\u03b5\u03bd \u03b1\u03ba\u03c5\u03c1\u03ce\u03bd\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b9\u03b4\u03ad\u03b1 \u03b5\u03bd\u03cc\u03c2 shared disk sweep. \u0394\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03c4\u03bf \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf llama.cpp setup \u03b4\u03b5\u03bd \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03c3\u03b5 \u03bd\u03b1 \u03c4\u03b7\u03bd \u03b5\u03ba\u03c6\u03c1\u03ac\u03c3\u03b5\u03b9 \u03b1\u03c0\u03bf\u03b4\u03bf\u03c4\u03b9\u03ba\u03ac \u03c0\u03ac\u03bd\u03c9 \u03b1\u03c0\u03cc \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 batch 8 \u03ba\u03b1\u03b9 \u03cc\u03c4\u03b9 \u03ad\u03bd\u03b1\u03c2 purpose-built sequential-sweep engine \u03b8\u03b1 \u03ae\u03c4\u03b1\u03bd \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03cc \u03c0\u03b5\u03af\u03c1\u03b1\u03bc\u03b1.<\/p>\n<div class=\"td-comparison\">\n<p class=\"td-comparison-title\">\u03a4\u03ad\u03c3\u03c3\u03b5\u03c1\u03b9\u03c2 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ad\u03c2 \u03ad\u03bd\u03bd\u03bf\u03b9\u03b5\u03c2 \u03c0\u03bf\u03c5 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03bc\u03b5\u03c4\u03c1\u03ce\u03bd\u03c4\u03b1\u03b9 \u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03ac<\/p>\n<div class=\"td-comparison-cards td-comparison-cards--horizontal\">\n<div class=\"td-comparison-grid td-comparison-grid--four\">\n<div class=\"td-platform-card\">\n<h3>Model fit<\/h3>\n<p>\u0391\u03bd \u03c4\u03bf checkpoint, \u03c4\u03b1 experts \u03ba\u03b1\u03b9 \u03c4\u03bf context \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03bd\u03b1 \u03c6\u03bf\u03c1\u03c4\u03c9\u03b8\u03bf\u03cd\u03bd \u03c3\u03c4\u03b1 \u03b4\u03b9\u03b1\u03b8\u03ad\u03c3\u03b9\u03bc\u03b1 tiers \u03c7\u03c9\u03c1\u03af\u03c2 \u03b1\u03c0\u03bf\u03c4\u03c5\u03c7\u03af\u03b1.<\/p>\n<\/div>\n<div class=\"td-platform-card\">\n<h3>Single-stream latency<\/h3>\n<p>\u03a0\u03cc\u03c3\u03bf \u03c0\u03b5\u03c1\u03b9\u03bc\u03ad\u03bd\u03b5\u03b9 \u03ad\u03bd\u03b1\u03c2 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc\u03c2 \u03c7\u03c1\u03ae\u03c3\u03c4\u03b7\u03c2 \u03b1\u03bd\u03ac token \u03cc\u03c4\u03b1\u03bd \u03b7 \u03c1\u03bf\u03ae \u03b4\u03b5\u03bd \u03bc\u03bf\u03b9\u03c1\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c4\u03bf I\/O \u03bc\u03b5 \u03ac\u03bb\u03bb\u03b1 requests.<\/p>\n<\/div>\n<div class=\"td-platform-card\">\n<h3>Aggregate throughput<\/h3>\n<p>\u03a0\u03cc\u03c3\u03b1 token \u03b5\u03be\u03c5\u03c0\u03b7\u03c1\u03b5\u03c4\u03b5\u03af \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03ac \u03c4\u03bf \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03bc\u03b5 batching, \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03ba\u03c1\u03cd\u03b2\u03b5\u03b9 latency \u03ae paging collapse.<\/p>\n<\/div>\n<div class=\"td-platform-card td-platform-card--navy\">\n<h3>Quality at operating point<\/h3>\n<p>\u0391\u03bd caching, rerouting \u03ae quantization \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03bf\u03cd\u03bd \u03c4\u03b7\u03bd \u03b1\u03c0\u03bf\u03b4\u03b5\u03ba\u03c4\u03ae \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c3\u03c4\u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc workload.<\/p>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<h2 id=\"llama-moe-trace-routing\">\u03a4\u03b9 \u03b1\u03c0\u03bf\u03ba\u03ac\u03bb\u03c5\u03c8\u03b5 \u03c4\u03bf llama-moe-trace \u03b3\u03b9\u03b1 \u03c4\u03bf routing<\/h2>\n<p>\u0393\u03b9\u03b1 \u03bd\u03b1 \u03b5\u03c0\u03b9\u03bb\u03b5\u03b3\u03b5\u03af \u03c4\u03b9 \u03b8\u03b1 \u03bc\u03b5\u03af\u03bd\u03b5\u03b9 \u03c3\u03c4\u03b7\u03bd cache, \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c0\u03c1\u03ce\u03c4\u03b1 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc routing profile. \u039f \u03c3\u03c5\u03b3\u03b3\u03c1\u03b1\u03c6\u03ad\u03b1\u03c2 \u03c0\u03c1\u03cc\u03c3\u03b8\u03b5\u03c3\u03b5 \u03c3\u03c4\u03bf evaluation callback \u03c4\u03bf\u03c5 llama.cpp \u03c4\u03bf llama-moe-trace, \u03ad\u03bd\u03b1 \u03b5\u03c1\u03b3\u03b1\u03bb\u03b5\u03af\u03bf \u03c0\u03bf\u03c5 \u03ba\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c6\u03b5\u03b9 \u03c4\u03b9\u03c2 top-k \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ad\u03c2 \u03c4\u03bf\u03c5 router \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03c4\u03b1 \u03b2\u03ac\u03c1\u03b7. \u039a\u03b1\u03c4\u03ad\u03b3\u03c1\u03b1\u03c8\u03b5 Qwen3-30B-A3B \u03c3\u03b5 8.000 tokens \u03b3\u03b9\u03b1 \u03ba\u03b1\u03b8\u03b5\u03bc\u03af\u03b1 \u03b1\u03c0\u03cc \u03c4\u03ad\u03c3\u03c3\u03b5\u03c1\u03b9\u03c2 \u03ba\u03b1\u03c4\u03b7\u03b3\u03bf\u03c1\u03af\u03b5\u03c2: \u03c0\u03b5\u03b6\u03cc \u03bb\u03cc\u03b3\u03bf, \u03ba\u03ce\u03b4\u03b9\u03ba\u03b1, \u03bc\u03b1\u03b8\u03b7\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac \u03ba\u03b1\u03b9 \u03b9\u03b1\u03c4\u03c1\u03b9\u03ba\u03cc \u03ba\u03b5\u03af\u03bc\u03b5\u03bd\u03bf.<\/p>\n<p>\u0397 \u03c0\u03b9\u03b8\u03b1\u03bd\u03cc\u03c4\u03b7\u03c4\u03b1 \u03b5\u03c0\u03b1\u03bd\u03b1\u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7\u03c2 expert \u03c3\u03c4\u03bf \u03b5\u03c0\u03cc\u03bc\u03b5\u03bd\u03bf token \u03ae\u03c4\u03b1\u03bd 0,444, \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 \u03c4\u03c5\u03c7\u03b1\u03af\u03b1\u03c2 \u03c0\u03b9\u03b8\u03b1\u03bd\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2 0,223. \u03a4\u03bf 95% \u03c4\u03b7\u03c2 \u03ba\u03af\u03bd\u03b7\u03c3\u03b7\u03c2 \u03c0\u03b5\u03c1\u03bd\u03bf\u03cd\u03c3\u03b5 \u03b1\u03c0\u03cc \u03c4\u03bf 52,5% \u03c4\u03c9\u03bd experts. \u03a4\u03bf code workload \u03b5\u03af\u03c7\u03b5 \u03bc\u03b9\u03ba\u03c1\u03ae \u03b5\u03c0\u03b9\u03ba\u03ac\u03bb\u03c5\u03c8\u03b7 \u03bc\u03b5 \u03c4\u03b1 \u03ac\u03bb\u03bb\u03b1 domains, \u03b5\u03bd\u03ce general prose, \u03bc\u03b1\u03b8\u03b7\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac \u03ba\u03b1\u03b9 \u03b9\u03b1\u03c4\u03c1\u03b9\u03ba\u03cc \u03ba\u03b5\u03af\u03bc\u03b5\u03bd\u03bf \u03b5\u03bc\u03c6\u03ac\u03bd\u03b9\u03b6\u03b1\u03bd \u03bc\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b7 \u03bc\u03b5\u03c4\u03b1\u03be\u03cd \u03c4\u03bf\u03c5\u03c2 \u03bf\u03bc\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1. \u0386\u03c1\u03b1 \u03bc\u03af\u03b1 \u03b5\u03bd\u03b9\u03b1\u03af\u03b1 cache policy \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b1\u03c0\u03bf\u03b4\u03af\u03b4\u03b5\u03b9 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ac \u03cc\u03c4\u03b1\u03bd \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03c4\u03bf \u03bc\u03af\u03b3\u03bc\u03b1 \u03b5\u03c1\u03b3\u03b1\u03c3\u03b9\u03ce\u03bd.<\/p>\n<p>\u0397 \u03c4\u03b5\u03c7\u03bd\u03b9\u03ba\u03ae \u03b9\u03c3\u03c4\u03bf\u03c1\u03af\u03b1 \u03c4\u03bf\u03c5 tracer \u03b5\u03af\u03bd\u03b1\u03b9 \u03b5\u03c0\u03af\u03c3\u03b7\u03c2 \u03c3\u03b7\u03bc\u03b1\u03bd\u03c4\u03b9\u03ba\u03ae \u03b3\u03b9\u03b1 \u03c4\u03bf E-E-A-T \u03c4\u03b7\u03c2 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7\u03c2. \u039c\u03b9\u03b1 \u03c0\u03c1\u03ce\u03c4\u03b7 \u03c5\u03bb\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 \u03b1\u03bd\u03c4\u03ad\u03b3\u03c1\u03b1\u03c8\u03b5 \u03bb\u03b1\u03bd\u03b8\u03b1\u03c3\u03bc\u03ad\u03bd\u03b1 \u03ad\u03bd\u03b1 \u03bc\u03b7 \u03c3\u03c5\u03bd\u03b5\u03c7\u03cc\u03bc\u03b5\u03bd\u03bf tensor view \u03ba\u03b1\u03b9 \u03ad\u03b4\u03c9\u03c3\u03b5 \u03bf\u03bc\u03bf\u03b9\u03cc\u03bc\u03bf\u03c1\u03c6\u03b1 \u03b1\u03bb\u03bb\u03ac \u03ac\u03c7\u03c1\u03b7\u03c3\u03c4\u03b1 \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03b1. \u039f\u03b9 \u03b1\u03bd\u03c9\u03bc\u03b1\u03bb\u03af\u03b5\u03c2 \u03b5\u03bd\u03c4\u03bf\u03c0\u03af\u03c3\u03c4\u03b7\u03ba\u03b1\u03bd, \u03b7 \u03b1\u03bd\u03c4\u03b9\u03b3\u03c1\u03b1\u03c6\u03ae \u03b4\u03b9\u03bf\u03c1\u03b8\u03ce\u03b8\u03b7\u03ba\u03b5 \u03ce\u03c3\u03c4\u03b5 \u03bd\u03b1 \u03c3\u03ad\u03b2\u03b5\u03c4\u03b1\u03b9 \u03c4\u03b1 strides \u03ba\u03b1\u03b9 \u03c4\u03bf \u03b1\u03c1\u03c7\u03b9\u03ba\u03cc run \u03b1\u03c0\u03bf\u03c1\u03c1\u03af\u03c6\u03b8\u03b7\u03ba\u03b5 \u03c3\u03c4\u03bf \u03b4\u03b7\u03bc\u03cc\u03c3\u03b9\u03bf corrections record.<\/p>\n<div class=\"td-article-note\"><strong>\u039a\u03c1\u03af\u03c3\u03b9\u03bc\u03b7 \u03b4\u03b9\u03ac\u03ba\u03c1\u03b9\u03c3\u03b7:<\/strong> \u03c4\u03b1 routing traces \u03ba\u03b1\u03b9 \u03c4\u03b1 cache hit rates \u03c0\u03c1\u03bf\u03ad\u03c1\u03c7\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03c4\u03bf Qwen3-30B-A3B \u03ba\u03b1\u03b9 \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03bf\u03cd\u03bd\u03c4\u03b1\u03b9 \u03c9\u03c2 same-family proxy \u03b3\u03b9\u03b1 \u03c4\u03b7\u03bd \u03c0\u03c1\u03bf\u03b2\u03bf\u03bb\u03ae \u03c3\u03c4\u03bf \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03c4\u03c9\u03bd 235B. \u0397 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7 \u03b4\u03b7\u03bb\u03ce\u03bd\u03b5\u03b9 \u03c1\u03b7\u03c4\u03ac \u03cc\u03c4\u03b9 \u03af\u03b4\u03b9\u03bf trace \u03c0\u03ac\u03bd\u03c9 \u03c3\u03c4\u03bf Qwen3-235B GGUF \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03bc\u03b5\u03bb\u03bb\u03bf\u03bd\u03c4\u03b9\u03ba\u03ae \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1.<\/div>\n<h2 id=\"lru-cache-path-mapped-serving\">\u03a0\u03cc\u03c3\u03bf \u03b2\u03bf\u03b7\u03b8\u03bf\u03cd\u03bd \u03b7 LRU cache \u03ba\u03b1\u03b9 \u03c4\u03bf Path-Mapped Serving<\/h2>\n<p>\u03a3\u03c4\u03bf replay \u03c4\u03c9\u03bd traces, \u03b1\u03bd\u03ac-layer LRU cache \u03c0\u03bf\u03c5 \u03c7\u03c9\u03c1\u03bf\u03cd\u03c3\u03b5 \u03c4\u03bf 13,4% \u03c4\u03c9\u03bd experts \u03b5\u03be\u03c5\u03c0\u03b7\u03c1\u03b5\u03c4\u03bf\u03cd\u03c3\u03b5 \u03c4\u03bf 65,9% \u03c4\u03c9\u03bd expert requests. \u03a3\u03c4\u03bf \u03af\u03b4\u03b9\u03bf budget, \u03c4\u03bf static pinning \u03ad\u03c6\u03c4\u03b1\u03c3\u03b5 59,2%, \u03b7 LFU 60,1% \u03ba\u03b1\u03b9 \u03bf \u03bc\u03b7 \u03c5\u03bb\u03bf\u03c0\u03bf\u03b9\u03ae\u03c3\u03b9\u03bc\u03bf\u03c2 clairvoyant Belady oracle 79,1%. \u0397 LRU \u03ba\u03ad\u03c1\u03b4\u03b9\u03c3\u03b5 \u03b5\u03c0\u03b5\u03b9\u03b4\u03ae \u03b1\u03be\u03b9\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b5 \u03c4\u03b7 \u03c7\u03c1\u03bf\u03bd\u03b9\u03ba\u03ae \u03c4\u03bf\u03c0\u03b9\u03ba\u03cc\u03c4\u03b7\u03c4\u03b1 \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03bc\u03cc\u03bd\u03bf \u03c4\u03b7 \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03ae \u03b4\u03b7\u03bc\u03bf\u03c4\u03b9\u03ba\u03cc\u03c4\u03b7\u03c4\u03b1 \u03ba\u03ac\u03b8\u03b5 expert.<\/p>\n<p>\u0397 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b7 cache \u03b1\u03c6\u03bf\u03c1\u03ac expert weights \u03ba\u03b1\u03b9 \u03b4\u03b5\u03bd \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c3\u03c5\u03b3\u03c7\u03ad\u03b5\u03c4\u03b1\u03b9 \u03bc\u03b5 prompt caching \u03ae KV cache. \u03a4\u03bf prompt cache \u03b1\u03c0\u03bf\u03c6\u03b5\u03cd\u03b3\u03b5\u03b9 \u03b5\u03c0\u03b1\u03bd\u03ac\u03bb\u03b7\u03c8\u03b7 \u03c5\u03c0\u03bf\u03bb\u03bf\u03b3\u03b9\u03c3\u03bc\u03bf\u03cd \u03b3\u03b9\u03b1 \u03ba\u03bf\u03b9\u03bd\u03cc prefix, \u03b5\u03bd\u03ce \u03b7 KV cache \u03b1\u03c0\u03bf\u03b8\u03b7\u03ba\u03b5\u03cd\u03b5\u03b9 \u03b5\u03bd\u03b4\u03b9\u03ac\u03bc\u03b5\u03c3\u03b5\u03c2 \u03ba\u03b1\u03c4\u03b1\u03c3\u03c4\u03ac\u03c3\u03b5\u03b9\u03c2 attention \u03b3\u03b9\u03b1 \u03c4\u03bf \u03c5\u03c0\u03ac\u03c1\u03c7\u03bf\u03bd context. \u0395\u03b4\u03ce \u03c4\u03bf \u03c0\u03c1\u03cc\u03b2\u03bb\u03b7\u03bc\u03b1 \u03b2\u03c1\u03af\u03c3\u03ba\u03b5\u03c4\u03b1\u03b9 \u03c3\u03c4\u03b7\u03bd \u03b1\u03bd\u03ac\u03ba\u03c4\u03b7\u03c3\u03b7 \u03c4\u03c9\u03bd \u03b5\u03bd\u03b5\u03c1\u03b3\u03ce\u03bd feed-forward experts \u03b1\u03c0\u03cc \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ac memory tiers \u03c3\u03b5 \u03ba\u03ac\u03b8\u03b5 \u03bd\u03ad\u03bf token. \u039c\u03b9\u03b1 \u03b5\u03c6\u03b1\u03c1\u03bc\u03bf\u03b3\u03ae \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03b5\u03c0\u03bf\u03bc\u03ad\u03bd\u03c9\u03c2 \u03bd\u03b1 \u03ad\u03c7\u03b5\u03b9 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03b5\u03c3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc prefix cache \u03ba\u03b1\u03b9 \u03bd\u03b1 \u03b5\u03be\u03b1\u03ba\u03bf\u03bb\u03bf\u03c5\u03b8\u03b5\u03af \u03bd\u03b1 \u03c3\u03ba\u03bf\u03bd\u03c4\u03ac\u03c6\u03c4\u03b5\u03b9 \u03c3\u03c4\u03bf expert-weight bandwidth.<\/p>\n<p>\u03a4\u03bf \u03c0\u03c1\u03bf\u03c4\u03b5\u03b9\u03bd\u03cc\u03bc\u03b5\u03bd\u03bf Path-Mapped Serving \u03ba\u03c1\u03b1\u03c4\u03ac \u03c4\u03bf\u03c5\u03c2 \u03b8\u03b5\u03c1\u03bc\u03bf\u03cd\u03c2 experts \u03c3\u03c4\u03b7 \u03b3\u03c1\u03ae\u03b3\u03bf\u03c1\u03b7 \u03bc\u03bd\u03ae\u03bc\u03b7 \u03ba\u03b1\u03b9 \u03c6\u03bf\u03c1\u03c4\u03ce\u03bd\u03b5\u03b9 \u03b1\u03c0\u03cc \u03c4\u03bf\u03bd \u03b4\u03af\u03c3\u03ba\u03bf \u03bc\u03cc\u03bd\u03bf \u03c4\u03b1 misses, \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03c4\u03b1 weights. \u039c\u03b5 \u03c0\u03c1\u03bf\u03b2\u03bf\u03bb\u03ae \u03c3\u03c4\u03bf Qwen3-235B, 18 GB fast-memory budget \u03ba\u03b1\u03b9 \u03c4\u03bf \u03bc\u03b5\u03c4\u03c1\u03b7\u03bc\u03ad\u03bd\u03bf I\/O, \u03b7 miss traffic \u03b5\u03ba\u03c4\u03b9\u03bc\u03ae\u03b8\u03b7\u03ba\u03b5 \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 \u03c3\u03c4\u03b1 2,7 GB \u03b1\u03bd\u03ac token \u03ba\u03b1\u03b9 \u03b7 single-stream \u03b1\u03c0\u03cc\u03b4\u03bf\u03c3\u03b7 \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 \u03c3\u03c4\u03b1 0,9 token \u03b1\u03bd\u03ac \u03b4\u03b5\u03c5\u03c4\u03b5\u03c1\u03cc\u03bb\u03b5\u03c0\u03c4\u03bf. \u0397 \u03b2\u03b5\u03bb\u03c4\u03af\u03c9\u03c3\u03b7 \u03b4\u03cd\u03bf \u03ad\u03c9\u03c2 \u03b5\u03c0\u03c4\u03ac \u03c6\u03bf\u03c1\u03ad\u03c2 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03b7, \u03b1\u03bb\u03bb\u03ac \u03b5\u03be\u03b1\u03ba\u03bf\u03bb\u03bf\u03c5\u03b8\u03b5\u03af \u03bd\u03b1 \u03b1\u03c0\u03ad\u03c7\u03b5\u03b9 \u03b1\u03c0\u03cc interactive \u03c1\u03c5\u03b8\u03bc\u03cc.<\/p>\n<p>\u0393\u03b9\u03b1 10 token \u03b1\u03bd\u03ac \u03b4\u03b5\u03c5\u03c4\u03b5\u03c1\u03cc\u03bb\u03b5\u03c0\u03c4\u03bf, \u03b7 \u03c0\u03c1\u03bf\u03b2\u03bf\u03bb\u03ae \u03b1\u03c0\u03b1\u03b9\u03c4\u03b5\u03af \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 95% cache hit rate, \u03c0\u03bf\u03c5 \u03b1\u03bd\u03c4\u03b9\u03c3\u03c4\u03bf\u03b9\u03c7\u03b5\u03af \u03c3\u03b5 \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 70 GB resident experts \u2014 \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 \u03c4\u03ad\u03c3\u03c3\u03b5\u03c1\u03b9\u03c2 \u03c6\u03bf\u03c1\u03ad\u03c2 \u03c4\u03bf fast-memory budget \u03c4\u03bf\u03c5 \u03c0\u03b5\u03b9\u03c1\u03ac\u03bc\u03b1\u03c4\u03bf\u03c2. \u0395\u03b4\u03ce \u03c6\u03b1\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03b3\u03b9\u03b1\u03c4\u03af \u03b7 <a href=\"https:\/\/twodots.gr\/inkling-open-weight-ai-epicheiriseis\/\">\u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7 \u03b3\u03b9\u03b1 self-hosted open-weight AI<\/a> \u03b4\u03b5\u03bd \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03be\u03b5\u03ba\u03b9\u03bd\u03ac \u03b1\u03c0\u03cc \u03c4\u03bf \u03b1\u03bd \u03ba\u03b1\u03c4\u03b5\u03b2\u03b1\u03af\u03bd\u03bf\u03c5\u03bd \u03c4\u03b1 weights, \u03b1\u03bb\u03bb\u03ac \u03b1\u03c0\u03cc capacity, latency, \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b1, monitoring \u03ba\u03b1\u03b9 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b9\u03ba\u03ae \u03b5\u03c5\u03b8\u03cd\u03bd\u03b7.<\/p>\n<h2 id=\"cache-aware-training-quality\">\u0393\u03b9\u03b1\u03c4\u03af \u03b7 cache-aware \u03b5\u03ba\u03c0\u03b1\u03af\u03b4\u03b5\u03c5\u03c3\u03b7 \u03b4\u03b5\u03bd \u03c0\u03ad\u03c1\u03b1\u03c3\u03b5 \u03c4\u03bf \u03cc\u03c1\u03b9\u03bf \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2<\/h2>\n<p>\u0397 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7 \u03c0\u03c1\u03bf\u03b5\u03b3\u03b3\u03c1\u03ac\u03c6\u03b7\u03ba\u03b5: \u03c5\u03c0\u03bf\u03b8\u03ad\u03c3\u03b5\u03b9\u03c2, metrics \u03ba\u03b1\u03b9 \u03cc\u03c1\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c4\u03c5\u03c7\u03af\u03b1\u03c2 \u03ba\u03b1\u03b8\u03bf\u03c1\u03af\u03c3\u03c4\u03b7\u03ba\u03b1\u03bd \u03c0\u03c1\u03b9\u03bd \u03b1\u03c0\u03cc \u03c4\u03b1 \u03ba\u03cd\u03c1\u03b9\u03b1 runs. \u0395\u03ba\u03c0\u03b1\u03b9\u03b4\u03b5\u03cd\u03c4\u03b7\u03ba\u03b1\u03bd MoE \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1 137 \u03b5\u03ba\u03b1\u03c4. \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03c4\u03c1\u03c9\u03bd \u03bc\u03b5 16 experts \u03ba\u03b1\u03b9 top-2 routing \u03c3\u03b5 300 \u03b5\u03ba\u03b1\u03c4. tokens \u03b1\u03c0\u03cc WikiText-103, codeparrot-clean \u03ba\u03b1\u03b9 OpenWebMath. \u0394\u03bf\u03ba\u03b9\u03bc\u03ac\u03c3\u03c4\u03b7\u03ba\u03b1\u03bd temporal-locality loss, \u03c0\u03bf\u03c5 \u03b5\u03c0\u03b9\u03b2\u03c1\u03b1\u03b2\u03b5\u03cd\u03b5\u03b9 \u03c0\u03b1\u03c1\u03cc\u03bc\u03bf\u03b9\u03bf routing \u03c3\u03b5 \u03b4\u03b9\u03b1\u03b4\u03bf\u03c7\u03b9\u03ba\u03ac tokens, \u03ba\u03b1\u03b9 domain-confinement loss, \u03c0\u03bf\u03c5 \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03af\u03b6\u03b5\u03b9 \u03ba\u03ac\u03b8\u03b5 domain \u03c3\u03b5 \u03c0\u03c1\u03bf\u03ba\u03b1\u03b8\u03bf\u03c1\u03b9\u03c3\u03bc\u03ad\u03bd\u03bf slice experts.<\/p>\n<p>\u03a4\u03bf locality \u03ba\u03c1\u03b9\u03c4\u03ae\u03c1\u03b9\u03bf \u03b6\u03b7\u03c4\u03bf\u03cd\u03c3\u03b5 \u03c4\u03bf\u03c5\u03bb\u03ac\u03c7\u03b9\u03c3\u03c4\u03bf\u03bd 30% \u03bc\u03b5\u03af\u03c9\u03c3\u03b7 cache misses \u03ba\u03b1\u03b9 \u03b1\u03cd\u03be\u03b7\u03c3\u03b7 perplexity \u03ad\u03c9\u03c2 1%. \u03a3\u03b5 \u03c7\u03b1\u03bc\u03b7\u03bb\u03cc \u03b2\u03ac\u03c1\u03bf\u03c2, \u03b7 \u03b1\u03cd\u03be\u03b7\u03c3\u03b7 perplexity \u03ae\u03c4\u03b1\u03bd 0,3%, \u03b1\u03bb\u03bb\u03ac \u03b7 \u03bc\u03b5\u03af\u03c9\u03c3\u03b7 misses \u03bc\u03cc\u03bb\u03b9\u03c2 15%. \u03a3\u03b5 \u03c5\u03c8\u03b7\u03bb\u03cc\u03c4\u03b5\u03c1\u03bf \u03b2\u03ac\u03c1\u03bf\u03c2, \u03c4\u03b1 misses \u03bc\u03b5\u03b9\u03ce\u03b8\u03b7\u03ba\u03b1\u03bd 59% \u03ad\u03c9\u03c2 60% \u03c3\u03b5 \u03b4\u03cd\u03bf seeds, \u03cc\u03bc\u03c9\u03c2 \u03c4\u03bf perplexity \u03b1\u03c5\u03be\u03ae\u03b8\u03b7\u03ba\u03b5 2,1% \u03ad\u03c9\u03c2 3,1%. \u0397 domain \u03b5\u03ba\u03b4\u03bf\u03c7\u03ae \u03ad\u03c6\u03c4\u03b1\u03c3\u03b5 static-pin hit rate 99,1%, \u03b1\u03bb\u03bb\u03ac \u03bc\u03b5 \u03b1\u03cd\u03be\u03b7\u03c3\u03b7 perplexity 3,1%, \u03c0\u03ac\u03bd\u03c9 \u03b1\u03c0\u03cc \u03c4\u03bf \u03cc\u03c1\u03b9\u03bf 2%.<\/p>\n<p>\u03a4\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c1\u03bd\u03b7\u03c4\u03b9\u03ba\u03cc \u03bc\u03cc\u03bd\u03bf \u03c9\u03c2 \u03c0\u03c1\u03bf\u03c2 \u03c4\u03bf \u03ba\u03bf\u03b9\u03bd\u03cc, \u03c0\u03c1\u03bf\u03b5\u03b3\u03b3\u03b5\u03b3\u03c1\u03b1\u03bc\u03bc\u03ad\u03bd\u03bf \u03cc\u03c1\u03b9\u03bf. \u0397 cacheability \u03cc\u03bd\u03c4\u03c9\u03c2 \u03b2\u03b5\u03bb\u03c4\u03b9\u03ce\u03b8\u03b7\u03ba\u03b5, \u03b1\u03bb\u03bb\u03ac \u03b4\u03b5\u03bd \u03ae\u03c4\u03b1\u03bd \u03b4\u03c9\u03c1\u03b5\u03ac\u03bd \u03c3\u03c4\u03bf \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf multi-domain scale. \u03a4\u03bf paper \u03b1\u03bd\u03c4\u03b9\u03c0\u03b1\u03c1\u03b1\u03b2\u03ac\u03bb\u03bb\u03b5\u03b9 \u03c0\u03c1\u03bf\u03c3\u03b5\u03ba\u03c4\u03b9\u03ba\u03ac \u03c4\u03bf StickyMoE, \u03c0\u03bf\u03c5 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7\u03c3\u03b5 \u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03b1 single-domain \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1. \u03a4\u03b1 \u03b4\u03cd\u03bf \u03c0\u03b5\u03b9\u03c1\u03ac\u03bc\u03b1\u03c4\u03b1 \u03b4\u03b5\u03bd \u03b4\u03b9\u03ba\u03b1\u03b9\u03bf\u03bb\u03bf\u03b3\u03bf\u03cd\u03bd \u03ba\u03b1\u03b8\u03bf\u03bb\u03b9\u03ba\u03ae \u03b5\u03c4\u03c5\u03bc\u03b7\u03b3\u03bf\u03c1\u03af\u03b1\u00b7 \u03b4\u03b5\u03af\u03c7\u03bd\u03bf\u03c5\u03bd \u03cc\u03c4\u03b9 corpus, \u03bc\u03ad\u03b3\u03b5\u03b8\u03bf\u03c2, budget \u03ba\u03b1\u03b9 \u03ba\u03c1\u03b9\u03c4\u03ae\u03c1\u03b9\u03bf \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2 \u03b1\u03bb\u03bb\u03ac\u03b6\u03bf\u03c5\u03bd \u03c4\u03bf trade-off.<\/p>\n<h2 id=\"training-rerouting-syndyasmos\">\u039f \u03c3\u03c5\u03bd\u03b4\u03c5\u03b1\u03c3\u03bc\u03cc\u03c2 training \u03ba\u03b1\u03b9 rerouting<\/h2>\n<p>\u0397 \u03c0\u03b9\u03bf \u03c5\u03c0\u03bf\u03c3\u03c7\u03cc\u03bc\u03b5\u03bd\u03b7 \u03ba\u03b1\u03c4\u03b5\u03cd\u03b8\u03c5\u03bd\u03c3\u03b7 \u03c0\u03c1\u03bf\u03ad\u03ba\u03c5\u03c8\u03b5 \u03cc\u03c4\u03b1\u03bd \u03b7 \u03b5\u03ba\u03c0\u03b1\u03b9\u03b4\u03b5\u03c5\u03bc\u03ad\u03bd\u03b7 locality \u03c3\u03c5\u03bd\u03b4\u03c5\u03ac\u03c3\u03c4\u03b7\u03ba\u03b5 \u03bc\u03b5 training-free cache-aware rerouting \u03c3\u03c4\u03bf inference. \u0391\u03bd \u03bf \u03ba\u03bf\u03c1\u03c5\u03c6\u03b1\u03af\u03bf\u03c2 expert \u03b4\u03b5\u03bd \u03b2\u03c1\u03b9\u03c3\u03ba\u03cc\u03c4\u03b1\u03bd \u03c3\u03c4\u03b7\u03bd cache, \u03c4\u03bf \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03c3\u03b5 \u03bd\u03b1 \u03b5\u03c0\u03b9\u03bb\u03ad\u03be\u03b5\u03b9 cached expert \u03bc\u03cc\u03bd\u03bf \u03cc\u03c4\u03b1\u03bd \u03c4\u03bf router score \u03c4\u03bf\u03c5 \u03ae\u03c4\u03b1\u03bd \u03bc\u03ad\u03c3\u03b1 \u03c3\u03b5 \u03c0\u03c1\u03bf\u03ba\u03b1\u03b8\u03bf\u03c1\u03b9\u03c3\u03bc\u03ad\u03bd\u03b7 \u03c3\u03c7\u03b5\u03c4\u03b9\u03ba\u03ae \u03b1\u03bd\u03bf\u03c7\u03ae.<\/p>\n<p>\u03a0\u03ac\u03bd\u03c9 \u03c3\u03c4\u03bf baseline, \u03c4\u03bf \u03ad\u03bd\u03c4\u03bf\u03bd\u03bf rerouting \u03c5\u03c0\u03bf\u03b2\u03ac\u03b8\u03bc\u03b9\u03b6\u03b5 \u03b3\u03c1\u03ae\u03b3\u03bf\u03c1\u03b1 \u03c4\u03b7\u03bd \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1. \u03a0\u03ac\u03bd\u03c9 \u03c3\u03c4\u03bf locality-trained \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf, \u03b7 \u03af\u03b4\u03b9\u03b1 \u03c0\u03b1\u03c1\u03ad\u03bc\u03b2\u03b1\u03c3\u03b7 \u03b5\u03af\u03c7\u03b5 \u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf \u03bf\u03c1\u03b9\u03b1\u03ba\u03cc \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2. \u03a3\u03c4\u03b1 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1 \u03c4\u03c9\u03bd 137 \u03b5\u03ba\u03b1\u03c4. \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03c4\u03c1\u03c9\u03bd, \u03b7 training-time locality \u03ad\u03b4\u03b9\u03bd\u03b5 58% \u03ad\u03c9\u03c2 61% \u03bc\u03b5\u03af\u03c9\u03c3\u03b7 misses \u03bc\u03b5 1,4% \u03ad\u03c9\u03c2 2,2% \u03b1\u03cd\u03be\u03b7\u03c3\u03b7 perplexity. \u039c\u03b5 \u03c4\u03bf \u03c0\u03c1\u03cc\u03c3\u03b8\u03b5\u03c4\u03bf rerouting, \u03b7 \u03bc\u03b5\u03af\u03c9\u03c3\u03b7 \u03ad\u03c6\u03c4\u03b1\u03bd\u03b5 80% \u03ad\u03c9\u03c2 82% \u03bc\u03b5 \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03cc \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 1,9% \u03ad\u03c9\u03c2 3,1%. \u03a3\u03c4\u03bf rung \u03c4\u03c9\u03bd 340 \u03b5\u03ba\u03b1\u03c4. \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03c4\u03c1\u03c9\u03bd, \u03c4\u03bf stacking \u03ad\u03c6\u03c4\u03b1\u03c3\u03b5 82% \u03bc\u03b5\u03af\u03c9\u03c3\u03b7 misses \u03bc\u03b5 \u03b1\u03cd\u03be\u03b7\u03c3\u03b7 perplexity 3,4%.<\/p>\n<p>\u0391\u03c5\u03c4\u03ac \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1 \u03c4\u03c9\u03bd \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03c9\u03bd checkpoints \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03b5\u03b3\u03b3\u03cd\u03b7\u03c3\u03b7 \u03b3\u03b9\u03b1 production Qwen3. \u0397 \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae \u03b9\u03b4\u03ad\u03b1 \u03cc\u03bc\u03c9\u03c2 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03b7: training objective, router, cache policy \u03ba\u03b1\u03b9 serving engine \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03bd\u03b1 \u03c3\u03c5\u03bd-\u03c3\u03c7\u03b5\u03b4\u03b9\u03b1\u03c3\u03c4\u03bf\u03cd\u03bd, \u03b1\u03bd\u03c4\u03af \u03ba\u03ac\u03b8\u03b5 \u03b5\u03c0\u03af\u03c0\u03b5\u03b4\u03bf \u03bd\u03b1 \u03b2\u03b5\u03bb\u03c4\u03b9\u03c3\u03c4\u03bf\u03c0\u03bf\u03b9\u03b5\u03af\u03c4\u03b1\u03b9 \u03b1\u03bd\u03b5\u03be\u03ac\u03c1\u03c4\u03b7\u03c4\u03b1.<\/p>\n<h2 id=\"epicheiriseis-edge-ai\">\u03a4\u03b9 \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 \u03b3\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03ae\u03c3\u03b5\u03b9\u03c2 \u03ba\u03b1\u03b9 \u03bf\u03bc\u03ac\u03b4\u03b5\u03c2 AI<\/h2>\n<p>\u0393\u03b9\u03b1 \u03bc\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03af\u03c1\u03b7\u03c3\u03b7, \u03c4\u03bf \u03b2\u03b1\u03c3\u03b9\u03ba\u03cc \u03bc\u03ac\u03b8\u03b7\u03bc\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03bd\u03b1 \u03bc\u03b7 \u03c3\u03c5\u03b3\u03c7\u03ad\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b9\u03b4\u03b9\u03c9\u03c4\u03b9\u03ba\u03ae \u03ae \u03c4\u03bf\u03c0\u03b9\u03ba\u03ae \u03b5\u03ba\u03c4\u03ad\u03bb\u03b5\u03c3\u03b7 \u03bc\u03b5 \u03c4\u03b7\u03bd \u03ad\u03c4\u03bf\u03b9\u03bc\u03b7 \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03b9\u03ba\u03ae \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03af\u03b1. \u0388\u03bd\u03b1 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c6\u03bf\u03c1\u03c4\u03ce\u03bd\u03b5\u03c4\u03b1\u03b9 \u03c3\u03b5 desktop \u03ba\u03b1\u03b9 \u03bd\u03b1 \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03c5\u03c0\u03b5\u03c1\u03b2\u03bf\u03bb\u03b9\u03ba\u03ac \u03b1\u03c1\u03b3\u03cc \u03b3\u03b9\u03b1 support, live content assistance \u03ae interactive copilots. \u0391\u03bd\u03c4\u03af\u03c3\u03c4\u03c1\u03bf\u03c6\u03b1, \u03ad\u03bd\u03b1 \u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c0\u03c1\u03bf\u03c3\u03c6\u03ad\u03c1\u03b5\u03b9 \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03b1\u03bd\u03ac \u03b5\u03c0\u03b9\u03c4\u03c5\u03c7\u03b7\u03bc\u03ad\u03bd\u03b7 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1. \u0393\u03b9\u03b1 \u03b1\u03c5\u03c4\u03cc \u03b7 <a href=\"https:\/\/twodots.gr\/poio-ai-montelo-axizei-gia-tin-epicheirisi\/\">\u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ae AI \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf\u03c5 \u03b3\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03af\u03c1\u03b7\u03c3\u03b7<\/a> \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 workload-specific evaluation \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03bc\u03cc\u03bd\u03b9\u03bc\u03bf \u03c3\u03c4\u03bf\u03af\u03c7\u03b7\u03bc\u03b1 \u03c3\u03c4\u03bf \u03bc\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf checkpoint.<\/p>\n<p>\u0397 \u03bc\u03ad\u03c3\u03b7 \u03c4\u03b1\u03c7\u03cd\u03c4\u03b7\u03c4\u03b1 \u03b5\u03c0\u03af\u03c3\u03b7\u03c2 \u03b4\u03b5\u03bd \u03b1\u03c1\u03ba\u03b5\u03af \u03b3\u03b9\u03b1 \u03b5\u03bc\u03c0\u03b5\u03b9\u03c1\u03af\u03b1 \u03c7\u03c1\u03ae\u03c3\u03c4\u03b7. \u0388\u03bd\u03b1 cache hit \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c3\u03c7\u03b5\u03c4\u03b9\u03ba\u03ac \u03b3\u03c1\u03ae\u03b3\u03bf\u03c1\u03bf, \u03b5\u03bd\u03ce \u03ad\u03bd\u03b1 miss \u03bd\u03b1 \u03c0\u03b5\u03c1\u03b9\u03bc\u03ad\u03bd\u03b5\u03b9 SSD I\/O \u03ba\u03b1\u03b9 \u03bd\u03b1 \u03b4\u03b7\u03bc\u03b9\u03bf\u03c5\u03c1\u03b3\u03b5\u03af \u03b1\u03b9\u03c3\u03b8\u03b7\u03c4\u03ae \u03bf\u03c5\u03c1\u03ac. \u0393\u03b9\u03b1 interactive \u03c1\u03bf\u03ad\u03c2 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 percentiles latency, \u03c7\u03c1\u03cc\u03bd\u03bf\u03c2 \u03ad\u03c9\u03c2 \u03c4\u03bf \u03c0\u03c1\u03ce\u03c4\u03bf token, \u03c1\u03c5\u03b8\u03bc\u03cc\u03c2 \u03bc\u03b5\u03c4\u03ac \u03c4\u03bf \u03c0\u03c1\u03ce\u03c4\u03bf token \u03ba\u03b1\u03b9 \u03c3\u03c5\u03c7\u03bd\u03cc\u03c4\u03b7\u03c4\u03b1 \u03b1\u03ba\u03c1\u03b1\u03af\u03c9\u03bd \u03ba\u03b1\u03b8\u03c5\u03c3\u03c4\u03b5\u03c1\u03ae\u03c3\u03b5\u03c9\u03bd. \u0393\u03b9\u03b1 batch \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b5\u03c2, \u03b1\u03bd\u03c4\u03af\u03b8\u03b5\u03c4\u03b1, \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03ad\u03c7\u03b5\u03b9 \u03bc\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b7 \u03c3\u03b7\u03bc\u03b1\u03c3\u03af\u03b1 \u03b7 \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03ae \u03bf\u03bb\u03bf\u03ba\u03bb\u03ae\u03c1\u03c9\u03c3\u03b7 \u03b1\u03bd\u03ac \u03ce\u03c1\u03b1 \u03ba\u03b1\u03b9 \u03b7 \u03b5\u03bd\u03b5\u03c1\u03b3\u03b5\u03b9\u03b1\u03ba\u03ae \u03ba\u03b1\u03c4\u03b1\u03bd\u03ac\u03bb\u03c9\u03c3\u03b7. \u03a4\u03bf \u03af\u03b4\u03b9\u03bf hardware \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03ba\u03b1\u03c4\u03ac\u03bb\u03bb\u03b7\u03bb\u03bf \u03b3\u03b9\u03b1 live chat \u03b1\u03bb\u03bb\u03ac \u03b1\u03c0\u03bf\u03b4\u03b5\u03ba\u03c4\u03cc \u03b3\u03b9\u03b1 \u03c0\u03c1\u03bf\u03b3\u03c1\u03b1\u03bc\u03bc\u03b1\u03c4\u03b9\u03c3\u03bc\u03ad\u03bd\u03b7 \u03bd\u03c5\u03c7\u03c4\u03b5\u03c1\u03b9\u03bd\u03ae \u03b5\u03c0\u03b5\u03be\u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1.<\/p>\n<p>\u03a4\u03bf \u03b4\u03b5\u03cd\u03c4\u03b5\u03c1\u03bf \u03bc\u03ac\u03b8\u03b7\u03bc\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03cc\u03c4\u03b9 \u03c4\u03b1 averages \u03ba\u03c1\u03cd\u03b2\u03bf\u03c5\u03bd domain \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03ad\u03c2. \u03a3\u03c4\u03b1 traces, \u03bf \u03ba\u03ce\u03b4\u03b9\u03ba\u03b1\u03c2 \u03b5\u03af\u03c7\u03b5 \u03c3\u03c7\u03b5\u03b4\u03cc\u03bd \u03b4\u03b9\u03b1\u03ba\u03c1\u03b9\u03c4\u03cc \u03c3\u03cd\u03bd\u03bf\u03bb\u03bf experts. \u03a3\u03c4\u03b1 training runs, \u03c4\u03b1 penalties \u03b5\u03c0\u03af\u03c3\u03b7\u03c2 \u03b4\u03b9\u03ad\u03c6\u03b5\u03c1\u03b1\u03bd \u03b1\u03bd\u03ac domain. \u0391\u03bd \u03c4\u03bf \u03af\u03b4\u03b9\u03bf \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03b3\u03c1\u03ac\u03c6\u03b5\u03b9 product descriptions, \u03c3\u03c5\u03bd\u03bf\u03c8\u03af\u03b6\u03b5\u03b9 tickets \u03ba\u03b1\u03b9 \u03b2\u03bf\u03b7\u03b8\u03ac developers, \u03b7 cache policy \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b4\u03bf\u03ba\u03b9\u03bc\u03b1\u03c3\u03c4\u03b5\u03af \u03c3\u03c4\u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03bc\u03af\u03b3\u03bc\u03b1 \u03ba\u03b1\u03b9 \u03c3\u03c4\u03b9\u03c2 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ad\u03c2 \u03bc\u03b5\u03c4\u03b1\u03b2\u03ac\u03c3\u03b5\u03b9\u03c2 \u03bc\u03b5\u03c4\u03b1\u03be\u03cd \u03b5\u03c1\u03b3\u03b1\u03c3\u03b9\u03ce\u03bd.<\/p>\n<p>\u03a4\u03bf \u03c4\u03c1\u03af\u03c4\u03bf \u03bc\u03ac\u03b8\u03b7\u03bc\u03b1 \u03b1\u03c6\u03bf\u03c1\u03ac \u03c4\u03bf \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03cc \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2. \u03a0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03b7 RAM \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03bc\u03b5\u03b9\u03ce\u03c3\u03b5\u03b9 SSD I\/O, \u03b1\u03bb\u03bb\u03ac \u03b1\u03c5\u03be\u03ac\u03bd\u03b5\u03b9 \u03b5\u03c0\u03ad\u03bd\u03b4\u03c5\u03c3\u03b7 \u03ba\u03b1\u03b9 \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b1. Rerouting \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03bc\u03b5\u03b9\u03ce\u03c3\u03b5\u03b9 \u03bc\u03b5\u03c4\u03b1\u03c6\u03bf\u03c1\u03ad\u03c2, \u03b1\u03bb\u03bb\u03ac \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c0\u03b5\u03c1\u03ac\u03c3\u03b5\u03b9 quality gates. \u039c\u03b9\u03b1 rack-scale \u03bb\u03cd\u03c3\u03b7 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b4\u03ce\u03c3\u03b5\u03b9 \u03c0\u03bf\u03bb\u03cd \u03c5\u03c8\u03b7\u03bb\u03cc\u03c4\u03b5\u03c1\u03bf bandwidth, \u03b1\u03bb\u03bb\u03ac \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03b5\u03bd\u03c4\u03b5\u03bb\u03ce\u03c2 \u03bf\u03b9\u03ba\u03bf\u03bd\u03bf\u03bc\u03af\u03b1 \u03ba\u03b1\u03b9 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03af\u03b1\u00b7 \u03b1\u03c5\u03c4\u03cc \u03c6\u03b1\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03ba\u03b1\u03b9 \u03cc\u03c4\u03b1\u03bd \u03b7 <a href=\"https:\/\/twodots.gr\/en\/amd-ai-data-centers-symfonies-meta-openai\/\">AI \u03c5\u03c0\u03bf\u03b4\u03bf\u03bc\u03ae \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03b5\u03af\u03c4\u03b1\u03b9 \u03c3\u03b5 \u03b5\u03c0\u03af\u03c0\u03b5\u03b4\u03bf \u03c0\u03bb\u03ae\u03c1\u03bf\u03c5\u03c2 stack<\/a> \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03bc\u03b5\u03bc\u03bf\u03bd\u03c9\u03bc\u03ad\u03bd\u03bf\u03c5 accelerator.<\/p>\n<p>\u03a4\u03ad\u03bb\u03bf\u03c2, \u03c4\u03bf local \u03b4\u03b5\u03bd \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 \u03b1\u03c5\u03c4\u03bf\u03bc\u03ac\u03c4\u03c9\u03c2 \u03b1\u03be\u03b9\u03cc\u03c0\u03b9\u03c3\u03c4\u03bf. \u0397 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03ac \u03b1\u03bd\u03ac\u03bc\u03b5\u03c3\u03b1 \u03c3\u03c4\u03bf \u00ab\u03c4\u03c1\u03ad\u03c7\u03b5\u03b9\u00bb \u03ba\u03b1\u03b9 \u03c3\u03c4\u03bf \u00ab\u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b5\u03af \u03c3\u03c9\u03c3\u03c4\u03ac\u00bb \u03b9\u03c3\u03c7\u03cd\u03b5\u03b9 \u03c4\u03cc\u03c3\u03bf \u03b3\u03b9\u03b1 \u03c4\u03b5\u03c1\u03ac\u03c3\u03c4\u03b9\u03b1 MoE \u03cc\u03c3\u03bf \u03ba\u03b1\u03b9 \u03b3\u03b9\u03b1 <a href=\"https:\/\/twodots.gr\/en\/twil-lm-mikra-montela-typikis-logikis-topika\/\">\u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03b1 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1 \u03c0\u03bf\u03c5 \u03b5\u03ba\u03c4\u03b5\u03bb\u03bf\u03cd\u03bd\u03c4\u03b1\u03b9 \u03c4\u03bf\u03c0\u03b9\u03ba\u03ac<\/a>. \u03a7\u03c1\u03b5\u03b9\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b1\u03bd\u03b5\u03be\u03ac\u03c1\u03c4\u03b7\u03c4\u03b5\u03c2 \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03ae\u03c3\u03b5\u03b9\u03c2, observability, fallback \u03ba\u03b1\u03b9 \u03c3\u03b1\u03c6\u03ae\u03c2 \u03b9\u03b4\u03b9\u03bf\u03ba\u03c4\u03b7\u03c3\u03af\u03b1 \u03c4\u03b7\u03c2 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7\u03c2.<\/p>\n<div class=\"td-decision-band\">\n<div class=\"td-decision-band-content\"><strong>\u039c\u03b7\u03bd \u03b5\u03b3\u03ba\u03c1\u03af\u03bd\u03b5\u03c4\u03b5 edge MoE pilot \u03b5\u03c0\u03b5\u03b9\u03b4\u03ae \u03c4\u03bf checkpoint \u03c6\u03bf\u03c1\u03c4\u03ce\u03bd\u03b5\u03c4\u03b1\u03b9 \u03c3\u03b5 \u03bc\u03af\u03b1 GPU.<\/strong><span>\u039f\u03c1\u03af\u03c3\u03c4\u03b5 \u03c3\u03c4\u03cc\u03c7\u03bf latency \u03ba\u03b1\u03b9 \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2, \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03c4\u03b5 bytes \u03b1\u03bd\u03ac token \u03ba\u03b1\u03b9 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc routing, \u03b4\u03bf\u03ba\u03b9\u03bc\u03ac\u03c3\u03c4\u03b5 peak concurrency, \u03ba\u03bf\u03c3\u03c4\u03bf\u03bb\u03bf\u03b3\u03ae\u03c3\u03c4\u03b5 \u03cc\u03bb\u03b1 \u03c4\u03b1 memory tiers \u03ba\u03b1\u03b9 \u03ba\u03c1\u03b1\u03c4\u03ae\u03c3\u03c4\u03b5 \u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03ae managed endpoint \u03c9\u03c2 \u03c3\u03c5\u03b3\u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03bf baseline.<\/span><\/div>\n<\/div>\n<h2 id=\"elegchoi-pilot-moe\">\u0395\u03c0\u03c4\u03ac \u03ad\u03bb\u03b5\u03b3\u03c7\u03bf\u03b9 \u03c0\u03c1\u03b9\u03bd \u03b1\u03c0\u03cc pilot MoE \u03c3\u03c4\u03b7\u03bd \u03ac\u03ba\u03c1\u03b7<\/h2>\n<p>\u0388\u03bd\u03b1 \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf pilot \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b4\u03b9\u03b1\u03c8\u03b5\u03cd\u03c3\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b1\u03c1\u03c7\u03b9\u03ba\u03ae \u03c5\u03c0\u03cc\u03b8\u03b5\u03c3\u03b7. \u0394\u03b5\u03bd \u03be\u03b5\u03ba\u03b9\u03bd\u03ac \u03b1\u03c0\u03cc \u03b1\u03b3\u03bf\u03c1\u03ac hardware, \u03b1\u03bb\u03bb\u03ac \u03b1\u03c0\u03cc workload, \u03b1\u03c0\u03bf\u03b4\u03b5\u03ba\u03c4\u03cc \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03ba\u03b1\u03b9 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf operating point.<\/p>\n<div class=\"td-step-list\">\n<p class=\"td-step-list-title\">\u0395\u03c0\u03c4\u03ac workload-first \u03ad\u03bb\u03b5\u03b3\u03c7\u03bf\u03b9 \u03b3\u03b9\u03b1 edge MoE inference<\/p>\n<ol>\n<li><span class=\"td-step-kicker\">\u0392\u03ae\u03bc\u03b1 1<\/span><strong>\u039f\u03c1\u03af\u03c3\u03c4\u03b5 \u03c4\u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc workload<\/strong>\n<p>\u039a\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c8\u03c4\u03b5 input \u03ba\u03b1\u03b9 output length, \u03b3\u03bb\u03ce\u03c3\u03c3\u03b5\u03c2, domains, concurrent \u03c7\u03c1\u03ae\u03c3\u03c4\u03b5\u03c2, \u03ce\u03c1\u03b5\u03c2 \u03b1\u03b9\u03c7\u03bc\u03ae\u03c2 \u03ba\u03b1\u03b9 \u03c0\u03bf\u03b9\u03b1 \u03ba\u03b1\u03b8\u03c5\u03c3\u03c4\u03ad\u03c1\u03b7\u03c3\u03b7 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c0\u03bf\u03b4\u03b5\u03ba\u03c4\u03ae \u03b1\u03bd\u03ac \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u0392\u03ae\u03bc\u03b1 2<\/span><strong>\u039c\u03b5\u03c4\u03c1\u03ae\u03c3\u03c4\u03b5 quality \u03c0\u03c1\u03b9\u03bd \u03b1\u03c0\u03cc optimization<\/strong>\n<p>\u03a6\u03c4\u03b9\u03ac\u03be\u03c4\u03b5 holdout set \u03bc\u03b5 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ad\u03c2 \u03c0\u03b5\u03c1\u03b9\u03c0\u03c4\u03ce\u03c3\u03b5\u03b9\u03c2 \u03ba\u03b1\u03b9 \u03ba\u03bb\u03b5\u03b9\u03b4\u03ce\u03c3\u03c4\u03b5 \u03cc\u03c1\u03b9\u03b1 \u03b3\u03b9\u03b1 \u03b1\u03ba\u03c1\u03af\u03b2\u03b5\u03b9\u03b1, \u03bc\u03bf\u03c1\u03c6\u03ae, \u03b1\u03c3\u03c6\u03ac\u03bb\u03b5\u03b9\u03b1 \u03ba\u03b1\u03b9 \u03b1\u03bd\u03b8\u03c1\u03ce\u03c0\u03b9\u03bd\u03b5\u03c2 \u03b4\u03b9\u03bf\u03c1\u03b8\u03ce\u03c3\u03b5\u03b9\u03c2 \u03c0\u03c1\u03b9\u03bd \u03b1\u03bb\u03bb\u03ac\u03be\u03b5\u03c4\u03b5 routing \u03ae cache.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u0392\u03ae\u03bc\u03b1 3<\/span><strong>\u03a7\u03b1\u03c1\u03c4\u03bf\u03b3\u03c1\u03b1\u03c6\u03ae\u03c3\u03c4\u03b5 \u03ba\u03ac\u03b8\u03b5 memory tier<\/strong>\n<p>\u039c\u03b5\u03c4\u03c1\u03ae\u03c3\u03c4\u03b5 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc VRAM, RAM \u03ba\u03b1\u03b9 storage bandwidth, \u03b4\u03b9\u03b1\u03b8\u03ad\u03c3\u03b9\u03bc\u03b7 \u03c7\u03c9\u03c1\u03b7\u03c4\u03b9\u03ba\u03cc\u03c4\u03b7\u03c4\u03b1, bytes \u03b1\u03bd\u03ac token \u03ba\u03b1\u03b9 \u03c7\u03c1\u03cc\u03bd\u03bf \u03bc\u03b5\u03c4\u03b1\u03c6\u03bf\u03c1\u03ac\u03c2 \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03b2\u03b1\u03c3\u03af\u03b6\u03b5\u03c3\u03c4\u03b5 \u03bc\u03cc\u03bd\u03bf \u03c3\u03c4\u03b9\u03c2 \u03bf\u03bd\u03bf\u03bc\u03b1\u03c3\u03c4\u03b9\u03ba\u03ad\u03c2 \u03c0\u03c1\u03bf\u03b4\u03b9\u03b1\u03b3\u03c1\u03b1\u03c6\u03ad\u03c2.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u0392\u03ae\u03bc\u03b1 4<\/span><strong>\u03a3\u03c5\u03bb\u03bb\u03ad\u03be\u03c4\u03b5 routing traces \u03c4\u03bf\u03c5 \u03b4\u03b9\u03ba\u03bf\u03cd \u03c3\u03b1\u03c2 \u03bc\u03af\u03b3\u03bc\u03b1\u03c4\u03bf\u03c2<\/strong>\n<p>\u0395\u03bb\u03ad\u03b3\u03be\u03c4\u03b5 temporal locality, working set \u03ba\u03b1\u03b9 \u03b1\u03bb\u03bb\u03b1\u03b3\u03ad\u03c2 \u03b1\u03bd\u03ac domain. \u039c\u03b7\u03bd \u03bc\u03b5\u03c4\u03b1\u03c6\u03ad\u03c1\u03b5\u03c4\u03b5 cache hit rate \u03b1\u03c0\u03cc \u03ac\u03bb\u03bb\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03ae corpus \u03c3\u03b1\u03bd \u03bd\u03b1 \u03ae\u03c4\u03b1\u03bd \u03b4\u03b9\u03ba\u03ae \u03c3\u03b1\u03c2 \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b7.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u0392\u03ae\u03bc\u03b1 5<\/span><strong>\u0394\u03bf\u03ba\u03b9\u03bc\u03ac\u03c3\u03c4\u03b5 cache \u03ba\u03b1\u03b9 batching \u03bc\u03b1\u03b6\u03af<\/strong>\n<p>\u039c\u03b5\u03c4\u03c1\u03ae\u03c3\u03c4\u03b5 single-stream latency \u03ba\u03b1\u03b9 aggregate throughput \u03c3\u03b5 \u03c1\u03b5\u03b1\u03bb\u03b9\u03c3\u03c4\u03b9\u03ba\u03cc concurrency, \u03bc\u03b1\u03b6\u03af \u03bc\u03b5 page faults, disk utilization, queue time \u03ba\u03b1\u03b9 cache warm-up.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u0392\u03ae\u03bc\u03b1 6<\/span><strong>\u0395\u03c0\u03b1\u03bd\u03b1\u03bb\u03ac\u03b2\u03b5\u03c4\u03b5 \u03c4\u03bf quality gate \u03c3\u03c4\u03bf operating point<\/strong>\n<p>\u0391\u03bd \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af\u03c4\u03b5 rerouting, quantization \u03ae \u03ac\u03bb\u03bb\u03bf approximation, \u03c4\u03c1\u03ad\u03be\u03c4\u03b5 \u03be\u03b1\u03bd\u03ac \u03c4\u03bf \u03af\u03b4\u03b9\u03bf holdout \u03ba\u03b1\u03b9 \u03c3\u03c5\u03b3\u03ba\u03c1\u03af\u03bd\u03b5\u03c4\u03b5 \u03b1\u03bd\u03ac domain, \u03cc\u03c7\u03b9 \u03bc\u03cc\u03bd\u03bf \u03c3\u03b5 \u03ad\u03bd\u03b1\u03bd \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03cc \u03bc\u03ad\u03c3\u03bf \u03cc\u03c1\u03bf.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u0392\u03ae\u03bc\u03b1 7<\/span><strong>\u03a3\u03c5\u03b3\u03ba\u03c1\u03af\u03bd\u03b5\u03c4\u03b5 \u03b1\u03bd\u03b1\u03c3\u03c4\u03c1\u03ad\u03c8\u03b9\u03bc\u03b5\u03c2 \u03b5\u03bd\u03b1\u03bb\u03bb\u03b1\u03ba\u03c4\u03b9\u03ba\u03ad\u03c2<\/strong>\n<p>\u0392\u03ac\u03bb\u03c4\u03b5 \u03b4\u03af\u03c0\u03bb\u03b1 \u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf, managed endpoint \u03ba\u03b1\u03b9 \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf local hardware. \u03a3\u03c5\u03b3\u03ba\u03c1\u03af\u03bd\u03b5\u03c4\u03b5 \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03b1\u03bd\u03ac \u03b5\u03c0\u03b9\u03c4\u03c5\u03c7\u03b7\u03bc\u03ad\u03bd\u03b7 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1, \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b1, \u03c3\u03c5\u03bd\u03c4\u03ae\u03c1\u03b7\u03c3\u03b7, \u03b9\u03b4\u03b9\u03c9\u03c4\u03b9\u03ba\u03cc\u03c4\u03b7\u03c4\u03b1 \u03ba\u03b1\u03b9 \u03c3\u03c7\u03ad\u03b4\u03b9\u03bf fallback.<\/p>\n<\/li>\n<\/ol>\n<\/div>\n<h2 id=\"periorismoi-meletis\">\u03a0\u03b5\u03c1\u03b9\u03bf\u03c1\u03b9\u03c3\u03bc\u03bf\u03af \u03c0\u03bf\u03c5 \u03b4\u03b5\u03bd \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c7\u03b1\u03b8\u03bf\u03cd\u03bd<\/h2>\n<p>\u0397 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 arXiv v1 \u03c0\u03c1\u03bf\u03b4\u03b7\u03bc\u03bf\u03c3\u03af\u03b5\u03c5\u03c3\u03b7 \u03b5\u03bd\u03cc\u03c2 \u03b1\u03bd\u03b5\u03be\u03ac\u03c1\u03c4\u03b7\u03c4\u03bf\u03c5 \u03b5\u03c1\u03b5\u03c5\u03bd\u03b7\u03c4\u03ae \u03ba\u03b1\u03b9 \u03b4\u03b5\u03bd \u03ad\u03c7\u03b5\u03b9 \u03c0\u03b5\u03c1\u03ac\u03c3\u03b5\u03b9 peer review. \u039f\u03b9 serving \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b5\u03b9\u03c2 \u03ad\u03b3\u03b9\u03bd\u03b1\u03bd \u03c3\u03b5 \u03ad\u03bd\u03b1\u03bd \u03c5\u03c0\u03bf\u03bb\u03bf\u03b3\u03b9\u03c3\u03c4\u03ae, \u03ad\u03bd\u03b1\u03bd DRAM-less PCIe 3 SSD, \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf GGUF \u03ba\u03b1\u03b9 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b7 \u03b4\u03b9\u03b1\u03bc\u03cc\u03c1\u03c6\u03c9\u03c3\u03b7 llama.cpp. \u0386\u03bb\u03bb\u03bf hardware, engine, kernel, filesystem \u03ae sequential-sweep scheduler \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b1\u03bb\u03bb\u03ac\u03be\u03b5\u03b9 \u03bf\u03c5\u03c3\u03b9\u03b1\u03c3\u03c4\u03b9\u03ba\u03ac \u03c4\u03b1 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1.<\/p>\n<p>\u03a4\u03bf Path-Mapped Serving \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03b5\u03af \u03c4\u03b1 \u03ae\u03b4\u03b7 \u03c0\u03b1\u03c1\u03b1\u03b4\u03bf\u03bc\u03ad\u03bd\u03b1 \u03c4\u03b5\u03c4\u03c1\u03ac\u03bc\u03c0\u03b9\u03c4\u03b1 \u03b2\u03ac\u03c1\u03b7 bit-exact \u03ba\u03b1\u03b9 \u03b4\u03b5\u03bd \u03c0\u03c1\u03bf\u03c3\u03b8\u03ad\u03c4\u03b5\u03b9 \u03bd\u03ad\u03b1 quantization \u03ae distillation. \u0391\u03c5\u03c4\u03cc \u03b5\u03be\u03b7\u03b3\u03b5\u03af \u03c4\u03b7 \u03b4\u03b9\u03b1\u03c4\u03cd\u03c0\u03c9\u03c3\u03b7 \u00ab\u03c0\u03bb\u03ae\u03c1\u03b7\u03c2 \u03b1\u03ba\u03c1\u03af\u03b2\u03b5\u03b9\u03b1\u00bb \u03bc\u03ad\u03c3\u03b1 \u03c3\u03c4\u03bf paper: \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 \u03af\u03b4\u03b9\u03b1 weights \u03bc\u03b5 \u03c4\u03bf Q4_K_M artifact \u03c0\u03bf\u03c5 \u03c3\u03b5\u03c1\u03b2\u03af\u03c1\u03b5\u03c4\u03b1\u03b9, \u03cc\u03c7\u03b9 \u03b9\u03c3\u03bf\u03b4\u03c5\u03bd\u03b1\u03bc\u03af\u03b1 \u03bc\u03b5 BF16 checkpoint \u03bf\u03cd\u03c4\u03b5 \u03b1\u03c0\u03cc\u03b4\u03b5\u03b9\u03be\u03b7 \u03cc\u03c4\u03b9 \u03b7 \u03b1\u03c1\u03c7\u03b9\u03ba\u03ae \u03c4\u03b5\u03c4\u03c1\u03b1\u03bc\u03c0\u03b9\u03c4\u03b7 \u03ad\u03ba\u03b4\u03bf\u03c3\u03b7 \u03ad\u03c7\u03b5\u03b9 \u03bc\u03b7\u03b4\u03b5\u03bd\u03b9\u03ba\u03ae \u03b1\u03c0\u03cc\u03ba\u03bb\u03b9\u03c3\u03b7 \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2.<\/p>\n<p>\u03a4\u03b1 routing traces \u03ad\u03b3\u03b9\u03bd\u03b1\u03bd \u03c3\u03c4\u03bf Qwen3-30B-A3B, \u03cc\u03c7\u03b9 \u03c3\u03c4\u03bf artifact \u03c4\u03c9\u03bd 235B. \u03a4\u03b1 training \u03c0\u03b5\u03b9\u03c1\u03ac\u03bc\u03b1\u03c4\u03b1 \u03ad\u03b3\u03b9\u03bd\u03b1\u03bd \u03c3\u03b5 137M \u03ba\u03b1\u03b9 340M \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03c4\u03c1\u03bf\u03c5\u03c2, \u03c0\u03bf\u03bb\u03cd \u03ba\u03ac\u03c4\u03c9 \u03b1\u03c0\u03cc frontier scale, \u03ba\u03b1\u03b9 \u03c4\u03bf \u03bc\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf rung \u03ae\u03c4\u03b1\u03bd undertrained \u03c3\u03c4\u03bf matched budget. \u0397 perplexity \u03b5\u03af\u03bd\u03b1\u03b9 \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf proxy, \u03cc\u03c7\u03b9 \u03c0\u03bb\u03ae\u03c1\u03b7\u03c2 \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b7 \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2 \u03b3\u03b9\u03b1 business tasks. \u0395\u03c0\u03bf\u03bc\u03ad\u03bd\u03c9\u03c2, \u03bf\u03cd\u03c4\u03b5 \u03c4\u03bf quality tax \u03bf\u03cd\u03c4\u03b5 \u03c4\u03b1 hit rates \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03bc\u03b5\u03c4\u03b1\u03c6\u03b5\u03c1\u03b8\u03bf\u03cd\u03bd \u03b1\u03c5\u03c4\u03bf\u03cd\u03c3\u03b9\u03b1 \u03c3\u03b5 \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03b9\u03ba\u03cc \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1.<\/p>\n<p>\u03a4\u03b1 \u03b4\u03b7\u03bc\u03bf\u03c3\u03b9\u03b5\u03c5\u03bc\u03ad\u03bd\u03b1 hit rates \u03ba\u03b1\u03b9 \u03c4\u03bf quality tax \u03b4\u03b5\u03bd \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03bc\u03b5\u03c4\u03b1\u03c6\u03b5\u03c1\u03b8\u03bf\u03cd\u03bd \u03b1\u03c5\u03c4\u03bf\u03cd\u03c3\u03b9\u03b1 \u03c3\u03b5 \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03b9\u03ba\u03cc \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1. \u03a0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b1\u03bd\u03c4\u03b9\u03bc\u03b5\u03c4\u03c9\u03c0\u03b9\u03c3\u03c4\u03bf\u03cd\u03bd \u03c9\u03c2 \u03c5\u03c0\u03bf\u03b8\u03ad\u03c3\u03b5\u03b9\u03c2 \u03c0\u03c1\u03bf\u03c2 \u03b4\u03bf\u03ba\u03b9\u03bc\u03ae \u03bc\u03b5 \u03af\u03b4\u03b9\u03bf model artifact, \u03af\u03b4\u03b9\u03b1 \u03ba\u03b1\u03c4\u03b1\u03bd\u03bf\u03bc\u03ae \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03c9\u03bd \u03ba\u03b1\u03b9 \u03af\u03b4\u03b9\u03b1 \u03c0\u03bf\u03bb\u03b9\u03c4\u03b9\u03ba\u03ae cache.<\/p>\n<p>\u0397 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7 \u03b5\u03bd\u03b9\u03c3\u03c7\u03cd\u03b5\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03c4\u03b7\u03bd \u03c0\u03c1\u03bf\u03b5\u03b3\u03b3\u03c1\u03b1\u03c6\u03ae, \u03c4\u03bf\u03bd \u03b4\u03b7\u03bc\u03cc\u03c3\u03b9\u03bf \u03ba\u03ce\u03b4\u03b9\u03ba\u03b1, \u03c4\u03b1 traces, \u03c4\u03b1 seeds, \u03c4\u03b1 data hashes \u03ba\u03b1\u03b9 \u03c4\u03bf corrections log. \u0391\u03c5\u03c4\u03ae \u03b7 \u03b4\u03b9\u03b1\u03c6\u03ac\u03bd\u03b5\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c4\u03c1\u03ad\u03c0\u03b5\u03b9 \u03b1\u03bd\u03b1\u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae \u03ba\u03b1\u03b9 \u03ad\u03bb\u03b5\u03b3\u03c7\u03bf, \u03b1\u03bb\u03bb\u03ac \u03b4\u03b5\u03bd \u03b1\u03bd\u03c4\u03b9\u03ba\u03b1\u03b8\u03b9\u03c3\u03c4\u03ac \u03c4\u03b7\u03bd \u03b1\u03bd\u03b5\u03be\u03ac\u03c1\u03c4\u03b7\u03c4\u03b7 \u03b5\u03c0\u03b9\u03ba\u03cd\u03c1\u03c9\u03c3\u03b7 \u03c3\u03b5 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03cc \u03b5\u03be\u03bf\u03c0\u03bb\u03b9\u03c3\u03bc\u03cc \u03ba\u03b1\u03b9 workload.<\/p>\n<h2 id=\"cache-meros-systimatos\">\u0397 cache \u03b5\u03af\u03bd\u03b1\u03b9 \u03bc\u03ad\u03c1\u03bf\u03c2 \u03c4\u03b7\u03c2 \u03bb\u03cd\u03c3\u03b7\u03c2, \u03cc\u03c7\u03b9 \u03cc\u03bb\u03b7 \u03b7 \u03bb\u03cd\u03c3\u03b7<\/h2>\n<p>\u0397 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03bc\u03b5\u03c4\u03b1\u03c4\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03b5\u03bb\u03ba\u03c5\u03c3\u03c4\u03b9\u03ba\u03ae \u03c5\u03c0\u03cc\u03c3\u03c7\u03b5\u03c3\u03b7 \u2014 \u00ab\u03ba\u03ac\u03bd\u03b5 \u03c4\u03bf\u03bd router \u03c0\u03b9\u03bf \u03c6\u03b9\u03bb\u03b9\u03ba\u03cc \u03c3\u03c4\u03b7\u03bd cache\u00bb \u2014 \u03c3\u03b5 \u03b5\u03bb\u03ad\u03b3\u03be\u03b9\u03bc\u03b7 \u03c5\u03c0\u03cc\u03b8\u03b5\u03c3\u03b7. \u0397 LRU \u03b1\u03be\u03b9\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b5 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ae temporal locality, \u03b7 cache-aware \u03b5\u03ba\u03c0\u03b1\u03af\u03b4\u03b5\u03c5\u03c3\u03b7 \u03bc\u03b5\u03af\u03c9\u03c3\u03b5 \u03c4\u03b1 misses \u03ba\u03b1\u03b9 \u03c4\u03bf rerouting \u03c0\u03c1\u03cc\u03c3\u03b8\u03b5\u03c3\u03b5 \u03cc\u03c6\u03b5\u03bb\u03bf\u03c2. \u03a0\u03b1\u03c1\u2019 \u03cc\u03bb\u03b1 \u03b1\u03c5\u03c4\u03ac, \u03ba\u03b1\u03bd\u03ad\u03bd\u03b1 \u03bc\u03b5\u03bc\u03bf\u03bd\u03c9\u03bc\u03ad\u03bd\u03bf \u03c4\u03ad\u03c7\u03bd\u03b1\u03c3\u03bc\u03b1 \u03b4\u03b5\u03bd \u03b5\u03be\u03b1\u03c6\u03ac\u03bd\u03b9\u03c3\u03b5 \u03c4\u03b1\u03c5\u03c4\u03cc\u03c7\u03c1\u03bf\u03bd\u03b1 \u03c4\u03bf bandwidth wall \u03ba\u03b1\u03b9 \u03c4\u03bf \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2.<\/p>\n<p>\u0393\u03b9\u03b1 \u03bc\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03af\u03c1\u03b7\u03c3\u03b7, \u03b7 \u03c3\u03c9\u03c3\u03c4\u03ae \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c3\u03c5\u03bd-\u03c3\u03c7\u03b5\u03b4\u03b9\u03b1\u03c3\u03bc\u03cc\u03c2 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf\u03c5, router, cache, memory hierarchy, serving engine \u03ba\u03b1\u03b9 workload. \u03a4\u03bf edge \u03ae on-premise AI \u03b1\u03be\u03af\u03b6\u03b5\u03b9 \u03cc\u03c4\u03b1\u03bd \u03b7 \u03b9\u03b4\u03b9\u03c9\u03c4\u03b9\u03ba\u03cc\u03c4\u03b7\u03c4\u03b1, \u03c4\u03bf latency, \u03bf \u03c3\u03c4\u03b1\u03b8\u03b5\u03c1\u03cc\u03c2 \u03cc\u03b3\u03ba\u03bf\u03c2 \u03ae \u03bf \u03ad\u03bb\u03b5\u03b3\u03c7\u03bf\u03c2 \u03b4\u03b9\u03ba\u03b1\u03b9\u03bf\u03bb\u03bf\u03b3\u03bf\u03cd\u03bd \u03c4\u03b5\u03ba\u03bc\u03b7\u03c1\u03b9\u03c9\u03bc\u03ad\u03bd\u03b1 \u03c4\u03b7\u03bd \u03c5\u03c0\u03bf\u03b4\u03bf\u03bc\u03ae. \u0397 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03b1\u03bd\u03b1\u03c3\u03c4\u03c1\u03ad\u03c8\u03b9\u03bc\u03b7, \u03bc\u03b5 \u03c3\u03b1\u03c6\u03ad\u03c2 baseline \u03ba\u03b1\u03b9 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b5\u03b9\u03c2 \u03c0\u03bf\u03c5 \u03c3\u03c5\u03bd\u03b4\u03ad\u03bf\u03c5\u03bd \u03c4\u03b5\u03c7\u03bd\u03b9\u03ba\u03ae \u03b1\u03c0\u03cc\u03b4\u03bf\u03c3\u03b7 \u03bc\u03b5 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ae \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03c3\u03b9\u03b1\u03ba\u03ae \u03b1\u03be\u03af\u03b1.<\/p>\n<section class=\"td-service-cta\">\n<div class=\"td-service-cta-content\">\n<div class=\"td-service-cta-copy\"><span class=\"td-service-cta-kicker\">AI \u03c5\u03c0\u03bf\u03b4\u03bf\u03bc\u03ae \u03bc\u03b5 workload-first \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b5\u03b9\u03c2<\/span><\/p>\n<h3>\u03a3\u03c7\u03b5\u03b4\u03b9\u03ac\u03c3\u03c4\u03b5 \u03c4\u03bf AI workflow \u03c0\u03c1\u03b9\u03bd \u03b4\u03b5\u03c3\u03bc\u03b5\u03c5\u03c4\u03b5\u03af\u03c4\u03b5 \u03c3\u03b5 hardware<\/h3>\n<p>\u0397 TWO DOTS \u03c7\u03b1\u03c1\u03c4\u03bf\u03b3\u03c1\u03b1\u03c6\u03b5\u03af \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03b1, latency, \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1, \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2, integrations \u03ba\u03b1\u03b9 fallback \u03c0\u03c1\u03b9\u03bd \u03c3\u03c5\u03bd\u03b4\u03ad\u03c3\u03b5\u03b9 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1 \u03bc\u03b5 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ad\u03c2 \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03c3\u03b9\u03b1\u03ba\u03ad\u03c2 \u03c1\u03bf\u03ad\u03c2. \u0388\u03c4\u03c3\u03b9 \u03ad\u03bd\u03b1 edge \u03ae self-hosted pilot \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03b5\u03af\u03c4\u03b1\u03b9 \u03bc\u03b5 \u03c4\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03c0\u03bf\u03c5 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c0\u03b1\u03c1\u03b1\u03b4\u03ce\u03c3\u03b5\u03b9 \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03bc\u03cc\u03bd\u03bf \u03bc\u03b5 \u03c4\u03bf \u03bc\u03ad\u03b3\u03b5\u03b8\u03bf\u03c2 \u03c4\u03bf\u03c5 checkpoint.<\/p>\n<\/div>\n<div class=\"td-service-cta-actions\"><a class=\"td-service-cta-button\" href=\"https:\/\/twodots.gr\/aftomatismoi-epicheiriseon-ai\/\">\u0394\u03b5\u03af\u03c4\u03b5 \u03c4\u03bf\u03c5\u03c2 \u0391\u03c5\u03c4\u03bf\u03bc\u03b1\u03c4\u03b9\u03c3\u03bc\u03bf\u03cd\u03c2 \u0395\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03ae\u03c3\u03b5\u03c9\u03bd &amp; AI<\/a><\/div>\n<\/div>\n<\/section>\n<section id=\"sychnes-erotiseis\" class=\"td-faq-section\">\n<div class=\"td-faq\">\n<p class=\"td-faq-heading\">\u0427\u0435\u0441\u0442\u043e \u0437\u0430\u0434\u0430\u0432\u0430\u043d\u0438 \u0432\u044a\u043f\u0440\u043e\u0441\u0438<\/p>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a4\u03b9 \u03b5\u03af\u03bd\u03b1\u03b9 \u03ad\u03bd\u03b1 Mixture-of-Experts \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u0395\u03af\u03bd\u03b1\u03b9 \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae \u03cc\u03c0\u03bf\u03c5 \u03ad\u03bd\u03b1\u03c2 router \u03c3\u03c4\u03ad\u03bb\u03bd\u03b5\u03b9 \u03ba\u03ac\u03b8\u03b5 token \u03c3\u03b5 \u03bc\u03b9\u03ba\u03c1\u03cc \u03c5\u03c0\u03bf\u03c3\u03cd\u03bd\u03bf\u03bb\u03bf \u03c0\u03bf\u03bb\u03bb\u03ce\u03bd experts. \u0388\u03c4\u03c3\u03b9 \u03c4\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03b1\u03c0\u03bf\u03b8\u03b7\u03ba\u03b5\u03cd\u03b5\u03b9 \u03bc\u03b5\u03b3\u03ac\u03bb\u03b7 \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03ae \u03c7\u03c9\u03c1\u03b7\u03c4\u03b9\u03ba\u03cc\u03c4\u03b7\u03c4\u03b1, \u03b1\u03bb\u03bb\u03ac \u03b5\u03bd\u03b5\u03c1\u03b3\u03bf\u03c0\u03bf\u03b9\u03b5\u03af \u03bc\u03cc\u03bd\u03bf \u03bc\u03ad\u03c1\u03bf\u03c2 \u03c4\u03c9\u03bd \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03c4\u03c1\u03c9\u03bd \u03c3\u03b5 \u03ba\u03ac\u03b8\u03b5 \u03b2\u03ae\u03bc\u03b1.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u0393\u03b9\u03b1\u03c4\u03af \u03ad\u03bd\u03b1 MoE \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c7\u03c9\u03c1\u03ac \u03b1\u03bb\u03bb\u03ac \u03bd\u03b1 \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03b1\u03c1\u03b3\u03cc;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u0395\u03c0\u03b5\u03b9\u03b4\u03ae \u03c4\u03bf \u00ab\u03c7\u03c9\u03c1\u03ac\u00bb \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03bc\u03b5\u03b3\u03ac\u03bb\u03bf \u03bc\u03ad\u03c1\u03bf\u03c2 \u03c4\u03c9\u03bd experts \u03bc\u03ad\u03bd\u03b5\u03b9 \u03c3\u03b5 RAM \u03ae SSD. \u0391\u03bd \u03c4\u03b1 \u03b5\u03bd\u03b5\u03c1\u03b3\u03ac \u03b2\u03ac\u03c1\u03b7 \u03bc\u03b5\u03c4\u03b1\u03c6\u03ad\u03c1\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b3\u03b9\u03b1 \u03ba\u03ac\u03b8\u03b5 token \u03b1\u03c0\u03cc \u03b1\u03c1\u03b3\u03cc tier, \u03c4\u03bf decode \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03c4\u03bf bandwidth \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03b1\u03c0\u03cc \u03c4\u03b7\u03bd \u03c5\u03c0\u03bf\u03bb\u03bf\u03b3\u03b9\u03c3\u03c4\u03b9\u03ba\u03ae \u03b9\u03c3\u03c7\u03cd \u03c4\u03b7\u03c2 GPU.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a0\u03cc\u03c3\u03bf \u03b3\u03c1\u03ae\u03b3\u03bf\u03c1\u03bf \u03ae\u03c4\u03b1\u03bd \u03c4\u03bf Qwen3-235B \u03c3\u03c4\u03b7 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u03a4\u03bf warm single-stream decode \u03bc\u03b5\u03c4\u03c1\u03ae\u03b8\u03b7\u03ba\u03b5 \u03c3\u03c4\u03b1 0,441 token \u03b1\u03bd\u03ac \u03b4\u03b5\u03c5\u03c4\u03b5\u03c1\u03cc\u03bb\u03b5\u03c0\u03c4\u03bf \u03c3\u03c4\u03bf \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf RTX 3070, RAM, NVMe \u03ba\u03b1\u03b9 llama.cpp setup. \u0394\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03b3\u03b5\u03bd\u03b9\u03ba\u03cc benchmark \u03b3\u03b9\u03b1 \u03ba\u03ac\u03b8\u03b5 Qwen3 \u03ae MoE \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a0\u03cc\u03c3\u03bf \u03b2\u03bf\u03ae\u03b8\u03b7\u03c3\u03b5 \u03b7 LRU cache;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u03a3\u03c4\u03b7\u03bd \u03c0\u03c1\u03bf\u03c3\u03bf\u03bc\u03bf\u03af\u03c9\u03c3\u03b7 \u03c4\u03c9\u03bd Qwen3-30B traces, LRU cache \u03bc\u03b5 budget 13,4% \u03c4\u03c9\u03bd experts \u03c0\u03ad\u03c4\u03c5\u03c7\u03b5 hit rate 65,9%, \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 59,2% \u03b3\u03b9\u03b1 static pinning \u03ba\u03b1\u03b9 60,1% \u03b3\u03b9\u03b1 LFU. \u0397 \u03c0\u03c1\u03bf\u03b2\u03bf\u03bb\u03ae \u03c3\u03c4\u03bf \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03c4\u03c9\u03bd 235B \u03c0\u03b1\u03c1\u03ad\u03bc\u03b5\u03bd\u03b5 \u03ba\u03ac\u03c4\u03c9 \u03b1\u03c0\u03cc interactive \u03c4\u03b1\u03c7\u03cd\u03c4\u03b7\u03c4\u03b1.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u0393\u03b9\u03b1\u03c4\u03af \u03c4\u03bf batch 32 \u03ae\u03c4\u03b1\u03bd \u03c0\u03b9\u03bf \u03b1\u03c1\u03b3\u03cc;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u039f\u03b9 \u03c0\u03bf\u03bb\u03bb\u03ad\u03c2 \u03c0\u03b1\u03c1\u03ac\u03bb\u03bb\u03b7\u03bb\u03b5\u03c2 \u03c1\u03bf\u03ad\u03c2 \u03c0\u03c1\u03bf\u03ba\u03ac\u03bb\u03b5\u03c3\u03b1\u03bd page-fault storm \u03ba\u03b1\u03b9 paging thrash, \u03bc\u03b5\u03b9\u03ce\u03bd\u03bf\u03bd\u03c4\u03b1\u03c2 \u03c4\u03bf \u03b1\u03c0\u03bf\u03b4\u03bf\u03c4\u03b9\u03ba\u03cc I\/O. \u03a4\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03b1\u03c6\u03bf\u03c1\u03ac \u03c4\u03b7 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b7 \u03c5\u03bb\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 \u03ba\u03b1\u03b9 \u03b4\u03b5\u03bd \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03ba\u03bd\u03cd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03ba\u03ac\u03b8\u03b5 purpose-built batching engine \u03b8\u03b1 \u03b1\u03c0\u03bf\u03c4\u03cd\u03c7\u03b5\u03b9.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a0\u03ad\u03c1\u03b1\u03c3\u03b5 \u03b7 cache-aware \u03b5\u03ba\u03c0\u03b1\u03af\u03b4\u03b5\u03c5\u03c3\u03b7 \u03c4\u03b1 \u03ba\u03c1\u03b9\u03c4\u03ae\u03c1\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c4\u03c5\u03c7\u03af\u03b1\u03c2;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u038c\u03c7\u03b9. \u039f\u03b9 \u03b9\u03c3\u03c7\u03c5\u03c1\u03ad\u03c2 \u03c1\u03c5\u03b8\u03bc\u03af\u03c3\u03b5\u03b9\u03c2 \u03bc\u03b5\u03af\u03c9\u03c3\u03b1\u03bd \u03c3\u03b7\u03bc\u03b1\u03bd\u03c4\u03b9\u03ba\u03ac \u03c4\u03b1 cache misses, \u03b1\u03bb\u03bb\u03ac \u03be\u03b5\u03c0\u03ad\u03c1\u03b1\u03c3\u03b1\u03bd \u03c4\u03b1 \u03c0\u03c1\u03bf\u03b5\u03b3\u03b3\u03b5\u03b3\u03c1\u03b1\u03bc\u03bc\u03ad\u03bd\u03b1 \u03cc\u03c1\u03b9\u03b1 \u03b1\u03cd\u03be\u03b7\u03c3\u03b7\u03c2 perplexity. \u039f \u03bc\u03b7\u03c7\u03b1\u03bd\u03b9\u03c3\u03bc\u03cc\u03c2 \u03bb\u03b5\u03b9\u03c4\u03bf\u03cd\u03c1\u03b3\u03b7\u03c3\u03b5, \u03cc\u03bc\u03c9\u03c2 \u03b7 \u03b2\u03b5\u03bb\u03c4\u03af\u03c9\u03c3\u03b7 \u03b4\u03b5\u03bd \u03ae\u03c4\u03b1\u03bd \u03b4\u03c9\u03c1\u03b5\u03ac\u03bd \u03c3\u03c4\u03bf \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf multi-domain \u03c0\u03b5\u03af\u03c1\u03b1\u03bc\u03b1.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u039c\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03c4\u03b1 \u03b5\u03c5\u03c1\u03ae\u03bc\u03b1\u03c4\u03b1 \u03bd\u03b1 \u03b5\u03c6\u03b1\u03c1\u03bc\u03bf\u03c3\u03c4\u03bf\u03cd\u03bd \u03b1\u03c0\u03b5\u03c5\u03b8\u03b5\u03af\u03b1\u03c2 \u03c3\u03b5 production;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u038c\u03c7\u03b9 \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03ad\u03b1 \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b7. \u03a7\u03c1\u03b5\u03b9\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 traces, quality tests, latency, concurrency \u03ba\u03b1\u03b9 \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03c3\u03c4\u03bf \u03b4\u03b9\u03ba\u03cc \u03c3\u03b1\u03c2 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf, hardware, engine \u03ba\u03b1\u03b9 \u03bc\u03af\u03b3\u03bc\u03b1 \u03b5\u03c1\u03b3\u03b1\u03c3\u03b9\u03ce\u03bd. \u03a4\u03b1 \u03b4\u03b7\u03bc\u03bf\u03c3\u03b9\u03b5\u03c5\u03bc\u03ad\u03bd\u03b1 \u03bd\u03bf\u03cd\u03bc\u03b5\u03c1\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c6\u03b5\u03c4\u03b7\u03c1\u03af\u03b1 \u03b3\u03b9\u03b1 \u03c5\u03c0\u03cc\u03b8\u03b5\u03c3\u03b7, \u03cc\u03c7\u03b9 \u03b5\u03b3\u03b3\u03cd\u03b7\u03c3\u03b7 \u03b1\u03c0\u03cc\u03b4\u03bf\u03c3\u03b7\u03c2.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a4\u03b9 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03af\u03c1\u03b7\u03c3\u03b7 \u03c0\u03c1\u03b9\u03bd \u03b1\u03c0\u03cc edge MoE pilot;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u03a4\u03bf workload, \u03c4\u03bf \u03b1\u03c0\u03bf\u03b4\u03b5\u03ba\u03c4\u03cc quality threshold, bytes \u03b1\u03bd\u03ac token, bandwidth \u03ba\u03b1\u03b9 \u03c7\u03c9\u03c1\u03b7\u03c4\u03b9\u03ba\u03cc\u03c4\u03b7\u03c4\u03b1 \u03ba\u03ac\u03b8\u03b5 memory tier, routing locality, cache hit rate, single-stream latency, aggregate throughput, \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b1, \u03c3\u03c5\u03bd\u03c4\u03ae\u03c1\u03b7\u03c3\u03b7 \u03ba\u03b1\u03b9 \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03b1\u03bd\u03ac \u03b5\u03c0\u03b9\u03c4\u03c5\u03c7\u03b7\u03bc\u03ad\u03bd\u03b7 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1.<\/p>\n<\/div>\n<\/details>\n<\/div>\n<\/section>\n<div class=\"td-source-list\">\n<p id=\"piges\" class=\"td-source-list-title\">\u0418\u0437\u0442\u043e\u0447\u043d\u0438\u0446\u0438<\/p>\n<ul>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2608.18261\" target=\"_blank\" rel=\"noopener\">Suram \u2014 Cacheable by Design? Training Mixture-of-Experts Routers for Locality Against the Edge Memory-Bandwidth Wall<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/Shriniwas410\/cacheable-by-design\" target=\"_blank\" rel=\"noopener\">Shriniwas410 \u2014 Cacheable by Design: \u03ba\u03ce\u03b4\u03b9\u03ba\u03b1\u03c2, traces, \u03c0\u03c1\u03bf\u03b5\u03b3\u03b3\u03c1\u03b1\u03c6\u03ae \u03ba\u03b1\u03b9 corrections log<\/a><\/li>\n<li><a href=\"https:\/\/qwenlm.github.io\/blog\/qwen3\/\" target=\"_blank\" rel=\"noopener\">Qwen \u2014 Qwen3: Think Deeper, Act Faster<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2607.08780\" target=\"_blank\" rel=\"noopener\">Kayyam \u2014 Sticky Routing: Training MoE Models for Memory-Efficient Inference<\/a><\/li>\n<\/ul>\n<\/div>","protected":false},"excerpt":{"rendered":"<p>\u03a4\u03b1 MoE \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1 \u03c3\u03c4\u03b7\u03bd \u03ac\u03ba\u03c1\u03b7 \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03af\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03c4\u03bf bandwidth. \u0394\u03b5\u03af\u03c4\u03b5 \u03c4\u03b9 \u03ad\u03b4\u03b5\u03b9\u03be\u03b1\u03bd Qwen3, expert cache \u03ba\u03b1\u03b9 cache-aware routing \u03c3\u03b5 \u03c3\u03cd\u03b3\u03c7\u03c1\u03bf\u03bd\u03bf consumer hardware.<\/p>","protected":false},"author":1,"featured_media":89360,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"_gspb_post_css":"","content-type":"","footnotes":""},"categories":[199],"tags":[8988,7203,19427,7477,19426],"class_list":["post-89305","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-techniti-noimosyni","tag-ai-infrastructure","tag-edge-ai","tag-llm-inference","tag-machine-learning","tag-moe"],"blocksy_meta":{"styles_descriptor":{"styles":{"desktop":"","tablet":"","mobile":""},"google_fonts":[],"version":7}},"_links":{"self":[{"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/posts\/89305","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/comments?post=89305"}],"version-history":[{"count":1,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/posts\/89305\/revisions"}],"predecessor-version":[{"id":89361,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/posts\/89305\/revisions\/89361"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/media\/89360"}],"wp:attachment":[{"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/media?parent=89305"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/categories?post=89305"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/tags?post=89305"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}