{"id":87096,"date":"2026-07-29T12:01:09","date_gmt":"2026-07-29T09:01:09","guid":{"rendered":"https:\/\/twodots.gr\/?p=87096"},"modified":"2026-07-29T12:01:10","modified_gmt":"2026-07-29T09:01:10","slug":"jaxbench-ai-veltistopoiisi-kernels-tpu","status":"publish","type":"post","link":"https:\/\/twodots.gr\/en\/jaxbench-ai-veltistopoiisi-kernels-tpu\/","title":{"rendered":"JAXBench: \u03c0\u03ce\u03c2 \u03b7 AI \u03bc\u03b1\u03b8\u03b1\u03af\u03bd\u03b5\u03b9 \u03bd\u03b1 \u03b2\u03b5\u03bb\u03c4\u03b9\u03c3\u03c4\u03bf\u03c0\u03bf\u03b9\u03b5\u03af kernels \u03b3\u03b9\u03b1 TPU"},"content":{"rendered":"<div class=\"td-article-lede\">\n<p><strong>\u03a4\u03bf JAXBench \u03b5\u03af\u03bd\u03b1\u03b9 \u03ad\u03bd\u03b1 TPU-native benchmark \u03c0\u03bf\u03c5 \u03b5\u03be\u03b5\u03c4\u03ac\u03b6\u03b5\u03b9 \u03b1\u03bd AI agents \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03bd\u03b1 \u03c0\u03b1\u03c1\u03ac\u03b3\u03bf\u03c5\u03bd Pallas kernels \u03bf\u03b9 \u03bf\u03c0\u03bf\u03af\u03bf\u03b9 \u03cc\u03c7\u03b9 \u03bc\u03cc\u03bd\u03bf \u03bc\u03b5\u03c4\u03b1\u03b3\u03bb\u03c9\u03c4\u03c4\u03af\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9, \u03b1\u03bb\u03bb\u03ac \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c1\u03b9\u03b8\u03bc\u03b7\u03c4\u03b9\u03ba\u03ac \u03c3\u03c9\u03c3\u03c4\u03bf\u03af \u03ba\u03b1\u03b9 \u03c4\u03b1\u03c7\u03cd\u03c4\u03b5\u03c1\u03bf\u03b9 \u03b1\u03c0\u03cc \u03c4\u03bf XLA.<\/strong> \u0397 \u03c3\u03bf\u03c5\u03af\u03c4\u03b1 \u03ba\u03b1\u03bb\u03cd\u03c0\u03c4\u03b5\u03b9 50 JAX workloads \u03c3\u03b5 TPU v6e \u03ba\u03b1\u03b9 \u03c3\u03c5\u03bd\u03b4\u03ad\u03b5\u03b9 \u03ba\u03ac\u03b8\u03b5 \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03bc\u03b5 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf compilation, correctness \u03ba\u03b1\u03b9 performance gate.<\/p>\n<p>\u03a4\u03bf \u03b2\u03b1\u03c3\u03b9\u03ba\u03cc \u03b5\u03cd\u03c1\u03b7\u03bc\u03b1 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03cc\u03c4\u03b9 \u03ad\u03bd\u03b1 \u03bc\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03bb\u03cd\u03bd\u03b5\u03b9 \u03b1\u03c0\u03cc \u03bc\u03cc\u03bd\u03bf \u03c4\u03bf\u03c5 \u03c4\u03bf \u03c0\u03c1\u03cc\u03b2\u03bb\u03b7\u03bc\u03b1. \u0397 \u03b5\u03c0\u03b9\u03bc\u03b5\u03bb\u03b7\u03bc\u03ad\u03bd\u03b7 \u03b3\u03bd\u03ce\u03c3\u03b7 \u03b3\u03b9\u03b1 Pallas \u03ba\u03b1\u03b9 TPU \u03b1\u03cd\u03be\u03b7\u03c3\u03b5 \u03c4\u03b7\u03bd \u03bf\u03c1\u03b8\u03cc\u03c4\u03b7\u03c4\u03b1 \u03b1\u03bd\u03ac sample \u03b1\u03c0\u03cc 5,8% \u03c3\u03b5 37,3%, \u03b5\u03bd\u03ce \u03b7 \u03bf\u03c1\u03b3\u03b1\u03bd\u03c9\u03bc\u03ad\u03bd\u03b7 \u03b1\u03bd\u03b1\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7 \u03c4\u03bf\u03c5 Autocomp \u03ad\u03c6\u03c4\u03b1\u03c3\u03b5 \u03b3\u03b5\u03c9\u03bc\u03b5\u03c4\u03c1\u03b9\u03ba\u03cc \u03bc\u03ad\u03c3\u03bf speedup 1,36\u00d7 \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 XLA.<\/p>\n<\/div>\n<div class=\"td-article-note\">\n<p><strong>Short answer:<\/strong> \u03c4\u03bf JAXBench \u03b4\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03b7 \u03b1\u03c5\u03c4\u03cc\u03bc\u03b1\u03c4\u03b7 \u03b2\u03b5\u03bb\u03c4\u03b9\u03c3\u03c4\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 TPU kernels \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c0\u03bb\u03ae \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae \u03ba\u03ce\u03b4\u03b9\u03ba\u03b1. \u039f \u03c5\u03c0\u03bf\u03c8\u03ae\u03c6\u03b9\u03bf\u03c2 kernel \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03bc\u03b5\u03c4\u03b1\u03b3\u03bb\u03c9\u03c4\u03c4\u03af\u03b6\u03b5\u03c4\u03b1\u03b9, \u03bd\u03b1 \u03c3\u03c5\u03bc\u03c6\u03c9\u03bd\u03b5\u03af \u03b1\u03c1\u03b9\u03b8\u03bc\u03b7\u03c4\u03b9\u03ba\u03ac \u03bc\u03b5 \u03c4\u03bf reference \u03ba\u03b1\u03b9 \u03bd\u03b1 \u03bc\u03b5\u03c4\u03c1\u03b9\u03ad\u03c4\u03b1\u03b9 \u03c3\u03c4\u03bf \u03af\u03b4\u03b9\u03bf accelerator \u03b1\u03c0\u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 \u03c3\u03b5 \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc XLA baseline.<\/p>\n<\/div>\n<div class=\"td-article-toc\">\n<div class=\"td-toc-title\">Contents<\/div>\n<ul>\n<li><a href=\"#ti-metra-jaxbench\">\u03a4\u03b9 \u03bc\u03b5\u03c4\u03c1\u03ac \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac \u03c4\u03bf JAXBench<\/a><\/li>\n<li><a href=\"#veltistopoiisi-tpu-diafora-gpu\">\u0393\u03b9\u03b1\u03c4\u03af \u03b7 \u03b2\u03b5\u03bb\u03c4\u03b9\u03c3\u03c4\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 TPU \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03bc\u03b5\u03c4\u03b1\u03c6\u03bf\u03c1\u03ac \u03c4\u03b5\u03c7\u03bd\u03bf\u03b3\u03bd\u03c9\u03c3\u03af\u03b1\u03c2 \u03b1\u03c0\u03cc GPU<\/a><\/li>\n<li><a href=\"#workloads-expert-baselines\">\u03a4\u03b1 workloads \u03ba\u03b1\u03b9 \u03c4\u03b1 expert baselines<\/a><\/li>\n<li><a href=\"#correctness-epanalipsimotita\">\u03a0\u03ce\u03c2 \u03b5\u03be\u03b1\u03c3\u03c6\u03b1\u03bb\u03af\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 correctness \u03ba\u03b1\u03b9 \u03b5\u03c0\u03b1\u03bd\u03b1\u03bb\u03b7\u03c8\u03b9\u03bc\u03cc\u03c4\u03b7\u03c4\u03b1<\/a><\/li>\n<li><a href=\"#tesseris-proseggiseis\">\u039f\u03b9 \u03c4\u03ad\u03c3\u03c3\u03b5\u03c1\u03b9\u03c2 \u03c0\u03c1\u03bf\u03c3\u03b5\u03b3\u03b3\u03af\u03c3\u03b5\u03b9\u03c2 \u03c0\u03bf\u03c5 \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03ae\u03b8\u03b7\u03ba\u03b1\u03bd<\/a><\/li>\n<li><a href=\"#context-allazei-orthotita\">\u03a4\u03bf \u03b2\u03b1\u03c3\u03b9\u03ba\u03cc \u03b5\u03cd\u03c1\u03b7\u03bc\u03b1: \u03c4\u03bf \u03b5\u03be\u03b5\u03b9\u03b4\u03b9\u03ba\u03b5\u03c5\u03bc\u03ad\u03bd\u03bf context \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd \u03bf\u03c1\u03b8\u03cc\u03c4\u03b7\u03c4\u03b1<\/a><\/li>\n<li><a href=\"#agents-anthropinoi-kernels\">\u03a0\u03cc\u03c3\u03bf \u03ba\u03bf\u03bd\u03c4\u03ac \u03ad\u03c6\u03c4\u03b1\u03c3\u03b1\u03bd \u03bf\u03b9 agents \u03c3\u03c4\u03bf\u03c5\u03c2 \u03b1\u03bd\u03b8\u03c1\u03ce\u03c0\u03b9\u03bd\u03bf\u03c5\u03c2 kernels<\/a><\/li>\n<li><a href=\"#modelo-gnosi-stochou\">\u039c\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03ae \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b7 \u03b3\u03bd\u03ce\u03c3\u03b7 \u03c3\u03c4\u03cc\u03c7\u03bf\u03c5;<\/a><\/li>\n<li><a href=\"#ai-ypodomes-epicheiriseis\">\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 \u03c0\u03bf\u03c5 \u03c7\u03c4\u03af\u03b6\u03bf\u03c5\u03bd AI \u03c5\u03c0\u03bf\u03b4\u03bf\u03bc\u03ad\u03c2<\/a><\/li>\n<li><a href=\"#oria-meletis\">\u038c\u03c1\u03b9\u03b1 \u03c4\u03b7\u03c2 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7\u03c2 \u03ba\u03b1\u03b9 \u03b1\u03c3\u03c6\u03b1\u03bb\u03ae\u03c2 \u03b1\u03bd\u03ac\u03b3\u03bd\u03c9\u03c3\u03b7 \u03c4\u03c9\u03bd \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03b5\u03c3\u03bc\u03ac\u03c4\u03c9\u03bd<\/a><\/li>\n<li><a href=\"#plaisio-agentic-optimization\">\u0388\u03bd\u03b1 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b9\u03ba\u03cc \u03c0\u03bb\u03b1\u03af\u03c3\u03b9\u03bf \u03b3\u03b9\u03b1 agentic optimization<\/a><\/li>\n<\/ul>\n<\/div>\n<p>\u0397 \u03b1\u03c5\u03c4\u03cc\u03bc\u03b1\u03c4\u03b7 \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae \u03ba\u03ce\u03b4\u03b9\u03ba\u03b1 \u03b1\u03c0\u03cc \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1 \u03c4\u03b5\u03c7\u03bd\u03b7\u03c4\u03ae\u03c2 \u03bd\u03bf\u03b7\u03bc\u03bf\u03c3\u03cd\u03bd\u03b7\u03c2 \u03ad\u03c7\u03b5\u03b9 \u03c0\u03c1\u03bf\u03c7\u03c9\u03c1\u03ae\u03c3\u03b5\u03b9 \u03b5\u03bd\u03c4\u03c5\u03c0\u03c9\u03c3\u03b9\u03b1\u03ba\u03ac, \u03cc\u03bc\u03c9\u03c2 \u03b7 \u03b4\u03b7\u03bc\u03b9\u03bf\u03c5\u03c1\u03b3\u03af\u03b1 \u03b5\u03bd\u03cc\u03c2 \u03c3\u03c9\u03c3\u03c4\u03bf\u03cd \u03ba\u03b1\u03b9 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac \u03b3\u03c1\u03ae\u03b3\u03bf\u03c1\u03bf\u03c5 kernel \u03b3\u03b9\u03b1 \u03b5\u03be\u03b5\u03b9\u03b4\u03b9\u03ba\u03b5\u03c5\u03bc\u03ad\u03bd\u03bf hardware \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ae \u03ba\u03b1\u03c4\u03b7\u03b3\u03bf\u03c1\u03af\u03b1 \u03c0\u03c1\u03bf\u03b2\u03bb\u03ae\u03bc\u03b1\u03c4\u03bf\u03c2. \u03a4\u03bf JAXBench \u03c0\u03c1\u03bf\u03c4\u03b5\u03af\u03bd\u03b5\u03b9 \u03ad\u03bd\u03b1 \u03b1\u03bd\u03bf\u03b9\u03ba\u03c4\u03cc, TPU-native benchmark \u03c0\u03bf\u03c5 \u03b5\u03be\u03b5\u03c4\u03ac\u03b6\u03b5\u03b9 \u03b1\u03ba\u03c1\u03b9\u03b2\u03ce\u03c2 \u03b1\u03c5\u03c4\u03ae \u03c4\u03b7 \u03b4\u03c5\u03c3\u03ba\u03bf\u03bb\u03af\u03b1: \u03b1\u03bd \u03ad\u03bd\u03b1\u03c2 AI agent \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03cc\u03c7\u03b9 \u03bc\u03cc\u03bd\u03bf \u03bd\u03b1 \u03b3\u03c1\u03ac\u03c8\u03b5\u03b9 \u03ba\u03ce\u03b4\u03b9\u03ba\u03b1 Pallas \u03c0\u03bf\u03c5 \u03bc\u03b5\u03c4\u03b1\u03b3\u03bb\u03c9\u03c4\u03c4\u03af\u03b6\u03b5\u03c4\u03b1\u03b9, \u03b1\u03bb\u03bb\u03ac \u03ba\u03b1\u03b9 \u03bd\u03b1 \u03c0\u03b1\u03c1\u03ac\u03b3\u03b5\u03b9 \u03b1\u03c1\u03b9\u03b8\u03bc\u03b7\u03c4\u03b9\u03ba\u03ac \u03c3\u03c9\u03c3\u03c4\u03ac \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1 \u03ba\u03b1\u03b9 \u03bd\u03b1 \u03be\u03b5\u03c0\u03b5\u03c1\u03bd\u03ac \u03c3\u03b5 \u03c4\u03b1\u03c7\u03cd\u03c4\u03b7\u03c4\u03b1 \u03c4\u03bf \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc baseline \u03c4\u03bf\u03c5 XLA.<\/p>\n<p>\u0393\u03b9\u03b1 \u03b5\u03c0\u03b1\u03b3\u03b3\u03b5\u03bb\u03bc\u03b1\u03c4\u03af\u03b5\u03c2 \u03c0\u03bf\u03c5 \u03c3\u03c7\u03b5\u03b4\u03b9\u03ac\u03b6\u03bf\u03c5\u03bd AI \u03c0\u03c1\u03bf\u03ca\u03cc\u03bd\u03c4\u03b1, inference \u03c5\u03c0\u03b7\u03c1\u03b5\u03c3\u03af\u03b5\u03c2 \u03ae \u03c5\u03c0\u03bf\u03b4\u03bf\u03bc\u03ad\u03c2 \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03c9\u03bd, \u03b7 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03ad\u03c7\u03b5\u03b9 \u03b9\u03b4\u03b9\u03b1\u03af\u03c4\u03b5\u03c1\u03b7 \u03b1\u03be\u03af\u03b1. \u0394\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03b7 \u03b1\u03c0\u03cc\u03b4\u03bf\u03c3\u03b7 \u03b4\u03b5\u03bd \u03b5\u03be\u03b1\u03c1\u03c4\u03ac\u03c4\u03b1\u03b9 \u03b1\u03c0\u03bb\u03ce\u03c2 \u03b1\u03c0\u03cc \u03c4\u03bf \u03bc\u03ad\u03b3\u03b5\u03b8\u03bf\u03c2 \u03c4\u03bf\u03c5 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf\u03c5. \u0397 \u03b5\u03be\u03b5\u03b9\u03b4\u03b9\u03ba\u03b5\u03c5\u03bc\u03ad\u03bd\u03b7 \u03c4\u03b5\u03ba\u03bc\u03b7\u03c1\u03af\u03c9\u03c3\u03b7 \u03c4\u03bf\u03c5 \u03c3\u03c4\u03cc\u03c7\u03bf\u03c5, \u03b7 \u03c3\u03c9\u03c3\u03c4\u03ae \u03b4\u03bf\u03bc\u03ae \u03b1\u03bd\u03b1\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7\u03c2 \u03ba\u03b1\u03b9 \u03b7 \u03b1\u03c5\u03c3\u03c4\u03b7\u03c1\u03ae \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b7 \u03c3\u03c4\u03bf \u03af\u03b4\u03b9\u03bf \u03c4\u03bf accelerator \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03bd\u03b1 \u03b1\u03bb\u03bb\u03ac\u03be\u03bf\u03c5\u03bd \u03bf\u03c5\u03c3\u03b9\u03b1\u03c3\u03c4\u03b9\u03ba\u03ac \u03c4\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1.<\/p>\n<h2 id=\"ti-metra-jaxbench\">\u03a4\u03b9 \u03bc\u03b5\u03c4\u03c1\u03ac \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac \u03c4\u03bf JAXBench<\/h2>\n<p>\u03a4\u03bf JAXBench \u03c0\u03b5\u03c1\u03b9\u03bb\u03b1\u03bc\u03b2\u03ac\u03bd\u03b5\u03b9 50 workloads \u03c3\u03b5 JAX, \u03b5\u03c0\u03b9\u03bb\u03b5\u03b3\u03bc\u03ad\u03bd\u03b1 \u03ce\u03c3\u03c4\u03b5 \u03bd\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c3\u03c5\u03bd\u03b1\u03c6\u03ae \u03bc\u03b5 \u03c3\u03cd\u03b3\u03c7\u03c1\u03bf\u03bd\u03b5\u03c2 \u03b5\u03c6\u03b1\u03c1\u03bc\u03bf\u03b3\u03ad\u03c2 \u03ba\u03b1\u03b9 \u03bd\u03b1 \u03b1\u03c6\u03ae\u03bd\u03bf\u03c5\u03bd \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03c0\u03b5\u03c1\u03b9\u03b8\u03ce\u03c1\u03b9\u03bf \u03b2\u03b5\u03bb\u03c4\u03b9\u03c3\u03c4\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7\u03c2. \u03a4\u03b1 17 \u03c0\u03c1\u03bf\u03ad\u03c1\u03c7\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc production operators \u03c4\u03b7\u03c2 \u03b4\u03b7\u03bc\u03cc\u03c3\u03b9\u03b1\u03c2 \u03b2\u03b9\u03b2\u03bb\u03b9\u03bf\u03b8\u03ae\u03ba\u03b7\u03c2 MaxText \u03ba\u03b1\u03b9 \u03ba\u03b1\u03bb\u03cd\u03c0\u03c4\u03bf\u03c5\u03bd \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ad\u03c2 \u03cc\u03c0\u03c9\u03c2 Llama-3.1, DeepSeek-V3, Mixtral, Mamba-2 \u03ba\u03b1\u03b9 AlphaFold2. \u03a4\u03b1 \u03ac\u03bb\u03bb\u03b1 33 \u03b5\u03af\u03bd\u03b1\u03b9 fused operators \u03c4\u03bf\u03c5 KernelBench Level 2, \u03bc\u03b5\u03c4\u03b1\u03c6\u03c1\u03b1\u03c3\u03bc\u03ad\u03bd\u03b1 \u03b1\u03c0\u03cc PyTorch \u03c3\u03b5 \u03b9\u03c3\u03bf\u03b4\u03cd\u03bd\u03b1\u03bc\u03b5\u03c2 \u03c5\u03bb\u03bf\u03c0\u03bf\u03b9\u03ae\u03c3\u03b5\u03b9\u03c2 JAX.<\/p>\n<p>\u0397 \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ae \u03b1\u03c5\u03c4\u03ae \u03b1\u03c0\u03bf\u03c6\u03b5\u03cd\u03b3\u03b5\u03b9 \u03ad\u03bd\u03b1 \u03c3\u03c5\u03bd\u03b7\u03b8\u03b9\u03c3\u03bc\u03ad\u03bd\u03bf \u03c0\u03c1\u03cc\u03b2\u03bb\u03b7\u03bc\u03b1 \u03c4\u03c9\u03bd benchmarks: \u03bc\u03b9\u03ba\u03c1\u03ac \u03ae \u03c4\u03b5\u03c7\u03bd\u03b7\u03c4\u03ac workloads \u03c3\u03c4\u03b1 \u03bf\u03c0\u03bf\u03af\u03b1 \u03bf \u03c7\u03c1\u03cc\u03bd\u03bf\u03c2 \u03b5\u03ba\u03ba\u03af\u03bd\u03b7\u03c3\u03b7\u03c2 \u03ba\u03b1\u03b9 \u03b7 \u03bc\u03b5\u03c4\u03b1\u03c6\u03bf\u03c1\u03ac \u03bc\u03bd\u03ae\u03bc\u03b7\u03c2 \u03ba\u03c1\u03cd\u03b2\u03bf\u03c5\u03bd \u03c4\u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03c4\u03bf\u03c5 \u03c5\u03c0\u03bf\u03bb\u03bf\u03b3\u03b9\u03c3\u03bc\u03bf\u03cd. \u039f\u03b9 \u03b5\u03c1\u03b5\u03c5\u03bd\u03b7\u03c4\u03ad\u03c2 \u03c1\u03cd\u03b8\u03bc\u03b9\u03c3\u03b1\u03bd \u03be\u03b5\u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03ac \u03c4\u03b9\u03c2 \u03b4\u03b9\u03b1\u03c3\u03c4\u03ac\u03c3\u03b5\u03b9\u03c2 \u03ba\u03ac\u03b8\u03b5 workload \u03ce\u03c3\u03c4\u03b5 \u03c4\u03b1 compute-heavy \u03c0\u03c1\u03bf\u03b2\u03bb\u03ae\u03bc\u03b1\u03c4\u03b1 \u03bd\u03b1 \u03b1\u03be\u03b9\u03bf\u03c0\u03bf\u03b9\u03bf\u03cd\u03bd \u03ad\u03bd\u03c4\u03bf\u03bd\u03b1 \u03c4\u03b9\u03c2 Matrix Multiply Units \u03c4\u03b7\u03c2 TPU v6e. \u0388\u03c4\u03c3\u03b9, \u03ad\u03bd\u03b1 speedup \u03ad\u03c7\u03b5\u03b9 \u03bc\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b7 \u03c0\u03b9\u03b8\u03b1\u03bd\u03cc\u03c4\u03b7\u03c4\u03b1 \u03bd\u03b1 \u03c0\u03c1\u03bf\u03ad\u03c1\u03c7\u03b5\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03bf\u03c5\u03c3\u03b9\u03b1\u03c3\u03c4\u03b9\u03ba\u03ae \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ae tiling, pipeline \u03ae layout \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03b1\u03c0\u03cc \u03b8\u03cc\u03c1\u03c5\u03b2\u03bf \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b7\u03c2.<\/p>\n<h2 id=\"veltistopoiisi-tpu-diafora-gpu\">\u0393\u03b9\u03b1\u03c4\u03af \u03b7 \u03b2\u03b5\u03bb\u03c4\u03b9\u03c3\u03c4\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 TPU \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03bc\u03b5\u03c4\u03b1\u03c6\u03bf\u03c1\u03ac \u03c4\u03b5\u03c7\u03bd\u03bf\u03b3\u03bd\u03c9\u03c3\u03af\u03b1\u03c2 \u03b1\u03c0\u03cc GPU<\/h2>\n<p>\u039f\u03b9 TPU \u03b4\u03b9\u03b1\u03c6\u03ad\u03c1\u03bf\u03c5\u03bd \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ac \u03b1\u03c0\u03cc \u03c4\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03bc\u03b1\u03b6\u03b9\u03ba\u03bf\u03cd SIMT \u03c0\u03b1\u03c1\u03b1\u03bb\u03bb\u03b7\u03bb\u03b9\u03c3\u03bc\u03bf\u03cd \u03c4\u03c9\u03bd GPU. \u03a3\u03c4\u03b7\u03bd TPU v6e \u03c5\u03c0\u03ac\u03c1\u03c7\u03bf\u03c5\u03bd \u03b5\u03c5\u03c1\u03b5\u03af\u03c2 SIMD vector registers \u03ba\u03b1\u03b9 systolic MXUs 256\u00d7256, \u03b5\u03bd\u03ce \u03bf \u03c0\u03c1\u03bf\u03b3\u03c1\u03b1\u03bc\u03bc\u03b1\u03c4\u03b9\u03c3\u03bc\u03cc\u03c2 \u03c7\u03b1\u03bc\u03b7\u03bb\u03bf\u03cd \u03b5\u03c0\u03b9\u03c0\u03ad\u03b4\u03bf\u03c5 \u03c0\u03b5\u03c1\u03bd\u03ac \u03b1\u03c0\u03cc \u03c4\u03bf Pallas \u03ba\u03b1\u03b9 \u03c4\u03bf Mosaic backend. \u039f kernel author \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03bb\u03ac\u03b2\u03b5\u03b9 \u03c5\u03c0\u03cc\u03c8\u03b7 \u03c4\u03b9\u03c2 \u03b9\u03b5\u03c1\u03b1\u03c1\u03c7\u03af\u03b5\u03c2 VMEM, SMEM \u03ba\u03b1\u03b9 HBM, \u03c4\u03bf software pipelining, \u03c4\u03b1 block shapes \u03ba\u03b1\u03b9 \u03c4\u03b7 \u03bb\u03b5\u03be\u03b9\u03ba\u03bf\u03b3\u03c1\u03b1\u03c6\u03b9\u03ba\u03ae \u03c3\u03b5\u03b9\u03c1\u03ac \u03b4\u03b9\u03ac\u03c3\u03c7\u03b9\u03c3\u03b7\u03c2 \u03c4\u03bf\u03c5 grid.<\/p>\n<div class=\"td-comparison\">\n<p class=\"td-comparison-title\">\u03a4\u03ad\u03c3\u03c3\u03b5\u03c1\u03b9\u03c2 \u03c1\u03cc\u03bb\u03bf\u03b9 \u03c3\u03c4\u03b7 \u03b2\u03b5\u03bb\u03c4\u03b9\u03c3\u03c4\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 TPU kernels<\/p>\n<div class=\"td-comparison-cards td-comparison-cards--horizontal\">\n<div class=\"td-comparison-grid td-comparison-grid--two\">\n<div class=\"td-platform-card\">\n<h3>XLA baseline<\/h3>\n<p>\u0394\u03af\u03bd\u03b5\u03b9 \u03c4\u03bf \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc, \u03b1\u03bd\u03b1\u03c0\u03b1\u03c1\u03b1\u03b3\u03ce\u03b3\u03b9\u03bc\u03bf \u03c3\u03b7\u03bc\u03b5\u03af\u03bf \u03b1\u03bd\u03b1\u03c6\u03bf\u03c1\u03ac\u03c2 \u03c3\u03c4\u03bf \u03af\u03b4\u03b9\u03bf TPU \u03ba\u03b1\u03b9 \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03c4\u03bf \u03b1\u03c3\u03c6\u03b1\u03bb\u03ad\u03c2 fallback \u03cc\u03c4\u03b1\u03bd \u03bf \u03bd\u03ad\u03bf\u03c2 kernel \u03b4\u03b5\u03bd \u03ba\u03b5\u03c1\u03b4\u03af\u03b6\u03b5\u03b9.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">Baseline<\/span><span class=\"td-badge\">Fallback<\/span><\/div>\n<\/div>\n<div class=\"td-platform-card td-platform-card--navy\">\n<h3>Pallas kernel<\/h3>\n<p>\u0395\u03ba\u03b8\u03ad\u03c4\u03b5\u03b9 \u03c7\u03b1\u03bc\u03b7\u03bb\u03bf\u03cd \u03b5\u03c0\u03b9\u03c0\u03ad\u03b4\u03bf\u03c5 \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ad\u03c2 \u03b3\u03b9\u03b1 memory spaces, tiling, pipelining \u03ba\u03b1\u03b9 block shapes \u03c0\u03bf\u03c5 \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03bd\u03b1 \u03be\u03b5\u03c0\u03b5\u03c1\u03ac\u03c3\u03bf\u03c5\u03bd \u03c4\u03bf XLA.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">TPU-specific<\/span><span class=\"td-badge\">Mosaic<\/span><\/div>\n<\/div>\n<div class=\"td-platform-card\">\n<h3>AI optimization agent<\/h3>\n<p>\u03a0\u03b1\u03c1\u03ac\u03b3\u03b5\u03b9, \u03b4\u03b9\u03bf\u03c1\u03b8\u03ce\u03bd\u03b5\u03b9 \u03ba\u03b1\u03b9 \u03bc\u03b5\u03c4\u03c1\u03ac \u03c5\u03c0\u03bf\u03c8\u03b7\u03c6\u03af\u03bf\u03c5\u03c2 kernels, \u03b1\u03bb\u03bb\u03ac \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03ad\u03b3\u03ba\u03c5\u03c1\u03bf context, correctness gate \u03ba\u03b1\u03b9 device-side profiling.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">Search<\/span><span class=\"td-badge\">Verification<\/span><\/div>\n<\/div>\n<div class=\"td-platform-card td-platform-card--navy\">\n<h3>Selection policy<\/h3>\n<p>\u03a0\u03c1\u03bf\u03c9\u03b8\u03b5\u03af \u03bc\u03cc\u03bd\u03bf \u03c4\u03bf\u03bd \u03c3\u03c9\u03c3\u03c4\u03cc \u03ba\u03b1\u03b9 \u03c4\u03b1\u03c7\u03cd\u03c4\u03b5\u03c1\u03bf kernel, \u03ba\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c6\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7 \u03ba\u03b1\u03b9 \u03b5\u03c0\u03b9\u03c3\u03c4\u03c1\u03ad\u03c6\u03b5\u03b9 \u03c3\u03c4\u03bf XLA \u03cc\u03c4\u03b1\u03bd \u03c4\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03c0\u03b1\u03bb\u03b9\u03bd\u03b4\u03c1\u03bf\u03bc\u03b5\u03af.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">Promotion gate<\/span><span class=\"td-badge\">Rollback<\/span><\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<p>\u0391\u03c5\u03c4\u03cc \u03ad\u03c7\u03b5\u03b9 \u03ac\u03bc\u03b5\u03c3\u03b7 \u03c3\u03c5\u03bd\u03ad\u03c0\u03b5\u03b9\u03b1 \u03b3\u03b9\u03b1 \u03c4\u03b1 \u03b3\u03bb\u03c9\u03c3\u03c3\u03b9\u03ba\u03ac \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1. \u03a4\u03bf Pallas \u03b5\u03bc\u03c6\u03b1\u03bd\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c0\u03bf\u03bb\u03cd \u03bb\u03b9\u03b3\u03cc\u03c4\u03b5\u03c1\u03bf \u03c3\u03c5\u03c7\u03bd\u03ac \u03c3\u03c4\u03b1 training corpora \u03b1\u03c0\u03cc \u03c4\u03bf CUDA \u03ae \u03c4\u03bf Triton. \u0388\u03bd\u03b1 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b3\u03b5\u03bd\u03b9\u03ba\u03ac \u03b9\u03ba\u03b1\u03bd\u03cc \u03c3\u03c4\u03bf\u03bd \u03ba\u03ce\u03b4\u03b9\u03ba\u03b1 \u03b1\u03bb\u03bb\u03ac \u03bd\u03b1 \u03b5\u03c0\u03b9\u03bd\u03bf\u03b5\u03af \u03b1\u03bd\u03cd\u03c0\u03b1\u03c1\u03ba\u03c4\u03b1 APIs, \u03bd\u03b1 \u03b4\u03b7\u03bb\u03ce\u03bd\u03b5\u03b9 \u03bb\u03ac\u03b8\u03bf\u03c2 memory spaces \u03ae \u03bd\u03b1 \u03c0\u03b1\u03c1\u03b1\u03b2\u03b9\u03ac\u03b6\u03b5\u03b9 \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03b9\u03c3\u03bc\u03bf\u03cd\u03c2 tiling. \u03a4\u03b1 \u03c3\u03c6\u03ac\u03bb\u03bc\u03b1\u03c4\u03b1 \u03b1\u03c5\u03c4\u03ac \u03b4\u03b5\u03bd \u03b4\u03b9\u03bf\u03c1\u03b8\u03ce\u03bd\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c0\u03ac\u03bd\u03c4\u03b1 \u03bc\u03b5 \u03ad\u03bd\u03b1 \u03b3\u03b5\u03bd\u03b9\u03ba\u03cc \u03bc\u03ae\u03bd\u03c5\u03bc\u03b1 compiler, \u03b5\u03c0\u03b5\u03b9\u03b4\u03ae \u03b7 \u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03b7 \u03b3\u03bd\u03ce\u03c3\u03b7 \u03b3\u03b9\u03b1 \u03c4\u03bf hardware \u03b4\u03b5\u03bd \u03c0\u03b5\u03c1\u03b9\u03bb\u03b1\u03bc\u03b2\u03ac\u03bd\u03b5\u03c4\u03b1\u03b9 \u03b1\u03bd\u03b1\u03b3\u03ba\u03b1\u03c3\u03c4\u03b9\u03ba\u03ac \u03c3\u03c4\u03bf feedback.<\/p>\n<h2 id=\"workloads-expert-baselines\">\u03a4\u03b1 workloads \u03ba\u03b1\u03b9 \u03c4\u03b1 expert baselines<\/h2>\n<p>\u03a4\u03b1 17 production workloads \u03ba\u03b1\u03bb\u03cd\u03c0\u03c4\u03bf\u03c5\u03bd flash, grouped-query, multi-head latent, sparse, flex, paged \u03ba\u03b1\u03b9 ragged paged attention, \u03ba\u03b1\u03b8\u03ce\u03c2 \u03ba\u03b1\u03b9 GEMM, SwiGLU MLP, sparse mixture-of-experts, Megablox GMM \u03ba\u03b1\u03b9 ragged dot. \u03a0\u03b5\u03c1\u03b9\u03bb\u03b1\u03bc\u03b2\u03ac\u03bd\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b5\u03c0\u03af\u03c3\u03b7\u03c2 RMSNorm, cross-entropy, RetNet retention, Mamba-2 state space duality \u03ba\u03b1\u03b9 triangle multiplication \u03b1\u03c0\u03cc \u03c4\u03bf AlphaFold2. \u039f\u03b9 \u03b4\u03b9\u03b1\u03c3\u03c4\u03ac\u03c3\u03b5\u03b9\u03c2 \u03c0\u03c1\u03bf\u03ad\u03c1\u03c7\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03c4\u03b1 \u03b1\u03bd\u03c4\u03af\u03c3\u03c4\u03bf\u03b9\u03c7\u03b1 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1\u00b7 \u03b3\u03b9\u03b1 \u03c0\u03b1\u03c1\u03ac\u03b4\u03b5\u03b9\u03b3\u03bc\u03b1, \u03c4\u03bf GQA workload \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af 128 query heads, 8 key-value heads \u03ba\u03b1\u03b9 sequence length 4096.<\/p>\n<p>\u0393\u03b9\u03b1 \u03bf\u03ba\u03c4\u03ce priority kernels \u03c5\u03c0\u03ae\u03c1\u03c7\u03b1\u03bd hand-optimized Pallas implementations \u03b1\u03c0\u03cc \u03c4\u03bf Tokamax. \u0397 \u03bf\u03bc\u03ac\u03b4\u03b1 \u03b4\u03bf\u03ba\u03af\u03bc\u03b1\u03c3\u03b5 \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03ac 203 block-size configurations \u03ce\u03c3\u03c4\u03b5 \u03bd\u03b1 \u03b4\u03b7\u03bc\u03b9\u03bf\u03c5\u03c1\u03b3\u03ae\u03c3\u03b5\u03b9 \u03ad\u03bd\u03b1 \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc expert upper bound. \u0397 \u03c1\u03cd\u03b8\u03bc\u03b9\u03c3\u03b7 \u03b1\u03c5\u03c4\u03ae \u03ad\u03b4\u03c9\u03c3\u03b5 \u03ad\u03c9\u03c2 2,79\u00d7 \u03b2\u03b5\u03bb\u03c4\u03af\u03c9\u03c3\u03b7 \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 \u03c4\u03c9\u03bd \u03c0\u03c1\u03bf\u03b5\u03c0\u03b9\u03bb\u03b5\u03b3\u03bc\u03ad\u03bd\u03c9\u03bd \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03c4\u03c1\u03c9\u03bd Pallas \u03b3\u03b9\u03b1 \u03c4\u03bf Megablox GMM. \u03a4\u03bf benchmark \u03b5\u03c0\u03bf\u03bc\u03ad\u03bd\u03c9\u03c2 \u03c3\u03c5\u03b3\u03ba\u03c1\u03af\u03bd\u03b5\u03b9 \u03c4\u03bf\u03c5\u03c2 agents \u03c4\u03cc\u03c3\u03bf \u03bc\u03b5 \u03c4\u03bf XLA \u03cc\u03c3\u03bf \u03ba\u03b1\u03b9, \u03c3\u03b5 \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03b9\u03c3\u03bc\u03ad\u03bd\u03bf \u03c5\u03c0\u03bf\u03c3\u03cd\u03bd\u03bf\u03bb\u03bf, \u03bc\u03b5 \u03b1\u03bd\u03b8\u03c1\u03ce\u03c0\u03b9\u03bd\u03b1 \u03b2\u03b5\u03bb\u03c4\u03b9\u03c3\u03c4\u03bf\u03c0\u03bf\u03b9\u03b7\u03bc\u03ad\u03bd\u03bf \u03ba\u03ce\u03b4\u03b9\u03ba\u03b1.<\/p>\n<h2 id=\"correctness-epanalipsimotita\">\u03a0\u03ce\u03c2 \u03b5\u03be\u03b1\u03c3\u03c6\u03b1\u03bb\u03af\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 correctness \u03ba\u03b1\u03b9 \u03b5\u03c0\u03b1\u03bd\u03b1\u03bb\u03b7\u03c8\u03b9\u03bc\u03cc\u03c4\u03b7\u03c4\u03b1<\/h2>\n<p>\u039a\u03ac\u03b8\u03b5 workload \u03b4\u03b9\u03b1\u03b8\u03ad\u03c4\u03b5\u03b9 \u03ba\u03bf\u03b9\u03bd\u03cc interface \u03bc\u03b5 configuration, \u03b4\u03b7\u03bc\u03b9\u03bf\u03c5\u03c1\u03b3\u03af\u03b1 bf16 inputs \u03ba\u03b1\u03b9 \u03c3\u03c5\u03bd\u03ac\u03c1\u03c4\u03b7\u03c3\u03b7 \u03c5\u03c0\u03bf\u03bb\u03bf\u03b3\u03b9\u03c3\u03bc\u03bf\u03cd. \u039f\u03b9 \u03bc\u03b5\u03c4\u03b1\u03c6\u03c1\u03ac\u03c3\u03b5\u03b9\u03c2 \u03b1\u03c0\u03cc PyTorch \u03b5\u03bb\u03ad\u03b3\u03c7\u03b8\u03b7\u03ba\u03b1\u03bd \u03c3\u03c4\u03b7\u03bd TPU \u03bc\u03b5 \u03b1\u03bd\u03c4\u03b9\u03c3\u03c4\u03bf\u03b9\u03c7\u03b9\u03c3\u03bc\u03ad\u03bd\u03b5\u03c2 \u03b5\u03b9\u03c3\u03cc\u03b4\u03bf\u03c5\u03c2 \u03ba\u03b1\u03b9 \u03b1\u03c0\u03b1\u03af\u03c4\u03b7\u03c3\u03b7 \u03c3\u03c5\u03bc\u03c6\u03c9\u03bd\u03af\u03b1\u03c2 \u03bc\u03ad\u03c3\u03c9 jnp.allclose \u03bc\u03b5 atol \u03ba\u03b1\u03b9 rtol 10<sup>-2<\/sup>. \u0391\u03bd \u03bc\u03b9\u03b1 \u03bc\u03b5\u03c4\u03ac\u03c6\u03c1\u03b1\u03c3\u03b7 \u03b1\u03c0\u03bf\u03c4\u03cd\u03b3\u03c7\u03b1\u03bd\u03b5, \u03b1\u03bd\u03b1\u03b4\u03b7\u03bc\u03b9\u03bf\u03c5\u03c1\u03b3\u03bf\u03cd\u03bd\u03c4\u03b1\u03bd \u03ae \u03b4\u03b9\u03bf\u03c1\u03b8\u03c9\u03bd\u03cc\u03c4\u03b1\u03bd \u03c0\u03c1\u03b9\u03bd \u03b5\u03bd\u03c4\u03b1\u03c7\u03b8\u03b5\u03af \u03c3\u03c4\u03b7 \u03c3\u03bf\u03c5\u03af\u03c4\u03b1.<\/p>\n<p>\u0393\u03b9\u03b1 \u03c4\u03bf\u03bd \u03c7\u03c1\u03cc\u03bd\u03bf \u03b5\u03ba\u03c4\u03ad\u03bb\u03b5\u03c3\u03b7\u03c2 \u03b4\u03b5\u03bd \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03ae\u03b8\u03b7\u03ba\u03b5 \u03b1\u03c0\u03bb\u03cc wall-clock timing. \u03a4\u03bf jax.profiler \u03ba\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c6\u03b5\u03b9 Perfetto-compatible traces, \u03bc\u03b5\u03c4\u03ac \u03b1\u03c0\u03cc \u03c0\u03ad\u03bd\u03c4\u03b5 warmup iterations \u03ba\u03b1\u03b9 \u03ba\u03b1\u03c4\u03ac \u03c4\u03b7 \u03b4\u03b9\u03ac\u03c1\u03ba\u03b5\u03b9\u03b1 50 timed iterations. \u0397 \u03ba\u03cd\u03c1\u03b9\u03b1 \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b7 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b7 \u03b4\u03b9\u03ac\u03bc\u03b5\u03c3\u03bf\u03c2 \u03c4\u03c9\u03bd device-side events. \u0397 \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ae \u03b5\u03af\u03bd\u03b1\u03b9 \u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03b7 \u03b3\u03b9\u03b1 kernels \u03ba\u03ac\u03c4\u03c9 \u03b1\u03c0\u03cc \u03ad\u03bd\u03b1 millisecond, \u03cc\u03c0\u03bf\u03c5 \u03c4\u03bf Python dispatch, \u03bf \u03c0\u03c1\u03bf\u03b3\u03c1\u03b1\u03bc\u03bc\u03b1\u03c4\u03b9\u03c3\u03bc\u03cc\u03c2 \u03c4\u03bf\u03c5 runtime \u03ba\u03b1\u03b9 \u03bf \u03c3\u03c5\u03b3\u03c7\u03c1\u03bf\u03bd\u03b9\u03c3\u03bc\u03cc\u03c2 \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03c3\u03cd\u03bc\u03c6\u03c9\u03bd\u03b1 \u03bc\u03b5 \u03c4\u03b7\u03bd \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03bd\u03b1 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03bf\u03cd\u03bd 10% \u03ad\u03c9\u03c2 20% \u03c4\u03bf\u03c5 \u03c6\u03b1\u03b9\u03bd\u03bf\u03bc\u03b5\u03bd\u03b9\u03ba\u03bf\u03cd \u03c7\u03c1\u03cc\u03bd\u03bf\u03c5.<\/p>\n<h2 id=\"tesseris-proseggiseis\">\u039f\u03b9 \u03c4\u03ad\u03c3\u03c3\u03b5\u03c1\u03b9\u03c2 \u03c0\u03c1\u03bf\u03c3\u03b5\u03b3\u03b3\u03af\u03c3\u03b5\u03b9\u03c2 \u03c0\u03bf\u03c5 \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03ae\u03b8\u03b7\u03ba\u03b1\u03bd<\/h2>\n<p>\u0397 \u03c0\u03c1\u03ce\u03c4\u03b7 \u03bc\u03ad\u03b8\u03bf\u03b4\u03bf\u03c2 \u03b5\u03af\u03bd\u03b1\u03b9 best-of-N: \u03b1\u03bd\u03b5\u03be\u03ac\u03c1\u03c4\u03b7\u03c4\u03b1 one-shot samples \u03bc\u03b5 TPU preamble \u03ba\u03b1\u03b9 \u03c4\u03bf\u03bd JAX source code, \u03b1\u03c0\u03cc \u03c4\u03b1 \u03bf\u03c0\u03bf\u03af\u03b1 \u03ba\u03c1\u03b1\u03c4\u03b5\u03af\u03c4\u03b1\u03b9 \u03c4\u03bf \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf \u03c3\u03c9\u03c3\u03c4\u03cc \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1. \u0397 \u03b4\u03b5\u03cd\u03c4\u03b5\u03c1\u03b7 \u03b5\u03af\u03bd\u03b1\u03b9 iterative refinement, \u03cc\u03c0\u03bf\u03c5 \u03bf agent \u03bb\u03b1\u03bc\u03b2\u03ac\u03bd\u03b5\u03b9 \u03b1\u03bd\u03ac \u03b3\u03cd\u03c1\u03bf compilation errors, correctness results \u03ba\u03b1\u03b9 profiler summaries. \u0397 \u03c4\u03c1\u03af\u03c4\u03b7 \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03b5\u03af \u03c4\u03b7\u03bd \u03af\u03b4\u03b9\u03b1 iterative \u03b4\u03b9\u03b1\u03b4\u03b9\u03ba\u03b1\u03c3\u03af\u03b1 \u03b1\u03bb\u03bb\u03ac \u03c0\u03c1\u03bf\u03c3\u03b8\u03ad\u03c4\u03b5\u03b9 \u03b5\u03c0\u03b9\u03bc\u03b5\u03bb\u03b7\u03bc\u03ad\u03bd\u03bf context \u03b1\u03c0\u03cc \u03c4\u03bf Autocomp.<\/p>\n<p>\u0397 \u03c4\u03ad\u03c4\u03b1\u03c1\u03c4\u03b7 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c4\u03bf \u03af\u03b4\u03b9\u03bf \u03c4\u03bf Autocomp, \u03ad\u03bd\u03b1 \u03b1\u03bd\u03bf\u03b9\u03ba\u03c4\u03cc framework \u03c0\u03bf\u03c5 \u03b1\u03c0\u03bf\u03c3\u03c4\u03ac\u03b6\u03b5\u03b9 \u03b4\u03b7\u03bc\u03cc\u03c3\u03b9\u03b1 \u03c4\u03b5\u03ba\u03bc\u03b7\u03c1\u03af\u03c9\u03c3\u03b7 JAX Pallas \u03ba\u03b1\u03b9 Cloud TPU \u03c3\u03b5 \u03c0\u03b5\u03c1\u03af\u03bb\u03b7\u03c8\u03b7 \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae\u03c2, \u03b1\u03bd\u03b1\u03c6\u03bf\u03c1\u03ac API, \u03c3\u03c7\u03bf\u03bb\u03b9\u03b1\u03c3\u03bc\u03ad\u03bd\u03b1 \u03c0\u03b1\u03c1\u03b1\u03b4\u03b5\u03af\u03b3\u03bc\u03b1\u03c4\u03b1 \u03ba\u03b1\u03b9 \u03ba\u03b1\u03bd\u03cc\u03bd\u03b5\u03c2 correctness. \u03a7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af \u03b4\u03cd\u03bf \u03c6\u03ac\u03c3\u03b5\u03b9\u03c2 beam search: \u03c0\u03c1\u03ce\u03c4\u03b1 \u03bc\u03b5\u03c4\u03ac\u03c6\u03c1\u03b1\u03c3\u03b7 \u03c4\u03bf\u03c5 XLA baseline \u03c3\u03b5 Pallas \u03ba\u03b1\u03b9 \u03bc\u03b5\u03c4\u03ac \u03b2\u03b5\u03bb\u03c4\u03b9\u03c3\u03c4\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 \u03b1\u03c0\u03cc\u03b4\u03bf\u03c3\u03b7\u03c2. \u038c\u03bb\u03b5\u03c2 \u03bf\u03b9 \u03ba\u03cd\u03c1\u03b9\u03b5\u03c2 \u03b4\u03bf\u03ba\u03b9\u03bc\u03ad\u03c2 \u03ad\u03b3\u03b9\u03bd\u03b1\u03bd \u03bc\u03b5 Gemini 3 Flash \u03ba\u03b1\u03b9 \u03c3\u03c5\u03b3\u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03bf budget \u03ad\u03c9\u03c2 144 samples \u03b1\u03bd\u03ac benchmark.<\/p>\n<h2 id=\"context-allazei-orthotita\">\u03a4\u03bf \u03b2\u03b1\u03c3\u03b9\u03ba\u03cc \u03b5\u03cd\u03c1\u03b7\u03bc\u03b1: \u03c4\u03bf \u03b5\u03be\u03b5\u03b9\u03b4\u03b9\u03ba\u03b5\u03c5\u03bc\u03ad\u03bd\u03bf context \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd \u03bf\u03c1\u03b8\u03cc\u03c4\u03b7\u03c4\u03b1<\/h2>\n<p>\u03a3\u03c4\u03bf \u03c0\u03bb\u03ae\u03c1\u03b5\u03c2 \u03c3\u03cd\u03bd\u03bf\u03bb\u03bf \u03c4\u03c9\u03bd 50 workloads, \u03c4\u03bf \u03b1\u03c0\u03bb\u03cc best-of-N \u03c0\u03b1\u03c1\u03ae\u03b3\u03b1\u03b3\u03b5 \u03c3\u03c9\u03c3\u03c4\u03cc kernel \u03c3\u03b5 13 \u03c0\u03b5\u03c1\u03b9\u03c0\u03c4\u03ce\u03c3\u03b5\u03b9\u03c2 \u03ba\u03b1\u03b9 \u03ad\u03c6\u03c4\u03b1\u03c3\u03b5 \u03b3\u03b5\u03c9\u03bc\u03b5\u03c4\u03c1\u03b9\u03ba\u03cc \u03bc\u03ad\u03c3\u03bf speedup 1,01\u00d7. \u03a4\u03bf iterative refinement \u03ad\u03bb\u03c5\u03c3\u03b5 32 \u03b1\u03c0\u03cc \u03c4\u03b1 50 \u03ba\u03b1\u03b9 \u03ad\u03c6\u03c4\u03b1\u03c3\u03b5 1,18\u00d7. \u039c\u03b5 \u03c4\u03b7\u03bd \u03c0\u03c1\u03bf\u03c3\u03b8\u03ae\u03ba\u03b7 \u03c4\u03bf\u03c5 \u03b5\u03c0\u03b9\u03bc\u03b5\u03bb\u03b7\u03bc\u03ad\u03bd\u03bf\u03c5 TPU context, \u03b7 iterative \u03bc\u03ad\u03b8\u03bf\u03b4\u03bf\u03c2 \u03ad\u03bb\u03c5\u03c3\u03b5 48 \u03b1\u03c0\u03cc \u03c4\u03b1 50 \u03ba\u03b1\u03b9 \u03b1\u03bd\u03ad\u03b2\u03b7\u03ba\u03b5 \u03c3\u03c4\u03bf 1,28\u00d7. \u03a4\u03bf Autocomp \u03ad\u03bb\u03c5\u03c3\u03b5 45, \u03b1\u03bb\u03bb\u03ac \u03c0\u03ad\u03c4\u03c5\u03c7\u03b5 \u03c4\u03bf\u03bd \u03c5\u03c8\u03b7\u03bb\u03cc\u03c4\u03b5\u03c1\u03bf \u03b3\u03b5\u03c9\u03bc\u03b5\u03c4\u03c1\u03b9\u03ba\u03cc \u03bc\u03ad\u03c3\u03bf, 1,36\u00d7, \u03b5\u03c0\u03b5\u03b9\u03b4\u03ae \u03b1\u03c6\u03b9\u03ad\u03c1\u03c9\u03c3\u03b5 \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03bf \u03b1\u03c0\u03cc \u03c4\u03bf budget \u03c3\u03c4\u03b7 \u03b2\u03b5\u03bb\u03c4\u03b9\u03c3\u03c4\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03bc\u03cc\u03bd\u03bf \u03c3\u03c4\u03bf debugging.<\/p>\n<div class=\"td-chart td-chart--metrics\">\n<p class=\"td-chart-title\">\u03a4\u03b1 \u03c4\u03b5\u03ba\u03bc\u03b7\u03c1\u03b9\u03c9\u03bc\u03ad\u03bd\u03b1 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1 \u03c4\u03bf\u03c5 JAXBench<\/p>\n<p class=\"td-chart-intro\">\u039f\u03b9 \u03c4\u03b9\u03bc\u03ad\u03c2 \u03b1\u03c6\u03bf\u03c1\u03bf\u03cd\u03bd \u03c4\u03b9\u03c2 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b5\u03c2 \u03b4\u03bf\u03ba\u03b9\u03bc\u03ad\u03c2 \u03c4\u03b7\u03c2 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1\u03c2 \u03c3\u03b5 TPU v6e \u03ba\u03b1\u03b9 \u03b4\u03b5\u03bd \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03bf\u03cd\u03bd \u03b3\u03b5\u03bd\u03b9\u03ba\u03ae \u03c0\u03c1\u03cc\u03b2\u03bb\u03b5\u03c8\u03b7 \u03b3\u03b9\u03b1 \u03ba\u03ac\u03b8\u03b5 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03ae workload.<\/p>\n<div class=\"td-metric-grid\">\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">50<\/span><strong>JAX workloads<\/strong><\/p>\n<p>17 production operators \u03ba\u03b1\u03b9 33 fused operators \u03bc\u03b5 \u03c0\u03c1\u03cc\u03b2\u03bb\u03b7\u03bc\u03b1 \u03bc\u03b5\u03b3\u03ad\u03b8\u03bf\u03c5\u03c2 \u03ba\u03b1\u03c4\u03ac\u03bb\u03bb\u03b7\u03bb\u03bf \u03b3\u03b9\u03b1 TPU.<\/p>\n<\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">48\/50<\/span><strong>\u03a3\u03c9\u03c3\u03c4\u03bf\u03af kernels \u03bc\u03b5 context<\/strong><\/p>\n<p>\u0397 iterative \u03bc\u03ad\u03b8\u03bf\u03b4\u03bf\u03c2 \u03bc\u03b5 \u03b5\u03c0\u03b9\u03bc\u03b5\u03bb\u03b7\u03bc\u03ad\u03bd\u03b7 \u03c4\u03b5\u03ba\u03bc\u03b7\u03c1\u03af\u03c9\u03c3\u03b7 TPU \u03ad\u03bb\u03c5\u03c3\u03b5 48 workloads.<\/p>\n<\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">1,36\u00d7<\/span><strong>Autocomp \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 XLA<\/strong><\/p>\n<p>\u039f \u03c5\u03c8\u03b7\u03bb\u03cc\u03c4\u03b5\u03c1\u03bf\u03c2 \u03b3\u03b5\u03c9\u03bc\u03b5\u03c4\u03c1\u03b9\u03ba\u03cc\u03c2 \u03bc\u03ad\u03c3\u03bf\u03c2 \u03c3\u03c4\u03bf \u03c0\u03bb\u03ae\u03c1\u03b5\u03c2 benchmark \u03bc\u03b5 Gemini 3 Flash.<\/p>\n<\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">1,60\u00d7<\/span><strong>\u03a3\u03c4\u03bf expert \u03c5\u03c0\u03bf\u03c3\u03cd\u03bd\u03bf\u03bb\u03bf<\/strong><\/p>\n<p>Autocomp \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 XLA \u03c3\u03c4\u03b1 \u03bf\u03ba\u03c4\u03ce workloads \u03bc\u03b5 hand-tuned Tokamax references.<\/p>\n<\/div>\n<\/div>\n<\/div>\n<p>\u0397 \u03ba\u03b1\u03b8\u03b1\u03c1\u03cc\u03c4\u03b5\u03c1\u03b7 \u03ad\u03bd\u03b4\u03b5\u03b9\u03be\u03b7 \u03b1\u03c6\u03bf\u03c1\u03ac \u03c4\u03b7\u03bd \u03bf\u03c1\u03b8\u03cc\u03c4\u03b7\u03c4\u03b1 \u03b1\u03bd\u03ac sample: \u03c4\u03bf curated context \u03c4\u03b7\u03bd \u03b1\u03cd\u03be\u03b7\u03c3\u03b5 \u03b1\u03c0\u03cc 5,8% \u03c3\u03b5 37,3% \u03c7\u03c9\u03c1\u03af\u03c2 \u03b1\u03bb\u03bb\u03b1\u03b3\u03ae \u03c3\u03c4\u03bf\u03bd \u03b1\u03bb\u03b3\u03cc\u03c1\u03b9\u03b8\u03bc\u03bf \u03b1\u03bd\u03b1\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7\u03c2. \u03a0\u03b1\u03c1\u03ac\u03bb\u03bb\u03b7\u03bb\u03b1, \u03b7 \u03ba\u03b1\u03c4\u03ac\u03c7\u03c1\u03b7\u03c3\u03b7 \u03c4\u03bf\u03c5 Pallas API \u03c0\u03b1\u03c1\u03ad\u03bc\u03b5\u03b9\u03bd\u03b5 \u03b7 \u03ba\u03c5\u03c1\u03af\u03b1\u03c1\u03c7\u03b7 \u03c0\u03b7\u03b3\u03ae \u03b1\u03c0\u03bf\u03c4\u03c5\u03c7\u03af\u03b1\u03c2. \u03a3\u03c4\u03bf best-of-N, 99,7% \u03c4\u03c9\u03bd samples \u03b1\u03c0\u03ad\u03c4\u03c5\u03c7\u03b1\u03bd \u03ba\u03b1\u03c4\u03ac \u03c4\u03b7 \u03bc\u03b5\u03c4\u03b1\u03b3\u03bb\u03ce\u03c4\u03c4\u03b9\u03c3\u03b7 \u03ae \u03c4\u03b7\u03bd \u03c0\u03c1\u03ce\u03c4\u03b7 \u03b5\u03ba\u03c4\u03ad\u03bb\u03b5\u03c3\u03b7 \u03bb\u03cc\u03b3\u03c9 API\/runtime \u03c0\u03c1\u03bf\u03b2\u03bb\u03b7\u03bc\u03ac\u03c4\u03c9\u03bd, \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 59,8% \u03c3\u03c4\u03bf iterative \u03bc\u03b5 context \u03ba\u03b1\u03b9 55,8% \u03c3\u03c4\u03bf Autocomp.<\/p>\n<h2 id=\"agents-anthropinoi-kernels\">\u03a0\u03cc\u03c3\u03bf \u03ba\u03bf\u03bd\u03c4\u03ac \u03ad\u03c6\u03c4\u03b1\u03c3\u03b1\u03bd \u03bf\u03b9 agents \u03c3\u03c4\u03bf\u03c5\u03c2 \u03b1\u03bd\u03b8\u03c1\u03ce\u03c0\u03b9\u03bd\u03bf\u03c5\u03c2 kernels<\/h2>\n<p>\u03a3\u03c4\u03b1 \u03bf\u03ba\u03c4\u03ce workloads \u03bc\u03b5 hand-tuned references, \u03c4\u03bf Tokamax \u03ad\u03b4\u03c9\u03c3\u03b5 \u03b3\u03b5\u03c9\u03bc\u03b5\u03c4\u03c1\u03b9\u03ba\u03cc \u03bc\u03ad\u03c3\u03bf 2,08\u00d7 \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 XLA, \u03bc\u03b5 floor \u03c3\u03c4\u03bf 1\u00d7 \u03b3\u03b9\u03b1 \u03c4\u03b7 \u03c3\u03c5\u03b3\u03ba\u03ad\u03bd\u03c4\u03c1\u03c9\u03c3\u03b7. \u03a4\u03bf Autocomp \u03ad\u03c6\u03c4\u03b1\u03c3\u03b5 1,60\u00d7, \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 \u03c4\u03bf 77% \u03c4\u03bf\u03c5 hand-tuned \u03b3\u03b5\u03c9\u03bc\u03b5\u03c4\u03c1\u03b9\u03ba\u03bf\u03cd \u03bc\u03ad\u03c3\u03bf\u03c5. \u039e\u03b5\u03c0\u03ad\u03c1\u03b1\u03c3\u03b5 \u03c4\u03bf Tokamax \u03c3\u03b5 \u03b4\u03cd\u03bf kernels \u03ba\u03b1\u03b9 \u03b2\u03c1\u03ad\u03b8\u03b7\u03ba\u03b5 \u03c3\u03c4\u03bf 68% \u03ad\u03c9\u03c2 91% \u03c4\u03b7\u03c2 \u03b5\u03c0\u03af\u03b4\u03bf\u03c3\u03ae\u03c2 \u03c4\u03bf\u03c5 \u03c3\u03b5 \u03c4\u03ad\u03c3\u03c3\u03b5\u03c1\u03b1 \u03b1\u03ba\u03cc\u03bc\u03b7.<\/p>\n<p>\u0397 \u03b1\u03c0\u03cc\u03c3\u03c4\u03b1\u03c3\u03b7 \u03c0\u03b1\u03c1\u03ad\u03bc\u03b5\u03b9\u03bd\u03b5 \u03bc\u03b5\u03b3\u03ac\u03bb\u03b7 \u03c3\u03c4\u03b1 paged \u03ba\u03b1\u03b9 ragged attention workloads, \u03cc\u03c0\u03bf\u03c5 \u03bf \u03c7\u03b5\u03b9\u03c1\u03bf\u03ba\u03af\u03bd\u03b7\u03c4\u03bf\u03c2 scheduling \u03b5\u03af\u03bd\u03b1\u03b9 \u03b9\u03b4\u03b9\u03b1\u03af\u03c4\u03b5\u03c1\u03b1 \u03c3\u03b7\u03bc\u03b1\u03bd\u03c4\u03b9\u03ba\u03cc\u03c2. \u03a4\u03bf hand-tuned ragged paged attention \u03ae\u03c4\u03b1\u03bd 6,91\u00d7 \u03c4\u03b1\u03c7\u03cd\u03c4\u03b5\u03c1\u03bf \u03b1\u03c0\u03cc XLA \u03c3\u03c4\u03bf \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf table, \u03b5\u03bd\u03ce \u03bf agent \u03b4\u03b5\u03bd \u03c0\u03b1\u03c1\u03ae\u03b3\u03b1\u03b3\u03b5 \u03c3\u03c9\u03c3\u03c4\u03cc kernel. \u0391\u03bd\u03c4\u03af\u03b8\u03b5\u03c4\u03b1, \u03c3\u03c4\u03bf sparse attention \u03c4\u03bf Autocomp \u03ad\u03c6\u03c4\u03b1\u03c3\u03b5 2,81\u00d7, \u03c0\u03ac\u03bd\u03c9 \u03b1\u03c0\u03cc \u03c4\u03bf 0,86\u00d7 \u03c4\u03b7\u03c2 \u03b4\u03b9\u03b1\u03b8\u03ad\u03c3\u03b9\u03bc\u03b7\u03c2 Tokamax \u03c0\u03b1\u03c1\u03b1\u03bb\u03bb\u03b1\u03b3\u03ae\u03c2, \u03b7 \u03bf\u03c0\u03bf\u03af\u03b1 \u03b5\u03af\u03c7\u03b5 \u03c1\u03c5\u03b8\u03bc\u03b9\u03c3\u03c4\u03b5\u03af \u03b3\u03b9\u03b1 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03cc sparsity pattern.<\/p>\n<h2 id=\"modelo-gnosi-stochou\">\u039c\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03ae \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b7 \u03b3\u03bd\u03ce\u03c3\u03b7 \u03c3\u03c4\u03cc\u03c7\u03bf\u03c5;<\/h2>\n<p>\u0397 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03b5\u03be\u03b5\u03c4\u03ac\u03b6\u03b5\u03b9 \u03b5\u03c0\u03af\u03c3\u03b7\u03c2 \u03ad\u03bd\u03b1 \u03c5\u03c0\u03bf\u03c3\u03cd\u03bd\u03bf\u03bb\u03bf \u03c0\u03ad\u03bd\u03c4\u03b5 kernels \u03bc\u03b5 Gemini 3.1 Pro. \u0397 \u03bc\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b7 \u03b9\u03ba\u03b1\u03bd\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c4\u03bf\u03c5 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf\u03c5 \u03b2\u03b5\u03bb\u03c4\u03af\u03c9\u03c3\u03b5 \u03cc\u03bb\u03b5\u03c2 \u03c4\u03b9\u03c2 \u03bc\u03b5\u03b8\u03cc\u03b4\u03bf\u03c5\u03c2: \u03c4\u03bf plain iterative \u03b1\u03bd\u03ad\u03b2\u03b7\u03ba\u03b5 \u03b1\u03c0\u03cc 1,07\u00d7 \u03bc\u03b5 Flash \u03c3\u03b5 2,43\u00d7 \u03bc\u03b5 Pro, \u03c4\u03bf iterative \u03bc\u03b5 context \u03b1\u03c0\u03cc 1,59\u00d7 \u03c3\u03b5 3,82\u00d7 \u03ba\u03b1\u03b9 \u03c4\u03bf Autocomp \u03b1\u03c0\u03cc 2,35\u00d7 \u03c3\u03b5 3,79\u00d7. \u03a3\u03c4\u03bf \u03bc\u03b9\u03ba\u03c1\u03cc \u03b1\u03c5\u03c4\u03cc \u03c5\u03c0\u03bf\u03c3\u03cd\u03bd\u03bf\u03bb\u03bf, \u03bf\u03b9 \u03b4\u03cd\u03bf context-aware \u03bc\u03ad\u03b8\u03bf\u03b4\u03bf\u03b9 \u03c0\u03b1\u03c1\u03ae\u03b3\u03b1\u03b3\u03b1\u03bd \u03c3\u03c9\u03c3\u03c4\u03cc Pallas \u03ba\u03b1\u03b9 \u03b3\u03b9\u03b1 \u03c4\u03b1 \u03c0\u03ad\u03bd\u03c4\u03b5 workloads.<\/p>\n<p>\u03a4\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03b4\u03b5\u03bd \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03c4\u03bf model scale \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03b4\u03b9\u03ac\u03c6\u03bf\u03c1\u03bf. \u0394\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03cc\u03bc\u03c9\u03c2 \u03cc\u03c4\u03b9 \u03b7 \u03b9\u03ba\u03b1\u03bd\u03cc\u03c4\u03b7\u03c4\u03b1 \u03ba\u03b1\u03b9 \u03c4\u03bf context \u03b1\u03bb\u03bb\u03b7\u03bb\u03b5\u03c0\u03b9\u03b4\u03c1\u03bf\u03cd\u03bd. \u03a3\u03c4\u03bf sparsely documented Pallas, \u03b7 \u03c4\u03b5\u03ba\u03bc\u03b7\u03c1\u03af\u03c9\u03c3\u03b7 \u03c0\u03c1\u03bf\u03c3\u03c6\u03ad\u03c1\u03b5\u03b9 \u03c4\u03b9\u03c2 \u03c0\u03c1\u03bf\u03cb\u03c0\u03bf\u03b8\u03ad\u03c3\u03b5\u03b9\u03c2 \u03b3\u03b9\u03b1 \u03bd\u03b1 \u03b3\u03af\u03bd\u03b5\u03b9 \u03b7 \u03b1\u03bd\u03b1\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7 \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03b9\u03ba\u03ae. \u039c\u03cc\u03bb\u03b9\u03c2 \u03ad\u03bd\u03b1 \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03c0\u03b1\u03c1\u03ac\u03b3\u03b5\u03b9 \u03c3\u03c9\u03c3\u03c4\u03ac seeds \u03c7\u03c9\u03c1\u03af\u03c2 \u03b5\u03ba\u03c4\u03b5\u03c4\u03b1\u03bc\u03ad\u03bd\u03bf debugging, \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03b1 samples \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03bd\u03b1 \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b7\u03b8\u03bf\u03cd\u03bd \u03b3\u03b9\u03b1 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ae \u03b2\u03b5\u03bb\u03c4\u03b9\u03c3\u03c4\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7.<\/p>\n<h2 id=\"ai-ypodomes-epicheiriseis\">\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 \u03c0\u03bf\u03c5 \u03c7\u03c4\u03af\u03b6\u03bf\u03c5\u03bd AI \u03c5\u03c0\u03bf\u03b4\u03bf\u03bc\u03ad\u03c2<\/h2>\n<p>\u03a4\u03bf \u03c0\u03c1\u03ce\u03c4\u03bf \u03c0\u03c1\u03b1\u03ba\u03c4\u03b9\u03ba\u03cc \u03bc\u03ac\u03b8\u03b7\u03bc\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03cc\u03c4\u03b9 \u03ad\u03bd\u03b1 coding agent \u03b4\u03b5\u03bd \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03b5\u03af\u03c4\u03b1\u03b9 \u03bc\u03cc\u03bd\u03bf \u03b1\u03c0\u03cc \u03c4\u03bf \u03b1\u03bd \u03bf\u03bb\u03bf\u03ba\u03bb\u03b7\u03c1\u03ce\u03bd\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03c3\u03c5\u03bd\u03ac\u03c1\u03c4\u03b7\u03c3\u03b7. \u03a3\u03b5 performance-critical \u03c3\u03c5\u03c3\u03c4\u03ae\u03bc\u03b1\u03c4\u03b1 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c4\u03c1\u03b5\u03b9\u03c2 \u03c0\u03cd\u03bb\u03b5\u03c2: compilability, \u03b1\u03c1\u03b9\u03b8\u03bc\u03b7\u03c4\u03b9\u03ba\u03ae \u03bf\u03c1\u03b8\u03cc\u03c4\u03b7\u03c4\u03b1 \u03ba\u03b1\u03b9 \u03b1\u03c0\u03cc\u03b4\u03bf\u03c3\u03b7 \u03b1\u03c0\u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 \u03c3\u03b5 \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc baseline. \u0397 \u03c0\u03b1\u03c1\u03ac\u03bb\u03b5\u03b9\u03c8\u03b7 \u03c4\u03b7\u03c2 \u03b4\u03b5\u03cd\u03c4\u03b5\u03c1\u03b7\u03c2 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b4\u03ce\u03c3\u03b5\u03b9 \u03b3\u03c1\u03ae\u03b3\u03bf\u03c1\u03bf \u03b1\u03bb\u03bb\u03ac \u03bb\u03ac\u03b8\u03bf\u03c2 \u03ba\u03ce\u03b4\u03b9\u03ba\u03b1\u00b7 \u03b7 \u03c0\u03b1\u03c1\u03ac\u03bb\u03b5\u03b9\u03c8\u03b7 \u03c4\u03b7\u03c2 \u03c4\u03c1\u03af\u03c4\u03b7\u03c2 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03bc\u03b5\u03c4\u03b1\u03c4\u03c1\u03ad\u03c8\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03c4\u03b5\u03c7\u03bd\u03b9\u03ba\u03ac \u03b5\u03c0\u03b9\u03c4\u03c5\u03c7\u03b7\u03bc\u03ad\u03bd\u03b7 \u03b1\u03c5\u03c4\u03bf\u03bc\u03b1\u03c4\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 \u03c3\u03b5 \u03b1\u03ba\u03c1\u03b9\u03b2\u03cc\u03c4\u03b5\u03c1\u03b7 \u03c5\u03c0\u03bf\u03b4\u03bf\u03bc\u03ae.<\/p>\n<p>\u03a4\u03bf \u03b4\u03b5\u03cd\u03c4\u03b5\u03c1\u03bf \u03bc\u03ac\u03b8\u03b7\u03bc\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b7 \u03b1\u03be\u03af\u03b1 \u03b5\u03bd\u03cc\u03c2 \u03b5\u03bb\u03b5\u03b3\u03c7\u03cc\u03bc\u03b5\u03bd\u03bf\u03c5 knowledge pack. \u0391\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae hardware, \u03ad\u03b3\u03ba\u03c5\u03c1\u03b1 API references, \u03bc\u03b9\u03ba\u03c1\u03ac \u03b4\u03bf\u03ba\u03b9\u03bc\u03b1\u03c3\u03bc\u03ad\u03bd\u03b1 examples \u03ba\u03b1\u03b9 \u03ba\u03b1\u03bd\u03cc\u03bd\u03b5\u03c2 correctness \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03b9\u03bf \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03b1 \u03b1\u03c0\u03cc \u03bc\u03b9\u03b1 \u03b1\u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03bf\u03c0\u03bf\u03af\u03b7\u03c4\u03b7 \u03c3\u03c5\u03bb\u03bb\u03bf\u03b3\u03ae \u03b5\u03b3\u03b3\u03c1\u03ac\u03c6\u03c9\u03bd. \u0393\u03b9\u03b1 \u03bc\u03b9\u03b1 \u03bf\u03bc\u03ac\u03b4\u03b1 platform engineering, \u03b1\u03c5\u03c4\u03cc \u03bc\u03b5\u03c4\u03b1\u03c6\u03c1\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c3\u03b5 versioned \u03c4\u03b5\u03ba\u03bc\u03b7\u03c1\u03af\u03c9\u03c3\u03b7 \u03ba\u03b1\u03b9 reproducible harness \u03b3\u03cd\u03c1\u03c9 \u03b1\u03c0\u03cc \u03c4\u03bf\u03bd agent, \u03bc\u03b5 \u03c4\u03b9\u03c2 \u03af\u03b4\u03b9\u03b5\u03c2 \u03b1\u03c1\u03c7\u03ad\u03c2 \u03c0\u03bf\u03c5 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03ad\u03bd\u03b1 <a href=\"https:\/\/twodots.gr\/llm-routing-latency-accuracy-cost\/\">\u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 LLM routing \u03b3\u03b9\u03b1 latency, \u03b1\u03ba\u03c1\u03af\u03b2\u03b5\u03b9\u03b1 \u03ba\u03b1\u03b9 \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2<\/a>.<\/p>\n<p>\u03a4\u03bf \u03c4\u03c1\u03af\u03c4\u03bf \u03b5\u03af\u03bd\u03b1\u03b9 \u03b7 \u03c3\u03c9\u03c3\u03c4\u03ae \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ae \u03c0\u03c1\u03bf\u03b2\u03bb\u03b7\u03bc\u03ac\u03c4\u03c9\u03bd. \u039f\u03b9 \u03b5\u03c1\u03b5\u03c5\u03bd\u03b7\u03c4\u03ad\u03c2 \u03b4\u03b5\u03af\u03c7\u03bd\u03bf\u03c5\u03bd \u03cc\u03c4\u03b9 compute-bound GEMM \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03ae\u03b4\u03b7 \u03bd\u03b1 \u03ba\u03bf\u03c1\u03ad\u03c3\u03b5\u03b9 \u03c4\u03b7\u03bd MXU \u03bc\u03ad\u03c3\u03c9 XLA \u03ba\u03b1\u03b9 \u03bd\u03b1 \u03bc\u03b7\u03bd \u03b1\u03c6\u03ae\u03bd\u03b5\u03b9 \u03b1\u03be\u03b9\u03cc\u03bb\u03bf\u03b3\u03bf \u03c0\u03b5\u03c1\u03b9\u03b8\u03ce\u03c1\u03b9\u03bf. \u0391\u03bd\u03c4\u03af\u03b8\u03b5\u03c4\u03b1, memory-bound attention kernels \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03bd\u03b1 \u03c9\u03c6\u03b5\u03bb\u03b7\u03b8\u03bf\u03cd\u03bd \u03b1\u03c0\u03cc tiling \u03c0\u03bf\u03c5 \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03af\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b5\u03c0\u03b1\u03bd\u03b1\u03bb\u03b1\u03bc\u03b2\u03b1\u03bd\u03cc\u03bc\u03b5\u03bd\u03b7 \u03ba\u03af\u03bd\u03b7\u03c3\u03b7 \u03c0\u03c1\u03bf\u03c2 HBM. \u039c\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03af\u03c1\u03b7\u03c3\u03b7 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bb\u03bf\u03b9\u03c0\u03cc\u03bd \u03bd\u03b1 \u03ba\u03ac\u03bd\u03b5\u03b9 profiling \u03c0\u03c1\u03b9\u03bd \u03b5\u03c0\u03b5\u03bd\u03b4\u03cd\u03c3\u03b5\u03b9 \u03c3\u03b5 agentic optimization \u03ba\u03b1\u03b9 \u03bd\u03b1 \u03ba\u03b1\u03c4\u03b5\u03c5\u03b8\u03cd\u03bd\u03b5\u03b9 \u03c4\u03bf budget \u03c3\u03c4\u03b1 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac bottlenecks.<\/p>\n<h2 id=\"oria-meletis\">\u038c\u03c1\u03b9\u03b1 \u03c4\u03b7\u03c2 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7\u03c2 \u03ba\u03b1\u03b9 \u03b1\u03c3\u03c6\u03b1\u03bb\u03ae\u03c2 \u03b1\u03bd\u03ac\u03b3\u03bd\u03c9\u03c3\u03b7 \u03c4\u03c9\u03bd \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03b5\u03c3\u03bc\u03ac\u03c4\u03c9\u03bd<\/h2>\n<p>\u03a4\u03b1 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1 \u03b1\u03c6\u03bf\u03c1\u03bf\u03cd\u03bd \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b7 \u03c3\u03bf\u03c5\u03af\u03c4\u03b1, TPU v6e, Pallas\/Mosaic \u03ba\u03b1\u03b9 \u03c4\u03b1 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1 \u03c0\u03bf\u03c5 \u03b4\u03bf\u03ba\u03b9\u03bc\u03ac\u03c3\u03c4\u03b7\u03ba\u03b1\u03bd. \u0397 \u03c3\u03cd\u03b3\u03ba\u03c1\u03b9\u03c3\u03b7 Gemini 3.1 Pro \u03ad\u03b3\u03b9\u03bd\u03b5 \u03bc\u03cc\u03bd\u03bf \u03c3\u03b5 \u03c0\u03ad\u03bd\u03c4\u03b5 kernels, \u03bb\u03cc\u03b3\u03c9 \u03ba\u03cc\u03c3\u03c4\u03bf\u03c5\u03c2, \u03ba\u03b1\u03b9 \u03bf\u03b9 \u03c3\u03c5\u03b3\u03b3\u03c1\u03b1\u03c6\u03b5\u03af\u03c2 \u03c4\u03b7 \u03c7\u03b1\u03c1\u03b1\u03ba\u03c4\u03b7\u03c1\u03af\u03b6\u03bf\u03c5\u03bd \u03c0\u03c1\u03bf\u03ba\u03b1\u03c4\u03b1\u03c1\u03ba\u03c4\u03b9\u03ba\u03ae. \u0395\u03c0\u03af\u03c3\u03b7\u03c2, \u03c4\u03b1 expert references \u03c5\u03c0\u03ac\u03c1\u03c7\u03bf\u03c5\u03bd \u03bc\u03cc\u03bd\u03bf \u03b3\u03b9\u03b1 \u03bf\u03ba\u03c4\u03ce \u03b1\u03c0\u03cc \u03c4\u03b1 17 priority workloads, \u03ac\u03c1\u03b1 \u03b4\u03b5\u03bd \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03bf\u03cd\u03bd \u03ba\u03b1\u03b8\u03bf\u03bb\u03b9\u03ba\u03cc \u03cc\u03c1\u03b9\u03bf \u03b3\u03b9\u03b1 \u03cc\u03bb\u03b7 \u03c4\u03b7 \u03c3\u03bf\u03c5\u03af\u03c4\u03b1.<\/p>\n<p>\u0397 \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b7 speedup \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af floor 1\u00d7 \u03c3\u03c4\u03b7\u03bd \u03bf\u03bc\u03b1\u03b4\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7: \u03ad\u03bd\u03b1\u03c2 \u03bb\u03b1\u03bd\u03b8\u03b1\u03c3\u03bc\u03ad\u03bd\u03bf\u03c2 \u03ae \u03c0\u03b9\u03bf \u03b1\u03c1\u03b3\u03cc\u03c2 kernel \u03b4\u03b5\u03bd \u03c1\u03af\u03c7\u03bd\u03b5\u03b9 \u03c4\u03bf\u03bd \u03b3\u03b5\u03c9\u03bc\u03b5\u03c4\u03c1\u03b9\u03ba\u03cc \u03bc\u03ad\u03c3\u03bf \u03ba\u03ac\u03c4\u03c9 \u03b1\u03c0\u03cc \u03c4\u03bf baseline. \u0391\u03c5\u03c4\u03cc \u03b5\u03af\u03bd\u03b1\u03b9 \u03bb\u03bf\u03b3\u03b9\u03ba\u03cc \u03b3\u03b9\u03b1 \u03c4\u03bf \u03c3\u03b5\u03bd\u03ac\u03c1\u03b9\u03bf \u03cc\u03c0\u03bf\u03c5 \u03c4\u03bf \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03ba\u03c1\u03b1\u03c4\u03ac \u03c4\u03bf\u03bd XLA kernel \u03b1\u03bd \u03bf \u03c5\u03c0\u03bf\u03c8\u03ae\u03c6\u03b9\u03bf\u03c2 \u03b4\u03b5\u03bd \u03b2\u03bf\u03b7\u03b8\u03ac, \u03b1\u03bb\u03bb\u03ac \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03bb\u03b1\u03bc\u03b2\u03ac\u03bd\u03b5\u03c4\u03b1\u03b9 \u03c5\u03c0\u03cc\u03c8\u03b7 \u03cc\u03c4\u03b1\u03bd \u03c3\u03c5\u03b3\u03ba\u03c1\u03af\u03bd\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c4\u03b1 aggregate \u03bd\u03bf\u03cd\u03bc\u03b5\u03c1\u03b1. \u0397 per-benchmark correctness \u03ba\u03b1\u03b9 \u03c4\u03bf fast@N \u03c3\u03c5\u03bc\u03c0\u03bb\u03b7\u03c1\u03ce\u03bd\u03bf\u03c5\u03bd \u03c4\u03b7\u03bd \u03b5\u03b9\u03ba\u03cc\u03bd\u03b1.<\/p>\n<h2 id=\"plaisio-agentic-optimization\">\u0388\u03bd\u03b1 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b9\u03ba\u03cc \u03c0\u03bb\u03b1\u03af\u03c3\u03b9\u03bf \u03b3\u03b9\u03b1 agentic optimization<\/h2>\n<p>\u039c\u03b9\u03b1 \u03c5\u03c0\u03b5\u03cd\u03b8\u03c5\u03bd\u03b7 \u03c5\u03bb\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b1\u03ba\u03bf\u03bb\u03bf\u03c5\u03b8\u03ae\u03c3\u03b5\u03b9 \u03c4\u03b7 \u03bb\u03bf\u03b3\u03b9\u03ba\u03ae \u03c4\u03bf\u03c5 JAXBench: \u03be\u03b5\u03ba\u03b9\u03bd\u03ac \u03bc\u03b5 production-relevant workloads, \u03bf\u03c1\u03af\u03b6\u03b5\u03b9 baseline \u03c3\u03c4\u03bf \u03af\u03b4\u03b9\u03bf hardware, \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af matched inputs \u03ba\u03b1\u03b9 \u03b1\u03bd\u03bf\u03c7\u03ad\u03c2 \u03ba\u03b1\u03c4\u03ac\u03bb\u03bb\u03b7\u03bb\u03b5\u03c2 \u03b3\u03b9\u03b1 \u03c4\u03bf\u03bd \u03b1\u03c1\u03b9\u03b8\u03bc\u03b7\u03c4\u03b9\u03ba\u03cc \u03c4\u03cd\u03c0\u03bf, \u03ba\u03ac\u03bd\u03b5\u03b9 warmup \u03ba\u03b1\u03b9 device-side profiling \u03ba\u03b1\u03b9 \u03ba\u03c1\u03b1\u03c4\u03ac \u03be\u03b5\u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03ac compilation, correctness \u03ba\u03b1\u03b9 latency outcomes. \u039f\u03b9 \u03c5\u03c0\u03bf\u03c8\u03ae\u03c6\u03b9\u03bf\u03b9 kernels \u03b4\u03b5\u03bd \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c0\u03b5\u03c1\u03bd\u03bf\u03cd\u03bd \u03c3\u03b5 \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae \u03c7\u03c9\u03c1\u03af\u03c2 \u03b5\u03c0\u03b1\u03bd\u03ac\u03bb\u03b7\u03c8\u03b7 \u03c4\u03c9\u03bd \u03b5\u03bb\u03ad\u03b3\u03c7\u03c9\u03bd \u03ba\u03b1\u03b9 \u03c3\u03cd\u03b3\u03ba\u03c1\u03b9\u03c3\u03b7 \u03bc\u03b5 \u03c4\u03bf \u03c5\u03c0\u03ac\u03c1\u03c7\u03bf\u03bd \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1. \u03a0\u03c1\u03cc\u03ba\u03b5\u03b9\u03c4\u03b1\u03b9 \u03b3\u03b9\u03b1 \u03c4\u03b7\u03bd \u03af\u03b4\u03b9\u03b1 \u03bb\u03bf\u03b3\u03b9\u03ba\u03ae \u03c0\u03bf\u03c5 \u03b1\u03c0\u03b1\u03b9\u03c4\u03b5\u03af \u03b7 <a href=\"https:\/\/twodots.gr\/epalithefsi-politikon-reinforcement-learning-ai-agent\/\">\u03b5\u03c0\u03b1\u03bb\u03ae\u03b8\u03b5\u03c5\u03c3\u03b7 \u03c0\u03bf\u03bb\u03b9\u03c4\u03b9\u03ba\u03ce\u03bd \u03c0\u03c1\u03b9\u03bd \u03b5\u03bc\u03c0\u03b9\u03c3\u03c4\u03b5\u03c5\u03c4\u03bf\u03cd\u03bc\u03b5 \u03ad\u03bd\u03b1\u03bd AI agent<\/a>.<\/p>\n<div class=\"td-step-list\">\n<p class=\"td-step-list-title\">\u0388\u03be\u03b9 \u03c0\u03cd\u03bb\u03b5\u03c2 \u03c0\u03c1\u03b9\u03bd \u03c0\u03b5\u03c1\u03ac\u03c3\u03b5\u03b9 \u03ad\u03bd\u03b1\u03c2 AI-generated TPU kernel \u03c3\u03b5 \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae<\/p>\n<ol>\n<li><span class=\"td-step-kicker\">Step 1<\/span><strong>\u0395\u03c0\u03b9\u03bb\u03ad\u03be\u03c4\u03b5 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc bottleneck<\/strong>\n<p>\u039a\u03ac\u03bd\u03c4\u03b5 profiling \u03c3\u03c4\u03bf production workload \u03ba\u03b1\u03b9 \u03b5\u03c0\u03b5\u03bd\u03b4\u03cd\u03c3\u03c4\u03b5 \u03bc\u03cc\u03bd\u03bf \u03b5\u03ba\u03b5\u03af \u03cc\u03c0\u03bf\u03c5 \u03c4\u03bf XLA \u03b1\u03c6\u03ae\u03bd\u03b5\u03b9 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf \u03c0\u03b5\u03c1\u03b9\u03b8\u03ce\u03c1\u03b9\u03bf \u03b2\u03b5\u03bb\u03c4\u03af\u03c9\u03c3\u03b7\u03c2.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Step 2<\/span><strong>\u039a\u03bb\u03b5\u03b9\u03b4\u03ce\u03c3\u03c4\u03b5 baseline \u03c3\u03c4\u03bf \u03af\u03b4\u03b9\u03bf TPU<\/strong>\n<p>\u03a7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03ae\u03c3\u03c4\u03b5 \u03af\u03b4\u03b9\u03b5\u03c2 \u03b5\u03b9\u03c3\u03cc\u03b4\u03bf\u03c5\u03c2, \u03c3\u03c7\u03ae\u03bc\u03b1\u03c4\u03b1, dtype \u03ba\u03b1\u03b9 hardware \u03ce\u03c3\u03c4\u03b5 \u03b7 \u03c3\u03cd\u03b3\u03ba\u03c1\u03b9\u03c3\u03b7 \u03bc\u03b5 \u03c4\u03bf\u03bd XLA kernel \u03bd\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b4\u03af\u03ba\u03b1\u03b9\u03b7.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Step 3<\/span><strong>\u0395\u03bb\u03ad\u03b3\u03be\u03c4\u03b5 \u03bc\u03b5\u03c4\u03b1\u03b3\u03bb\u03ce\u03c4\u03c4\u03b9\u03c3\u03b7 \u03ba\u03b1\u03b9 Pallas API<\/strong>\n<p>\u0391\u03c0\u03bf\u03c1\u03c1\u03af\u03c8\u03c4\u03b5 \u03b1\u03bd\u03cd\u03c0\u03b1\u03c1\u03ba\u03c4\u03b1 APIs, \u03bb\u03ac\u03b8\u03bf\u03c2 memory spaces \u03ba\u03b1\u03b9 block shapes \u03c0\u03bf\u03c5 \u03c0\u03b1\u03c1\u03b1\u03b2\u03b9\u03ac\u03b6\u03bf\u03c5\u03bd \u03c4\u03bf\u03c5\u03c2 \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03b9\u03c3\u03bc\u03bf\u03cd\u03c2 \u03c4\u03b7\u03c2 TPU.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Step 4<\/span><strong>\u0395\u03c0\u03b1\u03bb\u03b7\u03b8\u03b5\u03cd\u03c3\u03c4\u03b5 \u03b1\u03c1\u03b9\u03b8\u03bc\u03b7\u03c4\u03b9\u03ba\u03ae \u03bf\u03c1\u03b8\u03cc\u03c4\u03b7\u03c4\u03b1<\/strong>\n<p>\u03a3\u03c5\u03b3\u03ba\u03c1\u03af\u03bd\u03b5\u03c4\u03b5 \u03bc\u03b5 \u03c4\u03bf reference \u03c3\u03b5 matched inputs \u03ba\u03b1\u03b9 \u03b1\u03bd\u03bf\u03c7\u03ad\u03c2 \u03ba\u03b1\u03c4\u03ac\u03bb\u03bb\u03b7\u03bb\u03b5\u03c2 \u03b3\u03b9\u03b1 bf16 \u03c0\u03c1\u03b9\u03bd \u03b5\u03be\u03b5\u03c4\u03ac\u03c3\u03b5\u03c4\u03b5 \u03bf\u03c0\u03bf\u03b9\u03bf\u03b4\u03ae\u03c0\u03bf\u03c4\u03b5 speedup.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Step 5<\/span><strong>\u039c\u03b5\u03c4\u03c1\u03ae\u03c3\u03c4\u03b5 device-side \u03c7\u03c1\u03cc\u03bd\u03bf<\/strong>\n<p>\u039a\u03ac\u03bd\u03c4\u03b5 warmup, \u03c3\u03c5\u03bb\u03bb\u03ad\u03be\u03c4\u03b5 profiler traces \u03ba\u03b1\u03b9 \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03ae\u03c3\u03c4\u03b5 \u03c3\u03c4\u03b1\u03b8\u03b5\u03c1\u03cc \u03c3\u03c4\u03b1\u03c4\u03b9\u03c3\u03c4\u03b9\u03ba\u03cc \u03ce\u03c3\u03c4\u03b5 host overhead \u03ba\u03b1\u03b9 dispatch \u03bd\u03b1 \u03bc\u03b7 \u03bd\u03bf\u03b8\u03b5\u03cd\u03bf\u03c5\u03bd \u03c4\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Step 6<\/span><strong>\u03a0\u03c1\u03bf\u03c9\u03b8\u03ae\u03c3\u03c4\u03b5 \u03bc\u03cc\u03bd\u03bf \u03c4\u03bf\u03bd \u03b5\u03c0\u03b1\u03bb\u03b7\u03b8\u03b5\u03c5\u03bc\u03ad\u03bd\u03bf \u03bd\u03b9\u03ba\u03b7\u03c4\u03ae<\/strong>\n<p>\u039a\u03c1\u03b1\u03c4\u03ae\u03c3\u03c4\u03b5 selection policy, audit trail \u03ba\u03b1\u03b9 \u03ac\u03bc\u03b5\u03c3\u03bf fallback \u03c3\u03c4\u03bf\u03bd \u03c5\u03c0\u03ac\u03c1\u03c7\u03bf\u03bd\u03c4\u03b1 XLA kernel \u03cc\u03c4\u03b1\u03bd \u03bf \u03c5\u03c0\u03bf\u03c8\u03ae\u03c6\u03b9\u03bf\u03c2 \u03b1\u03c0\u03bf\u03c4\u03c5\u03b3\u03c7\u03ac\u03bd\u03b5\u03b9 \u03ae \u03c0\u03b1\u03bb\u03b9\u03bd\u03b4\u03c1\u03bf\u03bc\u03b5\u03af.<\/p>\n<\/li>\n<\/ol>\n<\/div>\n<p>\u0397 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03b4\u03b5\u03bd \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03ba\u03bd\u03cd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03bf\u03b9 AI agents \u03b1\u03bd\u03c4\u03b9\u03ba\u03b1\u03b8\u03b9\u03c3\u03c4\u03bf\u03cd\u03bd \u03c4\u03bf\u03c5\u03c2 kernel engineers. \u03a4\u03b5\u03ba\u03bc\u03b7\u03c1\u03b9\u03ce\u03bd\u03b5\u03b9 \u03cc\u03bc\u03c9\u03c2 \u03cc\u03c4\u03b9 \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03bd\u03b1 \u03b1\u03bd\u03b1\u03ba\u03c4\u03ae\u03c3\u03bf\u03c5\u03bd \u03c3\u03b7\u03bc\u03b1\u03bd\u03c4\u03b9\u03ba\u03cc \u03bc\u03ad\u03c1\u03bf\u03c2 \u03c4\u03b7\u03c2 \u03b1\u03bd\u03b8\u03c1\u03ce\u03c0\u03b9\u03bd\u03b7\u03c2 \u03b2\u03b5\u03bb\u03c4\u03b9\u03c3\u03c4\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7\u03c2 \u03cc\u03c4\u03b1\u03bd \u03ad\u03c7\u03bf\u03c5\u03bd \u03c3\u03c9\u03c3\u03c4\u03cc context \u03ba\u03b1\u03b9 \u03bf\u03c1\u03b3\u03b1\u03bd\u03c9\u03bc\u03ad\u03bd\u03b7 \u03b1\u03bd\u03b1\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7. \u03a4\u03bf \u03c0\u03b9\u03bf \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03c3\u03c5\u03bc\u03c0\u03ad\u03c1\u03b1\u03c3\u03bc\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03cc\u03c4\u03b9 \u03b7 \u03b1\u03be\u03af\u03b1 \u03b2\u03c1\u03af\u03c3\u03ba\u03b5\u03c4\u03b1\u03b9 \u03c3\u03c4\u03bf \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03b3\u03cd\u03c1\u03c9 \u03b1\u03c0\u03cc \u03c4\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf: benchmark, \u03c4\u03b5\u03ba\u03bc\u03b7\u03c1\u03af\u03c9\u03c3\u03b7, feedback, selection policy, \u03b1\u03c3\u03c6\u03b1\u03bb\u03ad\u03c2 fallback \u03ba\u03b1\u03b9 \u03c3\u03b1\u03c6\u03ae\u03c2 <a href=\"https:\/\/twodots.gr\/ai-agents-poios-ftaiei-otan-aftomatopoiisi-apotygchanei\/\">\u03b5\u03c5\u03b8\u03cd\u03bd\u03b7 \u03cc\u03c4\u03b1\u03bd \u03b7 \u03b1\u03c5\u03c4\u03bf\u03bc\u03b1\u03c4\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 \u03bc\u03b5 AI agents \u03b1\u03c0\u03bf\u03c4\u03c5\u03b3\u03c7\u03ac\u03bd\u03b5\u03b9<\/a>.<\/p>\n<div class=\"td-decision-band\">\n<p class=\"td-decision-label\">\u03a4\u03bf \u03ba\u03c1\u03b9\u03c4\u03ae\u03c1\u03b9\u03bf \u03b3\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03c3\u03b9\u03b1\u03ba\u03ae \u03c7\u03c1\u03ae\u03c3\u03b7<\/p>\n<p><strong>\u0388\u03bd\u03b1\u03c2 \u03b3\u03c1\u03ae\u03b3\u03bf\u03c1\u03bf\u03c2 kernel \u03ad\u03c7\u03b5\u03b9 \u03b1\u03be\u03af\u03b1 \u03bc\u03cc\u03bd\u03bf \u03b1\u03c6\u03bf\u03cd \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03c7\u03b8\u03b5\u03af \u03c3\u03c9\u03c3\u03c4\u03cc\u03c2 \u03ba\u03b1\u03b9 \u03bc\u03b5\u03c4\u03c1\u03b7\u03b8\u03b5\u03af \u03c3\u03c4\u03bf hardware \u03cc\u03c0\u03bf\u03c5 \u03b8\u03b1 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03ae\u03c3\u03b5\u03b9.<\/strong><\/p>\n<p>\u03a4\u03bf JAXBench \u03bc\u03b5\u03c4\u03b1\u03c6\u03ad\u03c1\u03b5\u03b9 \u03c4\u03bf \u03b2\u03ac\u03c1\u03bf\u03c2 \u03b1\u03c0\u03cc \u03c4\u03bf \u03b5\u03bd\u03c4\u03c5\u03c0\u03c9\u03c3\u03b9\u03b1\u03ba\u03cc code generation \u03c3\u03b5 \u03ad\u03bd\u03b1 \u03b5\u03bb\u03b5\u03b3\u03c7\u03cc\u03bc\u03b5\u03bd\u03bf \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1: \u03ad\u03b3\u03ba\u03c5\u03c1\u03b7 \u03c4\u03b5\u03ba\u03bc\u03b7\u03c1\u03af\u03c9\u03c3\u03b7, reproducible harness, device-side profiling, selection policy \u03ba\u03b1\u03b9 \u03b1\u03c3\u03c6\u03b1\u03bb\u03ad\u03c2 fallback.<\/p>\n<\/div>\n<section class=\"td-service-cta\">\n<div class=\"td-service-cta-content\">\n<p class=\"td-service-cta-eyebrow\">Business Automation &amp; AI by TWO DOTS<\/p>\n<p class=\"td-service-cta-title\">\u039c\u03b5\u03c4\u03b1\u03c4\u03c1\u03ad\u03c8\u03c4\u03b5 \u03c4\u03bf AI optimization \u03b1\u03c0\u03cc \u03c0\u03b5\u03af\u03c1\u03b1\u03bc\u03b1 \u03c3\u03b5 \u03b5\u03bb\u03b5\u03b3\u03c7\u03cc\u03bc\u03b5\u03bd\u03bf \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03c3\u03b9\u03b1\u03ba\u03cc workflow.<\/p>\n<p>\u0397 TWO DOTS \u03c3\u03c7\u03b5\u03b4\u03b9\u03ac\u03b6\u03b5\u03b9 AI \u03b1\u03c5\u03c4\u03bf\u03bc\u03b1\u03c4\u03b9\u03c3\u03bc\u03bf\u03cd\u03c2 \u03bc\u03b5 \u03c4\u03b5\u03ba\u03bc\u03b7\u03c1\u03b9\u03c9\u03bc\u03ad\u03bd\u03b1 baselines, validation gates, monitoring, \u03c3\u03b1\u03c6\u03ae \u03b4\u03b9\u03ba\u03b1\u03b9\u03ce\u03bc\u03b1\u03c4\u03b1 \u03ba\u03b1\u03b9 \u03b1\u03c3\u03c6\u03b1\u03bb\u03ae fallback \u03b3\u03b9\u03b1 \u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03b5\u03c2 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03af\u03b5\u03c2.<\/p>\n<div class=\"td-service-cta-actions\"><a class=\"td-service-cta-button\" href=\"https:\/\/twodots.gr\/aftomatismoi-epicheiriseon-ai\/\">Check out Business Automation &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\">Frequently Asked Questions (FAQs)<\/p>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a4\u03b9 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c4\u03bf JAXBench;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u0395\u03af\u03bd\u03b1\u03b9 \u03b1\u03bd\u03bf\u03b9\u03ba\u03c4\u03ae benchmark suite 50 JAX workloads \u03b3\u03b9\u03b1 \u03c4\u03b7\u03bd \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7 AI-generated Pallas kernels \u03c3\u03b5 Google Cloud TPU v6e, \u03bc\u03b5 \u03be\u03b5\u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03bf\u03cd\u03c2 \u03b5\u03bb\u03ad\u03b3\u03c7\u03bf\u03c5\u03c2 \u03bc\u03b5\u03c4\u03b1\u03b3\u03bb\u03ce\u03c4\u03c4\u03b9\u03c3\u03b7\u03c2, \u03bf\u03c1\u03b8\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2 \u03ba\u03b1\u03b9 \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\">\u0393\u03b9\u03b1\u03c4\u03af \u03b4\u03b5\u03bd \u03b1\u03c1\u03ba\u03b5\u03af \u03ad\u03bd\u03b1 benchmark \u03b3\u03b9\u03b1 GPU;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u039f\u03b9 TPU \u03ad\u03c7\u03bf\u03c5\u03bd \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ae \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae, memory hierarchy \u03ba\u03b1\u03b9 software stack. \u03a4\u03b1 workloads, \u03c4\u03b1 block constraints \u03ba\u03b1\u03b9 \u03bf\u03b9 \u03c4\u03b5\u03c7\u03bd\u03b9\u03ba\u03ad\u03c2 \u03b2\u03b5\u03bb\u03c4\u03b9\u03c3\u03c4\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7\u03c2 \u03b4\u03b5\u03bd \u03bc\u03b5\u03c4\u03b1\u03c6\u03ad\u03c1\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b1\u03c5\u03c4\u03bf\u03cd\u03c3\u03b9\u03b1 \u03b1\u03c0\u03cc CUDA \u03ae Triton.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a0\u03bf\u03b9\u03bf \u03ae\u03c4\u03b1\u03bd \u03c4\u03bf \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03c3\u03c4\u03bf \u03c0\u03bb\u03ae\u03c1\u03b5\u03c2 JAXBench;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u039c\u03b5 Gemini 3 Flash, \u03c4\u03bf Autocomp \u03c0\u03ad\u03c4\u03c5\u03c7\u03b5 \u03b3\u03b5\u03c9\u03bc\u03b5\u03c4\u03c1\u03b9\u03ba\u03cc \u03bc\u03ad\u03c3\u03bf speedup 1,36\u00d7 \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 XLA \u03ba\u03b1\u03b9 \u03c3\u03c9\u03c3\u03c4\u03bf\u03cd\u03c2 kernels \u03c3\u03b5 45 \u03b1\u03c0\u03cc \u03c4\u03b1 50 workloads.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a4\u03b9 \u03c0\u03c1\u03cc\u03c3\u03c6\u03b5\u03c1\u03b5 \u03c4\u03bf \u03b5\u03be\u03b5\u03b9\u03b4\u03b9\u03ba\u03b5\u03c5\u03bc\u03ad\u03bd\u03bf TPU context;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u03a3\u03c4\u03b7\u03bd iterative \u03bc\u03ad\u03b8\u03bf\u03b4\u03bf \u03b1\u03cd\u03be\u03b7\u03c3\u03b5 \u03c4\u03b7\u03bd \u03bf\u03c1\u03b8\u03cc\u03c4\u03b7\u03c4\u03b1 \u03b1\u03bd\u03ac sample \u03b1\u03c0\u03cc 5,8% \u03c3\u03b5 37,3% \u03c7\u03c9\u03c1\u03af\u03c2 \u03b1\u03bb\u03bb\u03b1\u03b3\u03ae \u03c4\u03bf\u03c5 \u03b1\u03bb\u03b3\u03bf\u03c1\u03af\u03b8\u03bc\u03bf\u03c5 \u03b1\u03bd\u03b1\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7\u03c2 \u03ba\u03b1\u03b9 \u03bf\u03b4\u03ae\u03b3\u03b7\u03c3\u03b5 \u03c3\u03b5 \u03c3\u03c9\u03c3\u03c4\u03bf\u03cd\u03c2 kernels \u03b3\u03b9\u03b1 48 \u03b1\u03c0\u03cc \u03c4\u03b1 50 workloads.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u039f\u03b9 AI agents \u03be\u03b5\u03c0\u03ad\u03c1\u03b1\u03c3\u03b1\u03bd \u03c4\u03bf\u03c5\u03c2 hand-tuned kernels;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u03a3\u03b5 \u03b4\u03cd\u03bf \u03b1\u03c0\u03cc \u03c4\u03b1 \u03bf\u03ba\u03c4\u03ce \u03c3\u03c5\u03b3\u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03b1 workloads \u03c4\u03bf Autocomp \u03be\u03b5\u03c0\u03ad\u03c1\u03b1\u03c3\u03b5 \u03c4\u03bf \u03b4\u03b9\u03b1\u03b8\u03ad\u03c3\u03b9\u03bc\u03bf Tokamax reference, \u03b1\u03bb\u03bb\u03ac \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03ac \u03ad\u03bc\u03b5\u03b9\u03bd\u03b5 \u03c7\u03b1\u03bc\u03b7\u03bb\u03cc\u03c4\u03b5\u03c1\u03b1 \u03ba\u03b1\u03b9 \u03b1\u03c0\u03ad\u03c4\u03c5\u03c7\u03b5 \u03c3\u03b5 \u03b5\u03be\u03b5\u03b9\u03b4\u03b9\u03ba\u03b5\u03c5\u03bc\u03ad\u03bd\u03b1 paged-attention cases.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u0393\u03b9\u03b1\u03c4\u03af \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af\u03c4\u03b1\u03b9 \u03c4\u03bf XLA \u03c9\u03c2 baseline;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u0395\u03c0\u03b5\u03b9\u03b4\u03ae \u03b5\u03af\u03bd\u03b1\u03b9 \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc \u03ba\u03b1\u03b9 \u03b1\u03bd\u03b1\u03c0\u03b1\u03c1\u03b1\u03b3\u03ce\u03b3\u03b9\u03bc\u03bf baseline \u03b3\u03b9\u03b1 JAX \u03c3\u03b5 TPU. \u0391\u03bd \u03bf \u03bd\u03ad\u03bf\u03c2 kernel \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03c3\u03c9\u03c3\u03c4\u03cc\u03c2 \u03ae \u03c4\u03b1\u03c7\u03cd\u03c4\u03b5\u03c1\u03bf\u03c2, \u03c4\u03bf \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03ae\u03c3\u03b5\u03b9 \u03c4\u03bf\u03bd XLA kernel \u03c9\u03c2 \u03b1\u03c3\u03c6\u03b1\u03bb\u03ad\u03c2 fallback.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a0\u03bf\u03b9\u03b1 metrics \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bc\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03af\u03c1\u03b7\u03c3\u03b7;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u03a4\u03bf\u03c5\u03bb\u03ac\u03c7\u03b9\u03c3\u03c4\u03bf\u03bd compilability, numerical correctness, latency \u03ae throughput \u03c3\u03c4\u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc hardware, \u03c3\u03cd\u03b3\u03ba\u03c1\u03b9\u03c3\u03b7 \u03bc\u03b5 production baseline, \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03b1\u03bd\u03b1\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7\u03c2 \u03ba\u03b1\u03b9 sample budget.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u0395\u03af\u03bd\u03b1\u03b9 \u03ad\u03c4\u03bf\u03b9\u03bc\u03b7 \u03b7 \u03b1\u03c5\u03c4\u03cc\u03bc\u03b1\u03c4\u03b7 \u03b2\u03b5\u03bb\u03c4\u03b9\u03c3\u03c4\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 TPU kernels \u03b3\u03b9\u03b1 \u03b1\u03bd\u03b5\u03be\u03ad\u03bb\u03b5\u03b3\u03ba\u03c4\u03b7 \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u038c\u03c7\u03b9 \u03c3\u03cd\u03bc\u03c6\u03c9\u03bd\u03b1 \u03bc\u03b5 \u03c4\u03b1 failure modes \u03c4\u03b7\u03c2 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7\u03c2. \u03a7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03b1\u03c5\u03c3\u03c4\u03b7\u03c1\u03cc evaluation harness, \u03ad\u03b3\u03ba\u03c5\u03c1\u03b7 \u03c4\u03b5\u03ba\u03bc\u03b7\u03c1\u03af\u03c9\u03c3\u03b7, \u03ad\u03bb\u03b5\u03b3\u03c7\u03bf\u03c2 \u03bf\u03c1\u03b8\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2, profiling \u03ba\u03b1\u03b9 fallback \u03c3\u03c4\u03bf\u03bd \u03c5\u03c0\u03ac\u03c1\u03c7\u03bf\u03bd\u03c4\u03b1 kernel.<\/p>\n<\/div>\n<\/details>\n<\/div>\n<\/section>\n<div class=\"td-source-list\">\n<p id=\"piges\" class=\"td-source-list-title\">Sources<\/p>\n<ul>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2607.20466\" target=\"_blank\" rel=\"noopener\">Tschand et al.: JAXBench \u2014 Benchmarking Autonomous TPU Kernel Optimization<\/a><\/li>\n<li><a href=\"https:\/\/docs.jax.dev\/en\/latest\/pallas\/\" target=\"_blank\" rel=\"noopener\">JAX Documentation: Pallas kernel language<\/a><\/li>\n<li><a href=\"https:\/\/docs.jax.dev\/en\/latest\/pallas\/tpu\/\" target=\"_blank\" rel=\"noopener\">JAX Documentation: Pallas \u03b3\u03b9\u03b1 TPU<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/ucb-bar\/autocomp\" target=\"_blank\" rel=\"noopener\">UC Berkeley Architecture Research: Autocomp<\/a><\/li>\n<\/ul>\n<\/div>","protected":false},"excerpt":{"rendered":"<p>\u03a4\u03bf JAXBench \u03b4\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03c0\u03ce\u03c2 AI agents \u03b2\u03b5\u03bb\u03c4\u03b9\u03c3\u03c4\u03bf\u03c0\u03bf\u03b9\u03bf\u03cd\u03bd TPU kernels \u03ba\u03b1\u03b9 \u03b3\u03b9\u03b1\u03c4\u03af \u03bf\u03c1\u03b8\u03cc\u03c4\u03b7\u03c4\u03b1, Pallas context \u03ba\u03b1\u03b9 device-side profiling \u03c0\u03c1\u03bf\u03b7\u03b3\u03bf\u03cd\u03bd\u03c4\u03b1\u03b9 \u03c4\u03bf\u03c5 speedup.<\/p>","protected":false},"author":1,"featured_media":87593,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"_gspb_post_css":"","content-type":"","footnotes":""},"categories":[17366],"tags":[6506,6810,18920,18921,3597],"class_list":["post-87096","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-digital-marketing","tag-ai-agents","tag-cloud-computing","tag-jax","tag-tpu","tag-techniti-noimosyni"],"blocksy_meta":[],"_links":{"self":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/posts\/87096","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/comments?post=87096"}],"version-history":[{"count":0,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/posts\/87096\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/media\/87593"}],"wp:attachment":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/media?parent=87096"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/categories?post=87096"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/tags?post=87096"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}