{"id":11356,"date":"2026-10-04T00:29:00","date_gmt":"2026-10-04T00:29:00","guid":{"rendered":"https:\/\/bitep.net\/blog\/2026\/10\/04\/microsoftcom\/"},"modified":"2026-10-04T00:29:00","modified_gmt":"2026-10-04T00:29:00","slug":"microsoftcom","status":"publish","type":"post","link":"https:\/\/bitep.net\/blog\/2026\/10\/04\/microsoftcom\/","title":{"rendered":"microsoft.com"},"content":{"rendered":"<h2>Hesablama G\u00fcc\u00fcn\u00fcn Yeni M\u0259rh\u0259l\u0259si: &#8220;AI Superfactories&#8221; v\u0259 M\u0259qs\u0259dli Terapevtika<\/h2>\n<p>S\u00fcni intellekt modell\u0259rinin miqyas\u0131 geni\u015fl\u0259ndikc\u0259, \u0259n\u0259n\u0259vi hesablama infrastrukturlar\u0131 artan t\u0259l\u0259bl\u0259ri \u00f6d\u0259m\u0259kd\u0259 \u00e7\u0259tinlik \u00e7\u0259kir. Microsoft ba\u015fda olmaqla qlobal texnologiya n\u0259h\u0259ngl\u0259ri, h\u0259m geni\u015fmiqyasl\u0131 modell\u0259rin t\u0259limi (training), h\u0259m d\u0259 a\u015fa\u011f\u0131 gecikm\u0259li n\u0259tic\u0259\u00e7\u0131xarma (inference) prosesl\u0259rini t\u0259min etm\u0259k \u00fc\u00e7\u00fcn paylanm\u0131\u015f superhesablama \u015f\u0259b\u0259k\u0259l\u0259rin\u0259 \u2014 mahiyy\u0259tc\u0259 &#8220;AI Superfactories&#8221; (S\u00fcni \u0130ntellekt Superfabrikalar\u0131) adland\u0131r\u0131lan qlobal infrastruktur vahidl\u0259rin\u0259 ke\u00e7id edir. Bu infrastruktur transformasiyas\u0131 x\u0259rcl\u0259ri k\u0259skin \u015f\u0259kild\u0259 a\u015fa\u011f\u0131 salmaqla yana\u015f\u0131, a\u00e7\u0131q elmi t\u0259\u015f\u0259bb\u00fcsl\u0259r v\u0259 ixtisasla\u015fm\u0131\u015f modell\u0259r (m\u0259s\u0259l\u0259n, MedGemma variantlar\u0131) vasit\u0259sil\u0259 onkologiya v\u0259 m\u0259qs\u0259dli terapevtika sah\u0259sind\u0259 inqilabi add\u0131mlar\u0131n at\u0131lmas\u0131na \u015f\u0259rait yarad\u0131r.<\/p>\n<h2>Qlobal Paylanm\u0131\u015f \u0130nfrastruktur v\u0259 &#8220;AI Superfabrika&#8221; Konsepsiyas\u0131<\/h2>\n<p>T\u0259k bir m\u0259rk\u0259zd\u0259 minl\u0259rl\u0259 akseleratoru (GPU\/NPU) c\u0259ml\u0259m\u0259k enerji paylanmas\u0131, soyutma v\u0259 \u015f\u0259b\u0259k\u0259 gecikm\u0259l\u0259ri bax\u0131m\u0131ndan fiziki m\u0259hdudiyy\u0259tl\u0259rl\u0259 qar\u015f\u0131la\u015f\u0131r. &#8220;AI Superfactory&#8221; yana\u015fmas\u0131 co\u011frafi c\u0259h\u0259td\u0259n paylanm\u0131\u015f, lakin ultra-y\u00fcks\u0259k burax\u0131l\u0131\u015f qabiliyy\u0259tin\u0259 malik optik magistrallarla bir-birin\u0259 ba\u011flanm\u0131\u015f n\u0259h\u0259ng hesablama klasterl\u0259rini t\u0259k bir vahid orkestrasiya qat\u0131 alt\u0131nda birl\u0259\u015fdirir.<\/p>\n<h3>\u0130nfrastrukturun Texniki X\u00fcsusiyy\u0259tl\u0259ri<\/h3>\n<ul>\n<li><strong>Y\u00fcks\u0259k S\u0131xl\u0131ql\u0131 Klasterl\u0259r:<\/strong> Minl\u0259rl\u0259 qabaqc\u0131l qrafik prosessorun (m\u0259s\u0259l\u0259n, NVIDIA H100\/H200 v\u0259 ya f\u0259rdi silikon \u00e7ipl\u0259r) maye soyutma texnologiyalar\u0131 il\u0259 t\u0259chiz edilmi\u015f server \u015fkaflar\u0131nda konsentrasiyas\u0131.<\/li>\n<li><strong>Qabaqc\u0131l \u015e\u0259b\u0259k\u0259 Qar\u015f\u0131l\u0131ql\u0131 \u018flaq\u0259si:<\/strong> Klasterl\u0259r daxilind\u0259 InfiniBand (Quantum-2) v\u0259 RoCE v2 (RDMA over Converged Ethernet) protokollar\u0131 il\u0259 mikrosaniy\u0259 s\u0259viyy\u0259sind\u0259 gecikm\u0259l\u0259rin v\u0259 3.2 Tbps-\u0259 q\u0259d\u0259r qov\u015faqlararas\u0131 \u00f6t\u00fcrm\u0259nin t\u0259mini.<\/li>\n<li><strong>X\u0259rcl\u0259rin Optimizasiyas\u0131:<\/strong> Dinamik i\u015f y\u00fck\u00fc balans\u0131 v\u0259 paylanm\u0131\u015f orkestrasiya (m\u0259s\u0259l\u0259n, Ray v\u0259 Kubernetes \u0259sasl\u0131 paylanm\u0131\u015f m\u00fchitl\u0259r) say\u0259sind\u0259 serverl\u0259rin bo\u015fdayanma vaxt\u0131 minimuma endirilir v\u0259 n\u0259tic\u0259\u00e7\u0131xarma x\u0259rcl\u0259ri k\u0259skin azal\u0131r.<\/li>\n<\/ul>\n<table>\n<thead>\n<tr>\n<th>X\u00fcsusiyy\u0259t<\/th>\n<th>\u018fn\u0259n\u0259vi HPC Klasterl\u0259ri<\/th>\n<th>M\u00fcasir &#8220;AI Superfactories&#8221;<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u015e\u0259b\u0259k\u0259 Topologiyas\u0131<\/td>\n<td>Lokal Fat-Tree v\u0259 ya Torus<\/td>\n<td>Qlobal Paylanm\u0131\u015f, \u00c7oxqatl\u0131 Ultra-Low Latency Fabric<\/td>\n<\/tr>\n<tr>\n<td>Soyutma Arxitekturas\u0131<\/td>\n<td>\u018fsas\u0259n Hava Soyutmas\u0131<\/td>\n<td>Birba\u015fa \u00c7ip\u0259 Maye Soyutma (Direct-to-Chip Liquid Cooling)<\/td>\n<\/tr>\n<tr>\n<td>\u0130\u015f Y\u00fck\u00fc \u0130dar\u0259etm\u0259si<\/td>\n<td>Statik \u0130\u015f N\u00f6vb\u0259l\u0259ri (SLURM)<\/td>\n<td>Dinamik, Paylanm\u0131\u015f Avtomatik Miqyaslanan Orkestrasiya<\/td>\n<\/tr>\n<tr>\n<td>Enerji Effektivliyi (PUE)<\/td>\n<td>1.3 \u2013 1.6<\/td>\n<td>1.1 \u2013 1.2 h\u0259d\u0259fli optimalla\u015fd\u0131r\u0131lm\u0131\u015f arxitektura<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h2>A\u00e7\u0131q Elm v\u0259 M\u0259qs\u0259dli Terapevtika: X\u0259r\u00e7\u0259ng T\u0259dqiqatlar\u0131nda S\u0131\u00e7ray\u0131\u015f<\/h2>\n<p>Paylanm\u0131\u015f hesablama g\u00fcc\u00fcn\u00fcn yaratd\u0131\u011f\u0131 resurs \u0259l\u00e7atanl\u0131\u011f\u0131 biotibb sah\u0259sind\u0259 fundamental d\u0259yi\u015fiklikl\u0259r\u0259 yol a\u00e7\u0131r. X\u00fcsusil\u0259 MedGemma tipli tibbi y\u00f6n\u00fcml\u00fc a\u00e7\u0131q m\u0259nb\u0259li t\u0259m\u0259l modell\u0259r (foundation models) klinik t\u0259dqiqat boru k\u0259m\u0259rl\u0259rini (pipelines) s\u00fcr\u0259tl\u0259ndirir.<\/p>\n<p>Bu modell\u0259r onkologiyada immun h\u00fcceyr\u0259l\u0259rinin (m\u0259s\u0259l\u0259n, CAR-T h\u00fcceyr\u0259 terapiyas\u0131 \u00fc\u00e7\u00fcn) genetik modifikasiyas\u0131n\u0131 simulyasiya edir. \u0130ll\u0259rl\u0259 davam ed\u0259n laboratoriya s\u0131naqlar\u0131 \u0259v\u0259zin\u0259, molekulyar s\u0259viyy\u0259d\u0259 h\u00fcceyr\u0259-antigen qar\u015f\u0131l\u0131ql\u0131 t\u0259siri s\u00fcni intellekt klasterl\u0259rind\u0259 in-silico simulyasiya olunur. Bu yana\u015fma h\u0259m yeni terapevtik metodlar\u0131n haz\u0131rlanma x\u0259rcl\u0259rini azald\u0131r, h\u0259m d\u0259 qlobal s\u0259hiyy\u0259 sisteml\u0259ri aras\u0131ndak\u0131 texnoloji u\u00e7urumu aradan qald\u0131r\u0131r.<\/p>\n<h2>Praktiki T\u0259tbiq: Biotibbi Modelin Paylanm\u0131\u015f M\u00fchitd\u0259 \u0130cra Edilm\u0259si<\/h2>\n<p>A\u00e7\u0131q m\u0259nb\u0259li biotibbi modell\u0259rin (m\u0259s\u0259l\u0259n, Hugging Face platformas\u0131nda yay\u0131mlanan ixtisasla\u015fm\u0131\u015f modell\u0259r) t\u0259dqiqat m\u0259qs\u0259dl\u0259ri \u00fc\u00e7\u00fcn paylanm\u0131\u015f GPU m\u00fchitind\u0259 qura\u015fd\u0131r\u0131lmas\u0131 v\u0259 n\u0259tic\u0259\u00e7\u0131xarma (inference) m\u0259rh\u0259l\u0259sinin ba\u015flad\u0131lmas\u0131 a\u015fa\u011f\u0131dak\u0131 add\u0131mlarla h\u0259yata ke\u00e7irilir.<\/p>\n<h3>Add\u0131m 1: T\u0259l\u0259b Olunan M\u00fchitin Haz\u0131rlanmas\u0131<\/h3>\n<p>\u018fvv\u0259lc\u0259 y\u00fcks\u0259k hesablama s\u00fcr\u0259ti t\u0259min ed\u0259n <code>vLLM<\/code> m\u00fch\u0259rrikini v\u0259 z\u0259ruri kitabxanalar\u0131 qura\u015fd\u0131r\u0131n:<\/p>\n<pre><code class=\"language-bash\"># Virtual m\u00fchitin yarad\u0131lmas\u0131 v\u0259 aktivl\u0259\u015fdirilm\u0259si\npython3 -m venv bio_ai_env\nsource bio_ai_env\/bin\/activate\n\n# Laz\u0131mi as\u0131l\u0131l\u0131qlar\u0131n qura\u015fd\u0131r\u0131lmas\u0131\npip install --upgrade pip\npip install torch torchvision torchaudio --index-url https:\/\/download.pytorch.org\/whl\/cu121\npip install vllm transformers accelerate<\/code><\/pre>\n<h3>Add\u0131m 2: Modelin \u00c7oxsayl\u0131 GPU-larda Paralell\u0259\u015fdirilm\u0259si v\u0259 \u0130\u015f\u0259 Sal\u0131nmas\u0131<\/h3>\n<p>A\u015fa\u011f\u0131dak\u0131 Python skripti modeli tensor paralelliyi (Tensor Parallelism) vasit\u0259sil\u0259 bir ne\u00e7\u0259 GPU-ya paylay\u0131r v\u0259 klinik m\u0259tn \u00fczr\u0259 analiz apar\u0131r:<\/p>\n<pre><code class=\"language-bash\">cat &lt;&lt; 'EOF' &gt; run_biomed_inference.py\nfrom vllm import LLM, SamplingParams\n\n# \u00c7oxlu GPU istifad\u0259si \u00fc\u00e7\u00fcn Tensor Parallelism d\u0259r\u0259c\u0259sinin t\u0259yini\ntensor_parallel_size = 2\n\n# Modelin y\u00fckl\u0259nm\u0259si (N\u00fcmun\u0259vi biotibbi model arxitekturas\u0131)\nllm = LLM(\n    model=\"google\/medgemma-2b\", \n    tensor_parallel_size=tensor_parallel_size,\n    trust_remote_code=True\n)\n\n# N\u0259tic\u0259\u00e7\u0131xarma parametrl\u0259ri\nsampling_params = SamplingParams(\n    temperature=0.2, \n    top_p=0.95, \n    max_tokens=256\n)\n\nprompts = [\n    \"Klinik analiz: EGFR mutasiyas\u0131 a\u015fkarlanm\u0131\u015f a\u011fciy\u0259r x\u0259r\u00e7\u0259ngi x\u0259st\u0259sind\u0259 h\u0259d\u0259fli terapevtik inhibitorlar\u0131n t\u0259sir mexanizmini izah edin.\"\n]\n\n# N\u0259tic\u0259l\u0259rin hesablanmas\u0131\noutputs = llm.generate(prompts, sampling_params)\n\nfor output in outputs:\n    prompt = output.prompt\n    generated_text = output.outputs[0].text\n    print(f\"\\nSual: {prompt}\\nN\u0259tic\u0259:\\n{generated_text}\")\nEOF\n\npython3 run_biomed_inference.py<\/code><\/pre>\n<h2>S\u0259nayey\u0259 v\u0259 \u0130stifad\u0259\u00e7il\u0259r\u0259 T\u0259siri<\/h2>\n<p>Paylanm\u0131\u015f hesablama infrastrukturunun geni\u015fl\u0259nm\u0259si v\u0259 a\u00e7\u0131q elmi modell\u0259rin biotibbi sah\u0259y\u0259 inteqrasiyas\u0131 bir s\u0131ra strateji n\u0259tic\u0259l\u0259r do\u011furur:<\/p>\n<ul>\n<li><strong>\u018fm\u0259liyyat X\u0259rcl\u0259rinin D\u00fc\u015fm\u0259si:<\/strong> &#8220;AI Superfactories&#8221; vasit\u0259sil\u0259 hesablama resurslar\u0131n\u0131n daha s\u0259m\u0259r\u0259li idar\u0259 olunmas\u0131 b\u00f6y\u00fck modell\u0259rin h\u0259r token \u00fczr\u0259 emal qiym\u0259tini azald\u0131r. Bu, startaplar v\u0259 elmi laboratoriyalar \u00fc\u00e7\u00fcn giri\u015f baryerini a\u015fa\u011f\u0131 sal\u0131r.<\/li>\n<li><strong>Klinik T\u0259dqiqatlar\u0131n S\u00fcr\u0259tl\u0259nm\u0259si:<\/strong> Yeni d\u0259rman molekullar\u0131n\u0131n k\u0259\u015ffi v\u0259 x\u0259r\u00e7\u0259ng h\u00fcceyr\u0259l\u0259rini h\u0259d\u0259fl\u0259y\u0259n immunoterapiya protokollar\u0131n\u0131n haz\u0131rlanma m\u00fcdd\u0259ti ill\u0259rd\u0259n aylara enir.<\/li>\n<li><strong>A\u00e7\u0131q Elm vasit\u0259sil\u0259 \u015e\u0259ffafl\u0131q:<\/strong> Qapal\u0131 kommersiya modell\u0259rind\u0259n f\u0259rqli olaraq, a\u00e7\u0131q elmi modell\u0259r qlobal aliml\u0259r birliyin\u0259 n\u0259tic\u0259l\u0259ri t\u0259krarlamaq, yoxlamaq v\u0259 inki\u015faf etdirm\u0259k imkan\u0131 verir.<\/li>\n<\/ul>\n<h2>N\u0259tic\u0259<\/h2>\n<p>Microsoft v\u0259 dig\u0259r texnoloji liderl\u0259rin formala\u015fd\u0131rd\u0131\u011f\u0131 paylanm\u0131\u015f hesablama \u015f\u0259b\u0259k\u0259l\u0259ri s\u00fcni intellekti sad\u0259c\u0259 alqoritmik s\u0259viyy\u0259d\u0259n qlobal s\u0259naye infrastrukturuna \u00e7evirir. Bu g\u00fccl\u00fc bazan\u0131n \u00fcz\u0259rind\u0259 qurulan ixtisasla\u015fm\u0131\u015f v\u0259 a\u00e7\u0131q elmi modell\u0259r is\u0259 x\u0259r\u00e7\u0259ngl\u0259 m\u00fcbariz\u0259 kimi b\u0259\u015f\u0259riyy\u0259tin \u0259n m\u00fcr\u0259kk\u0259b tibbi probleml\u0259rinin h\u0259llind\u0259 \u0259sas mexanizm\u0259 \u00e7evrilm\u0259kd\u0259dir.<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Hesablama G\u00fcc\u00fcn\u00fcn Yeni M\u0259rh\u0259l\u0259si: &#8220;AI Superfactories&#8221; v\u0259 M\u0259qs\u0259dli Terapevtika S\u00fcni intellekt modell\u0259rinin miqyas\u0131 geni\u015fl\u0259ndikc\u0259, \u0259n\u0259n\u0259vi hesablama infrastrukturlar\u0131 artan t\u0259l\u0259bl\u0259ri \u00f6d\u0259m\u0259kd\u0259 \u00e7\u0259tinlik \u00e7\u0259kir. Microsoft ba\u015fda olmaqla qlobal texnologiya n\u0259h\u0259ngl\u0259ri, h\u0259m geni\u015fmiqyasl\u0131 modell\u0259rin t\u0259limi (training), h\u0259m d\u0259 a\u015fa\u011f\u0131 gecikm\u0259li n\u0259tic\u0259\u00e7\u0131xarma (inference) prosesl\u0259rini t\u0259min etm\u0259k \u00fc\u00e7\u00fcn paylanm\u0131\u015f superhesablama \u015f\u0259b\u0259k\u0259l\u0259rin\u0259 \u2014 mahiyy\u0259tc\u0259 &#8220;AI Superfactories&#8221; (S\u00fcni \u0130ntellekt Superfabrikalar\u0131) adland\u0131r\u0131lan qlobal [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":11355,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"om_disable_all_campaigns":false,"_monsterinsights_skip_tracking":false,"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[1],"tags":[],"class_list":["post-11356","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-uncategorized"],"aioseo_notices":[],"jetpack_featured_media_url":"https:\/\/bitep.net\/blog\/wp-content\/uploads\/2026\/10\/bitep-ai-1791073735262-12.png","jetpack_sharing_enabled":true,"_links":{"self":[{"href":"https:\/\/bitep.net\/blog\/wp-json\/wp\/v2\/posts\/11356","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/bitep.net\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/bitep.net\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/bitep.net\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/bitep.net\/blog\/wp-json\/wp\/v2\/comments?post=11356"}],"version-history":[{"count":0,"href":"https:\/\/bitep.net\/blog\/wp-json\/wp\/v2\/posts\/11356\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/bitep.net\/blog\/wp-json\/wp\/v2\/media\/11355"}],"wp:attachment":[{"href":"https:\/\/bitep.net\/blog\/wp-json\/wp\/v2\/media?parent=11356"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/bitep.net\/blog\/wp-json\/wp\/v2\/categories?post=11356"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/bitep.net\/blog\/wp-json\/wp\/v2\/tags?post=11356"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}