{"id":74467,"date":"2024-08-29T18:48:05","date_gmt":"2024-08-29T18:48:05","guid":{"rendered":"https:\/\/ekamu.net\/?p=74467"},"modified":"2024-08-29T18:48:05","modified_gmt":"2024-08-29T18:48:05","slug":"nvidia-blackwell-b200-selefine-4-kat-performans-farki-atti-peki-gercekten-oyle-mi","status":"publish","type":"post","link":"https:\/\/ekamu.net\/index.php\/2024\/08\/29\/nvidia-blackwell-b200-selefine-4-kat-performans-farki-atti-peki-gercekten-oyle-mi\/","title":{"rendered":"Nvidia Blackwell B200, selefine 4 kat performans fark\u0131 att\u0131: Peki ger\u00e7ekten \u00f6yle mi?"},"content":{"rendered":"

<\/span> Nvidia, Blackwell B200<\/strong> i\u015flemcisinin ilk MLPerf 4.1 sonu\u00e7lar\u0131n\u0131 yay\u0131nlad\u0131 ve bu sonu\u00e7lar, yeni Blackwell GPU’nun \u00f6nceki Hopper mimarili H100<\/strong> modeline k\u0131yasla d\u00f6rt kat<\/strong> daha y\u00fcksek performans sundu\u011funu g\u00f6steriyor. Ancak, bu sonu\u00e7lar birka\u00e7 \u00f6nemli noktay\u0131 g\u00f6z \u00f6n\u00fcnde bulundurmay\u0131 gerektiriyor. <\/figure>\n

Nvidia’n\u0131n yay\u0131nlad\u0131\u011f\u0131 verilere g\u00f6re, Blackwell<\/strong> tabanl\u0131 B200<\/strong> GPU<\/strong>, bir sunucu do\u011frulama testinde tek bir GPU ile<\/strong> saniyede 10.755 token<\/strong>, \u00e7evrimd\u0131\u015f\u0131 bir referans testinde ise saniyede 11.264 token<\/strong> i\u015flem kapasitesine sahip. Mevcut MLPerf Llama 2 70B<\/strong> k\u0131yaslama sonu\u00e7lar\u0131na bakt\u0131\u011f\u0131m\u0131zda, d\u00f6rt adet<\/strong> Hopper H100<\/strong> tabanl\u0131 bir makinenin benzer sonu\u00e7lar elde etti\u011fi g\u00f6r\u00fcl\u00fcyor. Bu da Nvidia’n\u0131n, tek bir Blackwell i\u015flemcisinin tek bir Hopper H100 GPU’ya k\u0131yasla yakla\u015f\u0131k 3.7-4 kat daha h\u0131zl\u0131<\/strong> oldu\u011funu ortaya koyuyor.<\/p>\n

4 katl\u0131k h\u0131z fark\u0131 biraz aldat\u0131c\u0131<\/b><\/p>\n

<\/span> Peki bu y\u00fczde 400\u2019e ula\u015fan makine \u00f6\u011frenimi performans fark\u0131 nereden ve nas\u0131l geliyor? \u0130lk olarak, Nvidia’n\u0131n Blackwell i\u015flemcisi, be\u015finci nesil Tensor Core’lar\u0131yla FP4<\/strong> hassasiyetini desteklerken, Hopper tabanl\u0131 H100 yaln\u0131zca FP8<\/strong> format\u0131n\u0131 destekliyor. Haliyle Blackwell’in FP4 performans\u0131, FP8’e k\u0131yasla iki kat daha fazla i\u015fleme kapasitesi sunuyor. Bu, dikkate al\u0131nmas\u0131 gereken ilk \u00f6nemli nokta. <\/figure>\n

\u0130kinci olarak ise Nvidia\u2019n\u0131n asl\u0131nda elma ile armudu k\u0131yaslamas\u0131<\/strong>. Firma, tek bir B200 ile d\u00f6rt adet H100\u2019\u00fc k\u0131yasl\u0131yor, bu biraz yan\u0131lt\u0131c\u0131. Zira genellikle GPU ba\u015f\u0131na performans\u0131 \u00f6l\u00e7mek i\u00e7in GPU\u2019lar tekil olarak k\u0131yaslan\u0131r \u00e7\u00fcnk\u00fc, \u00f6l\u00e7ekleme \u00e7o\u011fu zaman m\u00fckemmel de\u011fildir. Bu, CrossFire veya SLI yap\u0131ld\u0131\u011f\u0131nda iki ekran kart\u0131n\u0131n performans\u0131n\u0131n tek ekran kart\u0131na k\u0131yasla 2X olmamas\u0131 gibi.<\/p>\n

MLPerf 4.1’de tekil halde H100 GPU sonucu bulunmamakta ancak H200 i\u00e7in bulunuyor. Tek bir H200<\/strong>, saniyede 4.488 token<\/strong> i\u015flem kapasitesine ula\u015fabiliyor, bu da B200’\u00fcn 2.5<\/strong> kat daha h\u0131zl\u0131 oldu\u011fu anlam\u0131na geliyor<\/strong>. Bellek kapasitesi ve bant geni\u015fli\u011fi de \u00f6nemli. Test edilen B200 GPU, 180 GB HBM3E belle\u011fi ta\u015f\u0131rken, H100 SXM 80 GB HBM (baz\u0131 konfig\u00fcrasyonlarda 96 GB’a kadar) ve H200 96 GB HBM3 ile 144 GB’a kadar HBM3E sunmakta. Dolay\u0131s\u0131yla “4 kata kadar” ifadesi, say\u0131 format\u0131, GPU say\u0131s\u0131 ve bellek kapasitesi gibi potansiyel farkl\u0131l\u0131klardan kaynaklan\u0131yor.<\/p>\n\n

<\/span><\/p>\n

Kaynak :\u00a0https:\/\/www.donanimhaber.com\/nvidia-blackwell-b200-selefine-4-kat-performans-farki-atti–181297<\/span><\/p>\n","protected":false},"excerpt":{"rendered":"

Nvidia, Blackwell B200 i\u015flemcisinin ilk MLPerf 4.1 sonu\u00e7lar\u0131n\u0131 yay\u0131nlad\u0131 ve bu sonu\u00e7lar, yeni Blackwell GPU’nun \u00f6nceki Hopper mimarili H100 modeline k\u0131yasla d\u00f6rt kat daha y\u00fcksek performans sundu\u011funu g\u00f6steriyor. Ancak, bu sonu\u00e7lar birka\u00e7 \u00f6nemli …<\/p>\n","protected":false},"author":1,"featured_media":74468,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[2563,523,1133,104,2832],"class_list":["post-74467","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-teknoloji","tag-blackwell","tag-gpu","tag-islem","tag-kat","tag-kiyasla"],"_links":{"self":[{"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/posts\/74467","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/comments?post=74467"}],"version-history":[{"count":1,"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/posts\/74467\/revisions"}],"predecessor-version":[{"id":74471,"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/posts\/74467\/revisions\/74471"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/media\/74468"}],"wp:attachment":[{"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/media?parent=74467"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/categories?post=74467"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/tags?post=74467"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}