{"id":122456,"date":"2025-03-18T08:22:21","date_gmt":"2025-03-18T06:22:21","guid":{"rendered":"https:\/\/prohoster.info\/blog\/novosti-interneta\/amd-pohvastalas-chto-ryzen-ai-max-395-do-12-raz-bystree-v-rabote-s-ii-chem-pryamoj-konkurent-ot-intel"},"modified":"2025-03-18T08:22:21","modified_gmt":"2025-03-18T06:22:21","slug":"amd-pohvastalas-chto-ryzen-ai-max-395-do-12-raz-bystree-v-rabote-s-ii-chem-pryamoj-konkurent-ot-intel","status":"publish","type":"post","link":"https:\/\/prohoster.info\/cs\/blog\/news\/amd-pohvastalas-chto-ryzen-ai-max-395-do-12-raz-bystree-v-rabote-s-ii-chem-pryamoj-konkurent-ot-intel","title":{"rendered":"AMD se pochlubilo, \u017ee Ryzen AI Max+ 395 je a\u017e 12kr\u00e1t rychlej\u0161\u00ed v pr\u00e1ci s AI ne\u017e p\u0159\u00edm\u00fd konkurent od Intelu.","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"

Nejnov\u011bj\u0161\u00ed vlajkov\u00fd mobiln\u00ed procesor AMD Ryzen AI Max+ 395 rodiny Strix Halo poskytuje a\u017e 12kr\u00e1t vy\u0161\u0161\u00ed v\u00fdkon p\u0159i pr\u00e1ci s r\u016fzn\u00fdmi velk\u00fdmi jazykov\u00fdmi modely AI ne\u017e \u010dipy Intel Lunar Lake. O tom informovala spole\u010dnost AMD ozn\u00e1mila<\/a><\/noindex> na sv\u00e9m ofici\u00e1ln\u00edm blogu, sd\u00edleje odpov\u00eddaj\u00edc\u00ed diagramy.<\/p>\n

<\/center> \t\t\t\t
<\/center> \"AMD <\/p>\n

D\u00edky 16 v\u00fdpo\u010detn\u00edm j\u00e1dr\u016fm Zen 5, 40 grafick\u00fdm blok\u016fm RDNA 3.5 a NPU XDNA 2 s v\u00fdkonem 50 TOPS (bilion operac\u00ed za sekundu) zaji\u0161\u0165uje procesor Ryzen AI Max+ 395 a\u017e 12,2kr\u00e1t vy\u0161\u0161\u00ed v\u00fdkon v n\u011bkter\u00fdch sc\u00e9n\u00e1\u0159\u00edch LLM ne\u017e Intel Core Ultra 258V. Je t\u0159eba p\u0159ipomenout, \u017ee \u010dip Intel Lunar Lake obsahuje pouze \u010dty\u0159i P-j\u00e1dra a \u010dty\u0159i E-j\u00e1dra, co\u017e je celkem o polovinu m\u00e9n\u011b ne\u017e u Ryzen AI Max+ 395. Rozd\u00edl v v\u00fdkonu mezi platformami je v\u0161ak mnohem v\u00fdrazn\u011bj\u0161\u00ed ne\u017e dvakr\u00e1t.<\/p>\n

\"AMD <\/p>\n

V\u00fdhoda \u010dipu Ryzen AI Max+ 395 se st\u00e1v\u00e1 je\u0161t\u011b patrn\u011bj\u0161\u00ed se zvy\u0161uj\u00edc\u00ed se slo\u017eitost\u00ed jazykov\u00fdch model\u016f. Nejv\u011bt\u0161\u00ed rozd\u00edl v v\u00fdkonu mezi platformami je vid\u011bt p\u0159i pr\u00e1ci s LLM s 14 miliardami parametr\u016f, kde je pot\u0159eba v\u00edce opera\u010dn\u00ed pam\u011bti. P\u0159ipome\u0148me, \u017ee Lunar Lake p\u0159edstavuj\u00ed hybridn\u00ed procesory vybaven\u00e9 a\u017e 32 GB vestav\u011bn\u00e9 RAM. <\/p>\n

<\/center> \"AMD <\/p>\n

V testech LM Studio s vyu\u017eit\u00edm za\u0159\u00edzen\u00ed Asus ROG Flow Z13 se 64 GB unifikovan\u00e9 pam\u011bti integrovan\u00e1 grafika Radeon 8060S procesoru Ryzen AI Max+ 395 dos\u00e1hla 2,2kr\u00e1t vy\u0161\u0161\u00ed propustnosti token\u016f ne\u017e grafika Intel Arc 140V v r\u016fzn\u00fdch AI modelech. V testech Time-to-First-Token (metrika v\u00fdkonnosti jazykov\u00fdch model\u016f, kter\u00e1 ukazuje, jak dlouho trv\u00e1 od odesl\u00e1n\u00ed po\u017eadavku do vygenerov\u00e1n\u00ed prvn\u00edho tokenu odpov\u011bdi) \u010dip AMD prok\u00e1zal \u010dty\u0159n\u00e1sobnou v\u00fdhodu nad konkurentem v takov\u00fdch modelech, jako je Llama 3.2 3B Instruct, a zv\u00fd\u0161il odstup a\u017e na 9,1kr\u00e1t v modelech podporuj\u00edc\u00edch 7\u20138 miliard parametr\u016f, nap\u0159\u00edklad DeepSeek R1 Distill.<\/p>\n

\"AMD <\/p>\n

Procesor AMD se zejm\u00e9na da\u0159il v \u00faloh\u00e1ch multimod\u00e1ln\u00edho vid\u011bn\u00ed, kde zpracov\u00e1val slo\u017eit\u00e9 vizu\u00e1ln\u00ed vstupy a\u017e 7kr\u00e1t rychleji v IBM Granite Vision 3.2 3B a 6kr\u00e1t rychleji v Google Gemma 3 12B ve srovn\u00e1n\u00ed s \u010dipem Intel. Podpora technologie Variable Graphics Memory platformou AMD umo\u017e\u0148uje p\u0159id\u011blit a\u017e 96 GB pam\u011bti jako VRAM z syst\u00e9m\u016f s unifikovanou pam\u011bt\u00ed o objemu a\u017e 128 GB, co\u017e zase umo\u017e\u0148uje nasazen\u00ed modern\u00edch jazykov\u00fdch model\u016f, jako je Google Gemma 3 27B Vision.<\/p>\n

V\u00fdhody v\u00fdkonu procesoru AMD nad konkurenc\u00ed jsou viditeln\u00e9 i v praktick\u00fdch AI aplikac\u00edch, jako je anal\u00fdza l\u00e9ka\u0159sk\u00fdch obraz\u016f a pomoc s k\u00f3dov\u00e1n\u00edm pomoc\u00ed vysoce p\u0159esn\u00e9ho 6bitov\u00e9ho kvantov\u00e1n\u00ed v modelu DeepSeek R1 Distill Qwen 32B.<\/p>\n

Zdroje:<\/strong><\/p>\n