Chii chaitika
NVIDIA yakaendesa data rekuona kwekuita kweVera Rubin NVL72 papuratifomu kune MLPerf v6.1 benchmark suite, ichiratidza kuvandudzwa kwakakosha pamusoro pechizvarwa chakapfuura GB300 NVL72. Mhedzisiro yacho inoratidza inosvika 3.7x yepamusoro-soro pabhenji reQwen3-VL uye 2.5x yepamusoro-soro paDeepSeek-R1, inofambiswa nehardware-software co-design, NVFP4 kunyatsoita, uye akapatsanurwa nzira dzekushumira.
NVIDIA yakaburitsa preview mhinduro yeVera Rubin NVL72 papuratifomu muMLPerf v6.1 suite, yakatarisana neDeepSeek-R1 uye Qwen3-VL mabhenji. Iyo kambani yakataura kuti Vera Rubin NVL72 inoburitsa kusvika 3.7x yakakwirira kupfuura iyo GB300 NVL72 paQwen3-VL pane iyo offline, server, uye inodyidzana mamiriro kana uchishandisa vLLM neNVIDIA Dynamo yakavhurika-sosi inference framework. Kune iyo DeepSeek-R1 bhenji, uchishandisa iyo NVIDIA TensorRT-LLM raibhurari, kubuda kwaive kunosvika 2.5x yakakwirira kupfuura iyo GB300 NVL72.
Kuwana kwekuita kunonzi kuzere-stack co-dhizaini, kusanganisira yakakwidziridzwa Tensor Cores, iyo Transformer Injini, uye NVFP4 chaiyo, iyo inoderedza ndangariro tsoka yemhando uremu, kutarisisa, uye KV cache. Zvinyorwa zvakashandiswa zvakanyanya kupatsanurwa kushumira, kupatsanura prefill uye decode nhanho, pamwe nehukuru-hukuru hwekufananidza kune musanganiswa-wenyanzvi-matanho. Iyo NVL72 scale-up domain, ine simba nechizvarwa chechitanhatu NVLink uye NVLink Switch, yakapa iyo yekubatanidza nheyo inodiwa kune aya matekiniki pa rack scale.
NVIDIA yakaratidzawo kuyera kushanda zvakanaka, ichicherekedza kuti iyo DeepSeek-R1 kuendesa yakayerwa kubva kune imwechete GB300 NVL72 rack kusvika mana racks (288 GPUs) iine 99% kuyera kugona mumamiriro ekunze. Mune agentic workloads, kunyanya iyo SemiAnalysis AgentX bhenji, Vera Rubin NVL72 yakaunza 30x kuita zvirinani pane GB300 NVL72 mukuedzwa kwekutarisa. Partner Nebius akaendesawo Vera Rubin NVL72 preview mhinduro, ichiratidza maitiro akafanana ekuita.
Kuburitswa uku kunosanganisira mibairo kubva kune yakakura NVIDIA ecosystem, ine gumi nevaviri vanoendesa data, kusanganisira ASUS, Azure, Cisco, CoreWeave, uye Oracle Cloud Infrastructure. NVIDIA yakaendesawo Jetson AGX Thor zvawanikwa uchishandisa TensorRT Edge-LLM pane itsva Edge-Agentic benchmark ine Qwen3.6-27B. Post-kutumira mhinduro yeGPT-OSS-120B uye DLRMv3 yakaratidza zvimwe zvakawanikwa asi haisati yasimbiswa neMLCommons.
Kwakabva mashoko: blogs.nvidia.com ↗
Nei zvichikosha
Mhedzisiro iyi inopa humbowo hwechokwadi hwemaitiro ekuita kwechinotevera-chizvarwa cheAI inference, zvichikanganisa zvakananga mutengo-per-token economics kumasangano anoshandisa mhando dzemitauro mikuru. Nekuratidza padhuze-mutsara kuyera kunyatsoshanda pamaraki akawanda uye zvakakura zvakawanikwa muajenti mabasa, iyo data inobatsira mabhizinesi kufanotaura zvivakwa zvezvivakwa uye kusimbisa kugona kwehupfumi hwekuyera kutumirwa kweAI. Izvi zvine basa nekuti mitengo yekufungidzira ndiyo yekutanga mutyairi weAI chigadzirwa purofiti uye kuwanikwa.
Chinonyanya kukosha chezvigumisiro izvi chiri mukuverengwa kweinference economics. Nekuratidza kuti imwe neimwe Vera Rubin NVL72 rack inogona kuburitsa zvakanyanya ma tokens uye kushandira vashandisi vakawanda kupfuura GB300 NVL72 rack, NVIDIA inopa yakajeka metric yekudzikisa mutengo-per-tokeni. Izvi zvakakosha kumasangano uko mitengo yekufungidzira inoumba chikamu chikuru chemari yekushandisa, sezvo ichidudzira zvakananga kune yakakwira mari inokwanisika kana kuderera kwesevhisi mutengo webasa rimwe chete.
Iko kusimbiswa kwekuyera kushanda zvakanaka kunogadzirisa marwadzo akajairika muAI zvivakwa: hukama husiri hwemutsara pakati pekuwedzera kwehardware uye kubudirira kwekuwana. Kuwana 99% kuyera mashandiro mumakumi maviri nemasere eGPUs zvinoratidza kuti dhizaini nekubatanidza zviri kudzikamisa mabhodhoro ekutaurirana, zvichibvumira masangano kufanotaura kubudirira kwekuita zvakanyanya kana vachiwedzera mafekitori avo eAI.
Iko kuiswa kweagentic workload mabenchmarks, akadai seSemiAnalysis AgentX, inoratidza kuchinja kwemaitiro eAI anoyerwa. Sezvo maAI masisitimu anofamba kubva kune imwe-yekutendeuka mhinduro kuenda kune akawanda-nhanho kufunga uye chiito, echinyakare throughput metrics anogona kusanyatsobata utility. Iyo 30x mashandiro ekuvandudza mune ino chaiyo domain inoratidza kuti inotevera-chizvarwa Hardware yakagadziridzwa yakanyatso kurongeka uye compute zvinodiwa zveagent AI, inova chikamu chiri kukura mumabhizinesi maapplication.
Interactive Mechanism: Iyo Inonyatsoshanda
Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.
Which component of an AI application is the machine-learning model itself?
Zvekutarisa zvinotevera
Tarisa kuongororwa kwekupedzisira kweaya mhedzisiro neMLCommons uye kunotevera kuburitswa kweVera Rubin chikuva kumusika. Pamusoro pezvo, tevera kutorwa kweiyo nyowani MPerf Endpoints bhenji yekumisikidza agent, iyo inomisa metrics ekuita kweakawanda-nhanho AI vamiririri, uye tarisa kuti vakwikwidzi vanopindura sei kune aya chaiwo mabudiro uye kuyera mashandiro ebhenji.
Kusimbisa kwekupedzisira kweaya ekutarisa mhinduro neMLCommons ndiyo nhanho inotevera. Kusvika asimbiswa, nhamba idzi ndedzekutanga uye dzinogona kuchinja. Iko kuburitswa kwepamutemo kunopa iyo yechokwadi yekuita baseline yeVera Rubin chikuva.
Kuwanikwa kwemusika uye mitengo yeVera Rubin NVL72 papuratifomu inozoona maitiro ayo anoshanda. Kunyange kuwanikwa kwekuita kwakanyorwa, mutengo wehardware uye nguva yekuchinja kubva kuGB300 ichapesvedzera mitengo yekugamuchirwa. Masangano anozoda kuyera mabhenefiti ekuita maringe nekushandisa mari inodiwa pakusimudzira.
Kuvandudzwa uye kutorwa kweMLPerf Endpoints benchmark yeagent ichave yakakosha. Sezvo iyi bhenji inove yakamira, ichapa maonero akazara eAI system kugona kupfuura yakasvibira token throughput, zvichigona kugadzirisa zvakare kuti vatengesi vanotengesa sei uye kuenzanisa mapuratifomu avo.