Subira ku makuru
Guhanga udushyaAI Understanding ibisobanuro

NVIDIA Vera Rubin NVL72 imyanya iyoboye MLPerf Yerekana v6.1 ibisubizo

NVIDIA yasohoye ibisubizo byerekana mbere yerekana Vera Rubin NVL72 igera kuri 3.7x yinjira cyane kuruta GB300 NVL72 kuri Qwen3-VL na 2.5x kuri DeepSeek-R1 muri suite ya MLPerf.

4 min readRead the primary source
Source-provided image accompanying NVIDIA Vera Rubin NVL72 posts leading MLPerf Inference v6.1 results
Inyandiko y'ibanzeInkomoko yanditse
Umwanditsi
blogs.nvidia.com
Ihuza ry'inkomoko
blogs.nvidia.comhttps://blogs.nvidia.com/blog/vera-rubin-nvl72-mlperf-inference/
Ubwoko bw'inkomoko
Inyandiko y'ibanze - itangazo ryemewe, impapuro, dosiye, cyangwa urupapuro rwambere-dusoma mu buryo butaziguye.
ImirongoSobanukirwa ibi mumasegonda 60

Tangira hano

Amagambo y'ingenzi

Umwanzuro
Icyiciro cyogukora aho icyitegererezo cyahuguwe gitanga ibyahanuwe cyangwa ibisubizo.
Ururimi runini (LLM)
Ururimi rwicyitegererezo rwahuguwe kumyandiko minini corpora kubyara no gusesengura inyandiko.
Kwibuka (Memory Memory)
Imiterere yabitswe umukozi wa AI akoresha intambwe cyangwa amasomo kugirango atezimbere.
IsuzumeModeri ya AI Yasobanuwe Ikibazo

Byagenze bite

NVIDIA yatanze amakuru yerekana imikorere yimikorere ya platform ya Vera Rubin NVL72 kuri site ya MLPerf v6.1 yerekana ibipimo ngenderwaho, byerekana iterambere ryinshi ryinjira mubisekuruza byabanjirije GB300 NVL72. Ibisubizo byerekana kugeza kuri 3,7x byinjira cyane mubipimo bya Qwen3-VL hamwe na 2,5x byinjira cyane kuri DeepSeek-R1, bigendanwa na software ikorana na software, NVFP4, hamwe nubuhanga bwo gutanga serivisi butandukanye.

NVIDIA yasohoye ibisubizo byibanze kuri platform ya Vera Rubin NVL72 muri suite ya MLPerf v6.1, yibanda ku bipimo bya DeepSeek-R1 na Qwen3-VL. Isosiyete yatangaje ko Vera Rubin NVL72 itanga ibicuruzwa bigera kuri 3.7x hejuru ya GB300 NVL72 kuri Qwen3-VL hakurya ya interineti, seriveri, hamwe n’imikoranire iyo ukoresheje vLLM hamwe na NVIDIA Dynamo. Kubipimo bya DeepSeek-R1, ukoresheje isomero rya NVIDIA TensorRT-LLM, ibyinjira byageze kuri 2.5x hejuru ya GB300 NVL72.

Imikorere yunguka yitirirwa hamwe-igishushanyo mbonera, harimo na Tensor Cores yongerewe imbaraga, Transformer Moteri, na NVFP4 neza, bigabanya ikirenge cyo kwibuka kuburemere bw'icyitegererezo, kwitondera, na KV cache. Ibyatanzwe byakoresheje cyane serivisi zidatandukanijwe, gutandukanya ibyiciro byambere na decode, hamwe ninzobere nini nini zinzobere zibangikanye nuruvange-rwinzobere. Indangantego ya NVL72, ikoreshwa nigisekuru cya gatandatu NVLink na NVLink Switch, yatanze umusingi uhuza nkenerwa kuri ubwo buhanga ku gipimo cya rack.

NVIDIA yanagaragaje imikorere yo gupima, avuga ko itangwa rya DeepSeek-R1 ryagabanijwe kuva kuri GB300 NVL72 imwe kugeza kuri bine (288 GPU) hamwe na 99% byo gupima muburyo bwa interineti. Mubikorwa byakazi, cyane cyane igipimo cya SemiAnalysis AgentX, Vera Rubin NVL72 yatanze 30x imikorere myiza kuruta GB300 NVL72 mugupima ibizamini. Umufatanyabikorwa Nebius yatanze kandi ibisubizo bya Vera Rubin NVL72, byerekana imikorere isa.

Irekurwa ririmo ibisubizo bivuye muri ecosystem yagutse ya NVIDIA, hamwe nabafatanyabikorwa 19 batanga amakuru, nka ASUS, Azure, Cisco, CoreWeave, nibikorwa remezo bya Oracle Cloud. NVIDIA yatanze kandi ibisubizo bya Jetson AGX Thor ukoresheje TensorRT Edge-LLM ku gipimo gishya cya Edge-Agentic hamwe na Qwen3.6-27B. Ibisubizo nyuma yo gutanga ibisubizo kuri GPT-OSS-120B na DLRMv3 byagaragaje izindi nyungu ariko ntibiremezwa na MLCommons.

Ibisobanuro birambuye: blogs.nvidia.com ↗

Impamvu ari ngombwa

Ibisubizo biratanga ibimenyetso bifatika byerekana imikorere yimikorere yibisekuruza bizakurikiraho ibikorwa remezo bya AI, bigira ingaruka zitaziguye kubiciro byubukungu bwamashyirahamwe akoresha imvugo nini yindimi. Mugaragaza uburyo bwo gupima hafi yumurongo munini hamwe ninyungu nyinshi mubikorwa byakazi, amakuru afasha ibigo guteganya ibikenerwa mubikorwa remezo no kwemeza imbaraga zubukungu zo gupima ibikorwa bya AI. Ibi bifite akamaro kuko ibiciro byumwanzuro nigikoresho cyambere cyibicuruzwa bya AI byunguka kandi bigerwaho.

Ubusobanuro bwibanze bwibisubizo biri mubipimo byubukungu bwimbitse. Mugaragaza ko buri rack ya Vera Rubin NVL72 ishobora kubyara ibimenyetso byinshi kandi igakoresha abakoresha benshi kuruta GB300 NVL72, NVIDIA itanga igipimo gisobanutse cyo kugabanya ikiguzi kuri buri kimenyetso. Ibi nibyingenzi mumashyirahamwe aho ibiciro byumwanzuro bigize igice kinini cyibikorwa byakazi, kuko bihindura muburyo butaziguye amafaranga yinjira cyangwa amafaranga make ya serivisi kumurimo umwe.

Kwibanda ku gupima imikorere ikemura ikibazo rusange kibabaza mubikorwa remezo bya AI: umubano utari umurongo hagati yo kongeramo ibyuma ninyungu zinjira. Kugera ku gipimo cya 99% muri 288 GPUs byerekana ko imyubakire n’imikoranire bigabanya neza inzitizi z’itumanaho, bigatuma amashyirahamwe ashobora guhanura imikorere neza mu gihe yagura inganda za AI.

Kwinjizamo ibipimo ngenderwaho byakazi, nka SemiAnalysis AgentX, byerekana impinduka muburyo imikorere ya AI ipimwa. Nka sisitemu ya AI yimuka kuva kumurongo umwe usubiza kubitekerezo byinshi hamwe nibikorwa, ibipimo byinjira mubisanzwe ntibishobora gufata neza akamaro. Iterambere rya 30x muri iyi domeni yihariye ryerekana ko ibyuma bizakurikiraho byateganijwe neza cyane kubukererwe no kubara ibyifuzo bya agent agent AI, urwego rukura mubikorwa bya entreprise.

Interactive Mechanism

Uburyo bukoreshwa: Uburyo bukora

Shakisha ikoranabuhanga ryihishe inyuma yiri terambere.

Thinking Budget (Test-Time Tokens):1,024 tokens
Complex Accuracy79%Math & Code Logic
Latency3.2sTime to first full output
Inference Cost$0.0092Per query estimated
Reasoning StyleStep VerificationInternal chain depth
Active Thinking Trace:
1Deconstruct user problem into formal constraints
2Propose candidate hypotheses & step-by-step calculation
3Self-correction: Backtrack and refute subtle edge cases
4Exhaustive consistency check & final output synthesis
Core takeaway: Test-time compute fundamentally changes AI economics. Instead of only scaling during pre-training, giving reasoning models more tokens at inference time allows them to systematically solve PhD-level STEM problems.
Kugenzura Ibitekerezo Byagenzuwe+10 Points
AI Models Explained Quiz

Which component of an AI application is the machine-learning model itself?

Ibyo kureba

Kurikirana igenzura rya nyuma ryibisubizo byakozwe na MLCommons hamwe nibisohoka nyuma ya platform ya Vera Rubin kumasoko. Byongeye kandi, kurikirana iyemezwa rya MLPerf Endpoints igipimo ngenderwaho kugirango harebwe abakozi, bizagaragaza ibipimo ngenderwaho kubikorwa byintambwe nyinshi za AI, kandi urebe uburyo abanywanyi bitabira ibyo bipimo byihariye kandi byerekana ibipimo ngenderwaho.

Igenzura ryanyuma ryibisubizo byerekanwe na MLCommons nintambwe ikurikira. Kugeza igihe bigenzuwe, iyi mibare irabanza kandi irashobora guhinduka. Isohora kumugaragaro rizatanga imikorere yibanze kumurongo wa Vera Rubin.

Kuboneka kw'isoko n'ibiciro bya platform ya Vera Rubin NVL72 bizagaragaza ingaruka zifatika. Mugihe ibyagezweho mubikorwa byanditse, ikiguzi cyibikoresho nigihe cyinzibacyuho kuva GB300 bizagira ingaruka kubiciro byo kurera. Amashyirahamwe azakenera gupima inyungu zikorwa n’amafaranga yakoreshejwe asabwa mu kuzamura.

Gutezimbere no kwemeza ibipimo bya MLPerf Endpoints kubisobanuro byabakozi bizaba ingenzi. Mugihe iki gipimo kimaze kuba gisanzwe, bizatanga ibisobanuro birambuye byubushobozi bwa sisitemu ya AI irenze ibicuruzwa biva mu mahanga, birashobora guhindura uburyo abacuruzi binjiza isoko kandi bakagereranya urubuga rwabo.

Ibijyanye nuyobora & ibibazo

Moderi ya AI YasobanuweAmahugurwa ya AIEjo hazaza ha AIGerageza ibyo uzi - gerageza ikibazo cya AI kubuntuReba ijambo AI mumagambo yacuKurikiza icyerekezo cya AI cyo kurekura
Basanze ari ingirakamaro?