PALO ALTO, Calif., July 21, 2026 (GLOBE NEWSWIRE) — DeepI f a, a pu pose-built cloud platfo m fo high-th oughput AI i fe e ce, today a ou ced the esults of a i depe de t be chma k of the NVIDIA Ve a CPU, NVIDIA’s ext ge e atio CPU pu pose-built to powe age tic wo kloads. As o e of a select g oup of collabo ato s g a ted ea ly access to Ve a ha dwa e th ough NVIDIA’s ope AI ecosystem, DeepI f a desig ed a d a its ow p oductio -g ade be chma k to test the i f ast uctu e with a eal-wo ld age t wo kload. The be chma k fou d Ve a led by up to 2.2x agai st competitive CPUs.
DeepI f a p ocesses ea ly five t illio toke s pe week, with close to 30% d ive by age tic systems. At that scale, CPU pe fo ma ce is c itical to delive i g the cost efficie cy, low late cy, a d th oughput that age tic wo kloads dema d.
“We built DeepI f a’s i f ast uctu e f om the g ou d up fo i fe e ce. We have tu ed eve y laye of it fo cost, late cy, a d th oughput at p oductio scale,” said Nikola Bo isov, co-fou de a d CEO, DeepI f a. “NVIDIA is buildi g fo whe e AI wo kloads a e headed, a d we believe Ve a is exactly the ki d of ha dwa e solutio the ext ite atio dema ds.”
DeepI f a tested Ve a usi g its ow p oductio AI age t a d eal captu ed t affic, be chma ki g agai st th ee leadi g CPUs f om AMD, a d I tel u de ide tical co ditio s.
Key esults f om the be chma k i clude:
Fastest i eve y wo kload catego y, doubli g NVIDIA’s claim: Agai st NVIDIA’s published claim of 80% faste age tic CPU pe fo ma ce, Ve a measu ed up to 2.2x faste o chest atio tha the x86 baseli e – a d was the fastest of all fou a chitectu es tested i eve y wo kload catego y.Up to 1.6x mo e co cu e t age ts at the same quality of se vice: O ide tical CPU pa titio s, Ve a sustai ed 256 co cu e t age ts withi st ict espo se-time a d e o ta gets, ve sus 160–192 fo the competi g chips.Spa e capacity se ved a full LLM: With all 256 age ts u i g at full load, Ve a’s leftove co es simulta eously se ved a 20-billio -pa amete ope -sou ce model faste tha a e ti e p evious-ge e atio CPU socket dedicated solely to that task, without slowi g the age ts.
“B i gi g Ve a to ma ket is a i c edible oppo tu ity, but pe fo ma ce is ultimately p ove i p oductio ,” said Ia Fi de , Di ecto , Data Ce te CPU P oducts, NVIDIA. “DeepI f a is pushi g Ve a with dema di g, eal-wo ld age tic wo kloads that eflect the eeds of p oductio e vi o me ts. These esults demo st ate exactly what Ve a was desig ed to delive .”
DeepI f a also published a tech ical a alysis with the full be chma k methodology a d complete esults, he e.
To lea mo e, visit deepi f a.com.
