Nvidia Vera HPC CPU claim faces a newer AMD generation | Magica
Nvidia expands Vera into scientific CPU racks, with 2.4× claim tied to older AMD baseline
Editorial Team
••📖8 min read
Nvidia plans scientific computing racks for early 2027, extending Vera beyond its initial AI deployments. Its 2.4× performance-per-watt claim comes from one airflow simulation against AMD Turin; performance against newer rivals and delivered costs remain unresolved.
Nvidia reports 1.9× performance and 2.4× performance per watt in one OpenFOAM airflow simulation against a two-socket AMD EPYC 9755 system.
The gains vary by application, and the comparison does not cover AMD’s newly available sixth-generation EPYC processors.
Scientific computing racks are expected in early 2027; a German climate supercomputer using Vera is due in early 2028.
Nvidia, the GPU supplier expanding its own general-purpose processors, announced a liquid-cooled rack of 192 Vera CPUs for scientific computing on November 16. Due from server makers in early 2027, it targets simulations that run without GPU acceleration. Nvidia’s headline efficiency result is narrower than that market: one airflow workload tested against an AMD processor from the preceding generation.
A custom successor to Grace, with more of the system attached
A CPU runs operating systems and application code; GPUs handle the parallel calculations behind AI and many scientific applications. Nvidia’s new platform announcement addresses the scientific work that remains on CPUs, including weather modeling and fluid simulation. Its performance brief treats CPUs and GPUs as complementary, with GPUs serving highly parallel, vector-intensive work.
Nvidia calls Vera its first fully custom data center CPU. Its earlier Grace processor used licensed Neoverse V2 cores from chip-architecture supplier Arm; Vera uses Nvidia-designed Olympus cores. Each processor has 88 cores and supports 176 threads. Nvidia says it retains compatibility with software written for Arm’s architecture, as described in a report on the launch.
Vera also predates this scientific rack. That report, citing an Nvidia delivery update from August, says initial systems went to Oracle Cloud Infrastructure and AI developers Anthropic, OpenAI and SpaceXAI. It also describes OpenAI running agent sandbox workloads on Vera in Oracle’s cloud, where the processor became generally available in October. The new announcement extends an existing processor and its AI-oriented deployment effort into a packaged scientific computing system.
The announced rack combines Vera with Nvidia’s Quantum-X800 InfiniBand interconnect, about 300TB of memory and a petabyte of storage. It reuses chassis, power distribution and cooling components from Nvidia’s NVL72 rack infrastructure. Nvidia is offering a package of processors, memory, networking and software, giving it a role in more of the scientific computing system.
Nvidia’s explanation for the efficiency gains centers on moving data efficiently. Each CPU supports up to 1.2TB/s of memory bandwidth using LPDDR5X memory. Its puts fully populated memory-subsystem power at approximately 30–40 watts. That measures a component, rather than a complete server or rack. On STREAM, a sustained-memory-bandwidth benchmark, Nvidia claims 3.4× the bandwidth per watt of its two-socket AMD Turin comparison system.
South Korea is strengthening device checks and considering public Wi-Fi shutdowns before its November 19 college entrance exam, after a tax-accountant exam top scorer admitted using AI to cheat. The response still depends heavily on human screening, with no electronic device detectors planned for this year's test.
Company-reported NVIDIA results compare two-socket Vera and AMD EPYC 9755 systems, with AMD set to 1. Gains vary by workload; BWA performance falls below the AMD baseline. Source: NVIDIA performance brief.
What the 2.4× figure measures
The headline result comes from OpenFOAM’s motorBike Large workload, which simulates airflow around a motorcycle. Nvidia compares two-socket Vera and AMD EPYC 9755 systems, with the AMD result normalized to one. The same brief shows substantially different outcomes across applications:
Workload
Vera performance versus AMD
Vera performance per watt versus AMD
OpenFOAM airflow simulation
1.9×
2.4×
OpenCourant structural dynamics
1.1×
1.4×
HPCG sparse linear algebra
2.1×
2.6×
BWA genome alignment
0.9×
1.2×
These are Nvidia-published results, not independent validation. BWA is a useful counterexample: the AMD system finishes sooner, while Vera retains a claimed efficiency advantage. The brief says the dual-socket AMD configuration has 256 cores against Vera’s 176, and that BWA’s independent alignment tasks benefit from the additional cores.
The listed platforms also differ in memory and operating-system configurations. These results compare the tested systems; they do not isolate the effect of Nvidia’s core design.
The 2.4× performance-per-watt figure also should not be read as a 2.4-fold reduction in instantaneous power. Dividing the reported 1.9× performance by 2.4× efficiency implies Vera used roughly 79% of the comparison system’s power, or about 21% less, within that benchmark’s measurement boundary. The published brief lists configurations but does not explain the power-measurement procedure or provide absolute power readings for that test. It therefore cannot establish a facility electricity saving or a purchase-cost payback.
For the Weather Research and Forecasting model, Nvidia’s announcement explicitly calls the 1.6× performance and 2× energy-efficiency gains projections. Those figures should remain distinct from the OpenFOAM test claim.
There is a further timing limitation. AMD supplies rival server CPUs based on the x86 architecture, also used by Intel. Although Nvidia’s brief describes its comparisons as against the latest competition, the named AMD processor belongs to the Turin generation. AMD announced availability of sixth-generation EPYC SP7 processors on November 10, saying systems were shipping from leading equipment manufacturers. Those processors use its newer Zen 6 architecture. Nvidia’s comparison establishes no result against that generation.
Packaging CPUs at rack scale is also a competitive arena. AMD’s same announcement describes Helion, its open blueprint for agentic AI infrastructure, which lets customers choose CPU and GPU capacity, cooling and connectivity. That is an alternative deployment approach, rather than a like-for-like scientific benchmark against Vera.
The brief also describes larger gains in selected customer case studies: 2.2× to 7.5× on three scientific applications against a separately described x86 baseline, and 4.4× on a scientific-agent task involving compilation and hydrodynamics simulations. For the agent task, the language model and inference infrastructure were held constant; the measured work was compiling and executing simulation code. Those results do not measure faster model inference. The brief’s multi-node evaluation, extending to 12 Vera nodes, compares against Nvidia’s prior Grace platform; competitive clusters were unavailable.
DKRZ’s €40 million investment includes the wider supercomputer and five-year support. Its target of around 75% less energy per simulated year versus Levante applies to the mixed CPU and GPU system. Source: German Climate Computing Center.
A climate customer, but a different procurement
DKRZ, the German Climate Computing Center, is a nonprofit provider of computing, storage and model-optimization support to German climate researchers. It has selected Vera for its next climate supercomputer, HLRE-5. The center’s procurement release dated October 21 disclosed a €40 million contract with its longstanding system supplier Bull, including five years of support. The selection predates the new rack announcement.
The center brings experience running energy-efficient climate infrastructure: it says the GPU extension of its current Levante system ranked third on the November 2025 and June 2026 Green500 efficiency lists. It also credits earlier work with Nvidia on optimizing the ICON climate model for GPUs. Its choice builds on existing relationships and operating experience.
Expected to enter service in early 2028, HLRE-5 combines a Vera CPU partition with a GB200 Grace Blackwell GPU partition that supplies most of its simulation and AI capability. DKRZ says this will bring Arm to its core system for the first time, serving workloads that do not run on GPUs.
The center targets around 75% less energy per simulated year than its current Levante system, alongside the ability to run two global climate simulations simultaneously at five-kilometer resolution. That is a target for the mixed supercomputer, not a measured Vera-only gain. The contract value likewise covers the wider system and support, including storage and liquid cooling.
Funding comes mainly from the Helmholtz Association, Max Planck Society and City of Hamburg. The release says additional funding from the Max Planck Society and Germany’s Federal Ministry of Research, Technology and Space enabled increased system capacity. This is a publicly backed research procurement, rather than a disclosed price for Nvidia’s new rack.
The infrastructure also extends beyond chips. DKRZ plans to continue Bull’s warm-water liquid cooling and upgrade its cooling infrastructure to make more waste heat available for reuse by the neighboring University of Hamburg. These site-level changes are separate from Nvidia’s processor benchmark.
The next test is deployment economics
Nvidia expects Vera CPU racks from Dell Technologies, HPE, Lenovo and Supermicro beginning in early 2027. Its HPC SDK 26.11 is available now, and a new Fortran compiler is entering public beta. The compiler is intended to help existing scientific programs run on Nvidia platforms without rewriting their code—an important part of making the hardware usable for established research workloads.
Buyers still need configuration-specific answers. The announcement describes 192 CPUs per rack, while the performance brief describes a 32-tray reference configuration with 256. The documents give different rack specifications without reconciling them. Nor does DKRZ’s selection establish that it will buy the newly announced 192-CPU rack.
Nvidia’s launch announcement discloses no rack prices, unit volumes or rack order values, a limitation also identified in an analysis of its announcements. DKRZ’s broader contract does not fill that gap. Buyers next need application tests against current alternatives, documented power measurements and delivered system costs, including the cooling and software work their sites require. Those results will determine whether the claimed efficiency translates into more useful science within a fixed budget.
Russia’s objections to observer participation stalled the opening of a Geneva weapons conference, where governments must decide whether to authorize negotiations on autonomous weapons and whether any resulting rules should be binding.
Apple’s iOS 27.3 beta ties higher allowances for server-based AI to existing iCloud+ plans. Basic Siri commands remain free, but undisclosed daily limits leave the value of paid access uncertain.
The partners plan to combine IBM quantum processors, NVIDIA computing and Fermilab physics expertise to study modular quantum systems. A separate decoder release promises faster error-correction research, with important limits on its performance comparisons.
Federal Reserve Chairman Kevin Warsh outlined AI scenarios spanning shared prosperity, job displacement and an investment downturn. The exercise offers no forecast or rate guidance as colleagues worry that the infrastructure buildout is adding to inflation.
OpenAI is offering approved organizations a free classifier for suspected child sexual abuse images, including AI-generated material. Early results suggest less review work, but existing detection alternatives and the need for human review qualify the launch.
OpenAI’s CEO expressed empathy for Nvidia founder Jensen Huang but withheld judgment on reinstatement. His confidence in continued AI financing comes with a direct commercial stake and unresolved deployment risks.
Google’s 8,192-chip Ironwood submission completed the MLPerf DeepSeek-V3 benchmark in 1.70 minutes, close to NVIDIA’s 1.53 minutes at the same accelerator count. Google’s scaling data also shows the extra chip-time required for its fastest result, leaving cost and energy comparisons unresolved.
TensorWave will supply dedicated AMD infrastructure under a six-year OpenAI agreement, with delivery expected from 2027. Its operating record and financial backing sit alongside an undisclosed contract price and a recently terminated site lease.
California alleges OpenAI endangered the public and misled consumers by testing powerful AI models without adequate controls. The lawsuit follows a Hugging Face intrusion that OpenAI’s own account traced to containment failures, missing monitors and mishandled warning signs.
Perplexity has launched Frontier Research to develop AI that can formulate questions, test hypotheses and produce interpretable discoveries. The company promises more than 10,000 GPUs, but existing automated-research systems show why computing resources and generated papers alone cannot establish scientific value.
The new lab led by Stanford professors Yejin Choi and Carlos Guestrin has backing from Nvidia, Cisco and Snowflake. It promises systems that improve on company work, but research benchmarks leave production costs and customer control unproven.
NVIDIA announced a Rubin HPC GPU with more than 50,000 CUDA cores, sparse data acceleration and seventh-generation NVLink. The design targets simulations used by scientists and AI agents, but missing performance, power and pricing details leave its practical advantage unproven.
OpenAI will open 50,000 new ChatGPT access grants on November 23, covering up to 150,000 researchers and collaborators. New recipients get Plus instead of Pro, with first-come selection replacing a lottery. Its pilot study reports perceived time savings but does not establish faster discovery.
OpenAI reports production healthcare workflows at CVS and Cigna and stronger computer-use evaluations. The results show progress across disconnected software, but leave complete-task reliability and total deployment costs unresolved.
President Donald Trump supported a return for Nvidia co-founder Jensen Huang, but the company has installed interim leadership and begun a successor search amid scrutiny of his conduct and alleged Chinese intelligence ties.
Intel disclosed the bandwidth target for its next Xeon server processors at SC26, alongside previously announced core and cache limits. Application performance, power requirements and launch timing remain unresolved.
Mercedes-Benz plans California early access to city-driving and hands-free highway assistance in December, ahead of a 2027 U.S. rollout. Mixed road tests and unannounced subscription prices leave buyers waiting for evidence of everyday value.
Researchers estimate that 403 large U.S. data centers consumed 81.8 TWh in a year. Their regional-grid analysis reveals substantial fossil-fuel dependence, but cannot isolate AI demand or fully expose the underlying facility data.
HSBC lowered its Zhipu share-price target to HK$500 and delayed forecast adjusted breakeven by two years. Improving API margins and substantial fundraising cushion a business whose gross profit still covers little of its research bill.