NVIDIA plans to offer its 88-core server CPU, Vera, as a single SKU (product model) and use it for everything from hosting AI GPUs to CPU-centric scientific computing. In an interview published by The Register on November 20, Ian Finder, who leads NVIDIA's data center CPU development, explained the thinking behind the design. The single-SKU approach was already outlined at GTC 2026, but the HPC rack and performance evaluation materials released in November have made clearer which workloads Vera targets and where it is strong or weak against competing products.
The notable point is that the same CPU shows very different gaps against competitors depending on the workload. In results published by NVIDIA, Vera substantially outperforms AMD EPYC in fluid dynamics, while EPYC finishes DNA sequence alignment in less time. Vera's design focuses not just on core counts and clock speeds, but on memory access and power efficiency when many cores run at the same time.
One SKU, aimed at performance with all cores running
Vera has 88 custom CPU cores and runs at up to 3.3 GHz. According to The Register, NVIDIA plans to use this specification as a common SKU for Vera Rubin GPU hosts, CPU-only racks, storage systems and more. Rather than preparing CPUs with different core counts and frequencies for each use, the company will use the same CPU and change the system configuration.
That said, NVIDIA is not trying to cover every use in the CPU market with one product. In a Q&A session at GTC 2026 published by Tom's Hardware, Ian Buck, NVIDIA's vice president of hyperscale and HPC, stated clearly that Vera will ship as a single SKU, but explained that the company does not intend to serve the entire CPU market with one product. The emphasis is on speeding up the CPU-side work that makes GPUs wait during AI training and inference. Partners will be offered a two-socket reference module as well as the standalone CPU.
What Finder stressed in this interview was performance with all 88 cores active at once. Many conventional server CPUs are designed to prioritize response time under light load, and as more cores become active, the internal interconnect can become congested and memory access latency can grow. Vera is designed to keep per-core performance and performance per watt while holding down memory latency even when every core is under load.
He said the goal is to combine per-core performance close to that of frequency-focused x86 CPUs with power efficiency close to that of density-focused CPUs. The main targets are workloads such as AI agents, databases and analytics, which need both high single-thread performance and the ability to run many tasks in parallel.
This design also bears on the efficiency of AI agents. According to a technical article NVIDIA published on July 7, when an AI model calls a tool or runs code partway through inference, the CPU mainly handles that work. If the CPU-side processing drags on, then in a busy environment the KV cache used for the original request is more likely to be evicted while the GPU handles other requests, forcing results to be regenerated. If the CPU returns results quickly, that waiting time and recomputation burden can be reduced.
Shared cache and wide memory bandwidth behind the 88 cores
Vera consolidates its 88 CPU cores on a single compute die, with memory and I/O functions split off onto separate smaller dies. According to NVIDIA's design explainer and its HPC performance document, the 88 cores are linked with 164 MB of shared L3 cache through an internal interconnect called the Scalable Coherency Fabric (SCF). The SCF keeps data coherent while moving it at high speed among the cores, shared cache, memory controllers and I/O.
The OS sees each CPU socket as a single NUMA domain. NUMA (Non-Uniform Memory Access) refers to a configuration in which access times and other characteristics vary depending on where memory is placed. By treating a socket as one domain, Vera makes it easier to tune the placement of program threads, parallel-computing processes and data. In a two-socket configuration, each socket becomes its own NUMA domain.
Main memory uses SOCAMM2-standard LPDDR5X, supporting up to 1.5 TB of capacity and up to 1.2 TB/s of memory bandwidth per CPU. Capacity determines how much data can be held, while bandwidth determines how much data can move between CPU and memory per unit of time. If many cores compute in parallel but the necessary data cannot be supplied fast enough, the computation ends up waiting on memory reads. Vera emphasizes not only the cores' computational performance but also its ability to feed data from memory.
Another feature is a proprietary multithreading technology called Spatial Multithreading. The custom Armv9.2-compatible CPU core, Olympus, can split its execution resources into two hardware threads as needed. It lets a core choose between using broad execution resources for one thread or advancing different tasks simultaneously with two threads, and the 88 physical cores can handle up to 176 hardware threads. However, doubling the thread count does not double performance; the effect varies with the characteristics of the program being run.
Vera leads in fluid dynamics, while EPYC comes out ahead in DNA alignment
The HPC performance document NVIDIA released at SC26 (Figures 4 and 7 and the system configuration table in the appendix) shows that the performance gap between Vera and competing CPUs changes with the workload. The comparison target is a system with two of AMD's 128-core EPYC 9755 CPUs (code-named Turin).
| Workload and test conditions | Performance | Performance per watt |
|---|---|---|
| Fluid dynamics: OpenFOAM v2606, motorBike Large | Vera 1.9x EPYC 9755 | Vera 2.4x |
| DNA sequence alignment: BWA, all datasets | EPYC 9755 slightly faster | Vera 1.2x |
Note: Measurement results published by NVIDIA in November 2026. Both systems are two-socket configurations: Vera has 2 × 88 cores, or 176 physical cores, and EPYC 9755 has 2 × 128 cores, or 256 physical cores. Memory configurations and operating systems also differ, so this is a comparison between systems rather than between individual CPUs.
Vera came out far ahead in fluid dynamics, while the EPYC 9755 was slightly faster at DNA sequence alignment. In performance per watt, Vera leads in both tests. How a CPU is rated changes depending on whether the priority is completion time or work done per unit of power consumed.
OpenFOAM divides the computational domain into a fine grid and solves for fluid behavior in each region. The calculation requires iterative solution of simultaneous equations using sparse matrices, so memory access, cache efficiency and synchronization between processes strongly affect performance. Vera's wide memory bandwidth, shared cache and single-domain-per-socket NUMA layout suit this kind of workload.
BWA's DNA sequence alignment, by contrast, matches large numbers of short DNA sequences against a reference sequence, each independently. Because each thread can proceed without waiting on others, the number of tasks that can run in parallel matters more than memory access latency or communication between cores. NVIDIA's document says Vera's per-core performance was 1.3 times that of the EPYC 9755, but the EPYC, with more total cores, finished slightly sooner.
Independent research also confirms Vera's strengths and weaknesses. On November 11, Joseph Moore, Thomas Green and Simon McIntosh-Smith of the University of Bristol posted a paper to arXiv titled "NVIDIA Vera: A Microarchitecture-Level Comparison with Grace, Turin, and Granite Rapids". The study, accepted to PMBS26, a workshop held alongside SC26, examines Vera's microarchitecture and scientific computing performance using an early evaluation system provided by NVIDIA.
The team set a power cap equivalent to a 450 W TDP on the air-cooled Vera evaluation system, and used a 96-core EPYC 9R45 and a 96-core Xeon 6975P-C on AWS as the AMD and Intel comparison points. All measurements were taken on a single socket, and the compiler was standardized on GCC 15.3.0.
In STREAM Triad, which measures memory bandwidth, Vera recorded about 1.09 TB/s with all 88 cores running one thread each. That is close to the 1.2 TB/s maximum bandwidth in the specifications and was the highest among the CPUs the team compared. On the other hand, competitors came out ahead on workloads that depend heavily on floating-point computation. In the GROMACS molecular dynamics simulation, the Xeon reached 1.65 times Vera's performance, and in the miniBUDE molecular docking workload, the EPYC reached 2.38 times.
The researchers cite the CPUs' vector computing performance and core counts as factors in the difference, along with the fact that software optimization for Arm is not yet as mature as for x86. In workloads like GROMACS and miniBUDE, where floating-point computation determines performance, Vera cannot fully exploit its memory bandwidth advantage. The study used an early evaluation system and a specific cloud environment in single-socket tests, so its conditions differ from the two-socket system comparison NVIDIA published.
The effect of Spatial Multithreading also varied by workload. In the scientific applications the team evaluated, the geometric mean performance gain with two threads was 1.11 times. However, in the FT benchmark (class D) of the NAS Parallel Benchmarks, performance with two threads under default settings fell to 0.61 times that with one thread. This is thought to be caused by factors such as multiple threads competing for cache. When the work-partitioning settings were changed, it improved to 1.13 times.
The result also shows that enabling all of Vera's 176 threads does not always improve performance, and that tuning to match the program's parallelization and data placement is necessary. The study also did not evaluate communication between two sockets, workloads combining CPUs and GPUs, or whole-cluster performance.
A 2027 HPC rack, and how DKRZ divides work between CPUs and GPUs
In an announcement on November 16, NVIDIA unveiled a liquid-cooled HPC rack system holding 192 Vera CPUs. It uses NVIDIA Quantum-X800 InfiniBand to connect the CPUs and carries roughly 300 TB of memory across the rack. A 1U tray holds eight Vera CPUs and also integrates storage and management functions.
The system is scheduled to be offered by Dell Technologies, HPE, Lenovo and Supermicro in early 2027. This differs in product scope and configuration from the general-purpose Vera systems NVIDIA indicated in March 2026 for the second half of 2026.
Plans for specific supercomputers using Vera are also advancing. On October 26, the German Climate Computing Center (DKRZ) announced the procurement contract for its next supercomputer, HLRE-5. The system, based on the BullSequana XH3500, will have a CPU partition with NVIDIA Vera and a GPU partition with GB200 Grace Blackwell.
DKRZ plans to assign most of the scientific computing and the AI workloads to the GPU partition, and to have Vera CPUs handle calculations that GPUs cannot process efficiently. Nodes will be connected with 400 Gbit/s NVIDIA Quantum-2 InfiniBand, so the network configuration also differs from the HPC Vera rack NVIDIA has just announced.
HLRE-5 is scheduled to begin operation in early 2028. DKRZ aims for a system that cuts the energy needed to compute one simulated year by about 75% compared with its current supercomputer, Levante, and that can run two global climate simulations at 5 km resolution at the same time. These are design goals for the system as a whole, including the GPU partition, not performance targets for the Vera CPU alone.
Vera's single-SKU strategy is an attempt to put a CPU developed as a GPU host to work in scientific computing and other uses as well. The actual performance comparisons, however, show that it excels where memory bandwidth and access latency matter, while competitors can come out ahead where raw compute performance or the number of cores that can run in parallel dominates. The 2027 HPC rack shipments and the 2028 start of HLRE-5 operation will show how these characteristics play out in large-scale systems.
