NVIDIA has expanded its Vera CPU platform with a new liquid-cooled, rack-scale system built for scientific computing and what the company calls “agentic science”. The company announced it on November 16, 2026, at SC26 in Chicago. NVIDIA says the Vera CPU, its first fully custom data center CPU, gives higher performance and up to 2x better energy efficiency than x86 CPUs on scientific workloads. Dell Technologies, HPE, Lenovo and Supermicro plan to ship Vera CPU racks from early 2027.
The news shows a clear shift. CPUs are becoming a first-class part of NVIDIA’s accelerated computing stack, alongside GPUs, not just hosts that feed them.
NVIDIA Vera CPU Platform at a Glance
| Announced | November 16, 2026 (SC26, Chicago) |
|---|---|
| CPU cores | 88 custom NVIDIA Olympus cores, 176 threads per CPU |
| Memory per CPU | SOCAMM2 LPDDR5X, up to 1.2 TB/s bandwidth |
| Rack | 192 Vera CPUs, about 300 TB of memory and 1 PB of storage |
| Rack memory bandwidth | More than 200 TB/s aggregate peak |
| Networking | NVIDIA Quantum-X800 InfiniBand with SHARP In-Network Computing |
| Design | Fully liquid-cooled NVIDIA MGX, eight Vera CPUs per 1U tray |
| Software | NVIDIA HPC SDK 26.11 (available now), new Fortran compiler (public beta) |
| Availability | From Dell, HPE, Lenovo and Supermicro, starting early 2027 |
What’s in the Vera CPU Rack
The new system packs 192 Vera CPUs into one rack, connected by NVIDIA Quantum-X800 InfiniBand. Each 1U tray holds eight Vera CPUs together with NVMe storage and management hardware.
- Olympus cores: every Vera CPU has 88 custom Arm-based NVIDIA Olympus cores and 176 threads.
- Memory bandwidth: each CPU uses compact, serviceable SOCAMM2 LPDDR5X memory with up to 1.2 TB/s of bandwidth. Across the rack, that adds up to more than 200 TB/s of aggregate peak memory bandwidth.
- Capacity: a full rack has roughly 300 TB of memory and 1 PB of storage.
- Shared rack parts: the system reuses NVL72 chassis, power distribution and cooling components, which should make it simpler to deploy and service in data centers that already run NVIDIA GPU racks.
- SHARP In-Network Computing: some collective operations run inside the network, so CPUs spend less time coordinating work with each other.
Memory bandwidth is the key number here. Many scientific codes, such as weather, climate and fluid simulations, are limited by how fast data moves rather than by raw compute. That is where NVIDIA is aiming Vera.
Performance and Energy Efficiency Claims
NVIDIA shared two benchmark results from its own testing:
- OpenFOAM (motorBike Large airflow simulation): Vera delivered 1.9x the performance and 2.4x the performance per watt of x86 systems.
- WRF (Weather Research and Forecasting model): NVIDIA projects 1.6x higher performance and 2x better energy efficiency than x86.
Keep in mind that these are vendor numbers, and the WRF figures are projections. The “2x” headline comes from the WRF projection. NVIDIA says the full test setups are in its Vera CPU HPC performance brief. Independent benchmarks will matter once systems ship in 2027.
First Customer: DKRZ HLRE-5 Climate Supercomputer
The German Climate Computing Center (DKRZ) has picked Vera CPUs for its next supercomputer, HLRE-5. Bull is building it, and it will run alongside NVIDIA GB200 Grace Blackwell systems and Quantum-2 InfiniBand. DKRZ wants to run global climate simulations at kilometer scale while using less energy than its current system. HLRE-5 is expected to be available in early 2028.
Software: HPC SDK 26.11 and a New Fortran Compiler
On the software side, NVIDIA announced two things:
- NVIDIA HPC SDK 26.11 is available now, with compilers and libraries tuned for Vera. You can get it from the NVIDIA HPC SDK page.
- A next-generation Fortran compiler is now in public beta. NVIDIA built it with leading HPC centers and national labs. It aims to make existing Fortran programs run faster on NVIDIA platforms without code rewrites.
The Fortran compiler may matter most in practice. A huge share of climate, weather and physics code is still written in Fortran, and rewriting it is slow and risky.
Why CPUs Matter for AI and HPC
GPUs get most of the attention in AI infrastructure, but much of the work around them still runs on CPUs: data preparation, simulation steps that don’t map well to GPUs, orchestration, and the tool calls that AI agents make. NVIDIA’s “agentic science” pitch is that AI agents will increasingly drive simulations and experiments, and that workflow needs fast CPUs with lots of memory bandwidth as well as GPUs.
With Vera, NVIDIA now offers a full stack: its own CPUs, GPUs, InfiniBand networking and compilers. This puts it in direct competition with Intel Xeon and AMD EPYC in the HPC CPU market, not just alongside them.
For Linux admins, this also means more Arm-based servers in the data center. Operating system vendors are already preparing. SUSE has launched a dedicated SLES edition for NVIDIA’s Arm platforms with day-one support for Vera Rubin.
Frequently Asked Questions
What is the NVIDIA Vera CPU?
Vera is NVIDIA’s first fully custom data center CPU. It is Arm-based, with 88 custom Olympus cores and 176 threads, and it uses LPDDR5X memory with up to 1.2 TB/s of bandwidth per CPU. It can run on its own in CPU racks or be paired with Rubin GPUs in Vera Rubin systems.
When can I buy a Vera CPU rack?
Dell Technologies, HPE, Lenovo and Supermicro plan to ship Vera CPU racks starting in early 2027. NVIDIA has not published pricing.
Is Vera really 2x more efficient than x86?
That’s NVIDIA’s claim, based on its own tests. It measured 2.4x better performance per watt on OpenFOAM and projects 2x on the WRF weather model. Real results will depend on the workload, the x86 system it’s compared with, and how well the code is tuned for Arm.
Conclusion
NVIDIA’s expanded Vera CPU platform brings 192 Arm-based CPUs, over 200 TB/s of memory bandwidth and InfiniBand into one liquid-cooled rack aimed at science and simulation. With OEM systems coming in early 2027, an HPC SDK and a Fortran compiler for existing code, NVIDIA is now selling a complete CPU platform, not just GPUs.
Read the full announcement in NVIDIA’s press release.
How did it make you feel?
