NVIDIA’s Vera processor was introduced as a data-center CPU for agentic AI. This week, the company is making a second case for it: scientific computing, where simulations can be limited by moving large amounts of data as much as by raw arithmetic.
At Supercomputing 2026 on November 16, NVIDIA announced a rack-scale Vera CPU system for high-performance computing. The clearest real-world example so far is Germany’s next climate supercomputer, HLRE-5, where the German Climate Computing Center (DKRZ) plans to include a Vera-based CPU partition alongside GPU systems.
What NVIDIA announced
The proposed rack contains 192 Vera CPUs, connected with NVIDIA Quantum-X800 InfiniBand networking. NVIDIA says each processor has 88 custom Olympus cores and 1.2 terabytes per second of memory bandwidth; the full system combines about 300 terabytes of memory and is liquid-cooled. The company says it is aimed at workloads such as climate modeling and molecular dynamics that do not map neatly onto GPU acceleration alone.
That does not mean CPUs are replacing GPUs. DKRZ’s planned HLRE-5 is a mixed system: its GPU partition, built around NVIDIA GB200 Grace Blackwell systems, is intended to carry much of the simulation and AI work. Vera CPUs form a separate CPU partition for workloads that do not run on GPUs. The practical question is how well the two kinds of hardware complement each other across researchers’ actual software.
What the climate-center project tells us
DKRZ and system builder Bull announced the HLRE-5 contract on October 21, before NVIDIA’s November 16 Vera rack announcement. DKRZ says the €40 million project is expected to enter service in early 2028 and aims to run two global climate simulations at five-kilometer resolution while reducing energy use per simulated year by around 75% compared with its current Levante system.
Those are plans for the complete system, not a measured saving from Vera CPUs alone. HLRE-5 combines CPU and GPU partitions, high-speed networking, large storage tiers and liquid cooling. DKRZ also expects to reuse more of the system’s waste heat on the University of Hamburg campus. Its release describes the CPU partition as bringing Arm architecture into the center’s core system for the first time and improving energy efficiency for workloads that do not run on GPUs.
How much of the performance story is proven?
NVIDIA reports that Vera delivered 1.9 times the performance and 2.4 times the performance per watt of unspecified x86 systems on one OpenFOAM motorBike Large airflow simulation. For the Weather Research and Forecasting model, NVIDIA reports projected gains of 1.6 times in performance and 2 times in energy efficiency. These are vendor test results and projections, not independent measurements from a deployed HLRE-5 system.
The distinction matters because performance depends on the exact program, compiler, memory setup, competing system and power measurement. A result on one airflow case cannot establish that every climate model or scientific application will be faster. The strongest evidence will come from published test configurations, independent evaluations and DKRZ’s results after the system is installed.
Why this is a notable shift
NVIDIA’s March 2026 introduction framed Vera around agentic AI and reinforcement learning. The new rack proposal broadens the pitch to traditional high-performance computing: weather, climate, molecular dynamics and other memory-heavy work. It also shows the company selling a whole computing platform—processors, memory, networking, software and cooling—rather than a standalone chip.
For scientists, software support may matter as much as hardware. NVIDIA says its HPC SDK 26.11 is available with compilers and libraries optimized for Vera and that a next-generation Fortran compiler is entering public beta. Fortran remains common in scientific codes, so easier access to tuned compilers could lower migration effort. The claim that existing programs can run faster without being rewritten remains something developers will need to verify against their own applications.
What to watch next
Vera CPU racks are expected from Dell Technologies, HPE, Lenovo and Supermicro beginning in early 2027; HLRE-5 is scheduled for early 2028. Until those systems are available, buyers and researchers should treat the specifications, benchmark advantages and deployment dates as company or project plans. Watch for complete benchmark methods, independent tests, software compatibility reports and measurements of energy per useful simulation—not headline throughput alone.
Sources
NVIDIA: Vera CPU rack-scale platform announcement (November 16, 2026)
DKRZ and Bull: HLRE-5 contract announcement (October 21, 2026)
NVIDIA: Vera CPU introduction for agentic AI (May 31, 2026)
Illustration: Original conceptual artwork generated for this article. It does not depict NVIDIA Vera hardware or the DKRZ system.
