NVIDIA Specifies 192-CPU Vera Rack for Scientific Computing
NVIDIA specified a 192-CPU Vera rack for scientific computing, says partner systems are expected in early 2027, and identified a DKRZ deployment for 2028.
NVIDIA announced a liquid-cooled Vera CPU rack-scale platform for scientific computing, specifying a 192-processor configuration for climate modeling, molecular dynamics, simulation and scientific agents.
The company said Dell Technologies, HPE, Lenovo and Supermicro expect to offer Vera CPU racks beginning in early 2027. The configuration combines about 300 TB of memory, 1 PB of storage and Quantum-X800 InfiniBand networking in an NVIDIA MGX rack design.
Each 1U compute tray contains eight Vera CPUs along with NVMe storage and management components, according to NVIDIA. The company specifies each CPU with 88 custom Olympus cores, 176 threads and serviceable SOCAMM2 LPDDR5X memory capable of up to 1.2 TB/s of bandwidth. Across the 192-CPU system, NVIDIA claims more than 200 TB/s of aggregate peak memory bandwidth — the rate at which the processors can move data to and from memory — for simulation and AI workloads.
The platform reuses chassis, power distribution and cooling components from NVIDIA’s NVL72 rack infrastructure. An earlier NVIDIA description of the Vera CPU Rack says the liquid-cooled architecture can scale to 256 CPUs per rack, while the scientific-computing announcement specifies 192. NVIDIA’s materials do not explain the difference between those configurations.
NVIDIA said its own tests showed Vera delivering 1.9 times the performance and 2.4 times the performance per watt of an x86 comparison system on the OpenFOAM motorBike Large workload. It also projected 1.6 times higher performance and twice the performance per watt on the Weather Research and Forecasting model. Those benchmark and efficiency figures have not been independently verified.
Germany’s climate-computing center DKRZ has selected Vera CPUs for the CPU partition of its planned HLRE-5 supercomputer, but its contract announcement does not say it will use the 192-CPU rack or disclose how many Vera CPUs it will deploy. HLRE-5 will also include a GPU partition based on NVIDIA GB200 Grace Blackwell Superchips and use Quantum-2 InfiniBand networking.
DKRZ awarded Bull a €40 million contract that includes five years of support. The center said HLRE-5 is intended for physics-based simulation and AI-driven climate research, including two simultaneous global climate simulations at 5-kilometer resolution, and is designed to use about 75% less energy per simulated year than its current Levante system. DKRZ expects HLRE-5 to enter service in early 2028.
More news

NVIDIA Introduces Four-Node DGX Spark HPC Developer Kit

AMD-powered Lux supercomputer opens to Oak Ridge researchers

NVIDIA, IBM and Fermilab plan modular quantum-classical platform
