Software & CAM

NVIDIA unveils Vera CPU rack-scale platform for HPC

NVIDIA unveils Vera CPU rack-scale platform for HPC

Key Takeaways

  • Vera CPU rack‑scale system packs 192 custom NVIDIA CPUs, delivering >200 TB/s aggregate memory bandwidth.
  • Each CPU offers 88 Olympus cores, 176 threads, and up to 1.2 TB/s of LPDDR5X bandwidth.
  • The liquid‑cooled chassis provides ≈300 TB of RAM and 1 PB of NVMe storage in a 1U‑per‑8‑CPU tray design.
  • Integrated Quantum‑X800 InfiniBand and SHARP in‑network computing cut inter‑processor latency dramatically.
  • In NVIDIA’s internal benchmarks, Vera out‑performs comparable x86 servers by 1.9 × in raw speed and 2.4 × in performance‑per‑watt on the OpenFOAM motorBike Large airflow case.

NVIDIA Introduces the Vera CPU Rack‑Scale Platform for Extreme HPC

NVIDIA has unveiled Vera, a rack‑scale solution that combines its first fully custom data‑center processor with the company’s flagship InfiniBand networking. The platform is built for the most demanding scientific workloads—climate modeling, molecular dynamics, and other “unaccelerated” simulations that rely heavily on memory bandwidth rather than GPU acceleration.

Architecture Overview

Custom CPU Core Block

  • 88 “Olympus” cores per socket, each supporting 2 threads (176 threads total).
  • SOCAMM2 LPDDR5X memory subsystem, delivering up to 1.2 TB/s per CPU.
  • Designed for high instructions‑per‑cycle (IPC) on data‑intensive kernels.

Rack Composition

  • 192 Vera CPUs per rack, organized into 24 trays (8 CPUs per 1U tray).
  • ≈300 TB of total DRAM capacity (≈1.5 TB per tray).
  • 1 PB of NVMe storage, all housed in an MGX liquid‑cooling enclosure.
  • Re‑uses NVL72 chassis, power‑distribution, and cooling modules to streamline deployment.

Networking & In‑Network Computing

  • Quantum‑X800 InfiniBand (HDR200, 200 Gb/s per port) interconnects every CPU, providing a non‑blocking fabric.
  • SHARP (Scalable Hierarchical Aggregation and Reduction Protocol) offloads collective operations, reducing the time spent on synchronization and data movement.

Performance Highlights

Metric Vera (NVIDIA) Reference x86 Server*
CPU cores per socket 88 Olympus 64‑core Xeon (Ice Lake)
Threads per socket 176 128
Peak memory bandwidth per socket 1.2 TB/s 0.5 TB/s
Aggregate rack bandwidth >200 TB/s ~80 TB/s
Power efficiency (perf/W) 2.4 × higher —
Raw application speed (OpenFOAM motorBike) 1.9 × faster —
Total rack power draw ~650 kW (liquid‑cooled) ~1.2 MW (air‑cooled)

*Benchmark based on NVIDIA’s internal OpenFOAM “motorBike Large” airflow simulation; exact x86 configuration not disclosed.

The combination of high‑bandwidth memory and low‑latency fabric enables Vera to sustain data‑intensive kernels that would otherwise stall on traditional server architectures. In NVIDIA’s own testing, the platform shaved over 50 % off the time required to complete the OpenFOAM benchmark, while consuming less than half the energy of an equivalent x86 rack.

Target Workloads

  • Climate science: Ultra‑high‑resolution atmospheric models that require petabytes of memory bandwidth.
  • Molecular dynamics: Large‑scale protein folding simulations where each timestep moves terabytes of data.
  • Agentic science & AI‑driven discovery: Workflows that blend traditional HPC kernels with AI inference, benefitting from the shared memory bandwidth and SHARP‑accelerated reductions.

Deployment & Serviceability

  • Modular 1U trays allow hot‑swap of CPU modules and NVMe drives without powering down the entire rack.
  • Liquid‑cooling loop is sealed and monitored by NVIDIA’s MGX management software, reducing fan noise and improving thermal headroom.
  • Compatibility with existing NVL72 rack infrastructure means data‑center operators can adopt Vera without major retrofit costs.

Bottom Line

NVIDIA’s Vera rack‑scale platform delivers a purpose‑built CPU solution that tackles the memory‑bound side of high‑performance computing—a segment traditionally underserved by GPU‑centric designs. By pairing a custom 88‑core processor with HDR InfiniBand and SHARP in‑network computing, Vera achieves nearly double the raw performance and more than double the energy efficiency of leading x86 servers on real‑world scientific codes. For organizations that run large‑scale climate, molecular dynamics, or hybrid AI/HPC workloads, Vera offers a compelling, turnkey alternative that maximizes throughput while keeping power and cooling footprints in check.

Related Articles