NVIDIA Vera CPU Rack
CPU for the age of agents at factory scale.
Overview
Rack-Scale Infrastructure for AI Factories
Built on NVIDIA MGX™, the NVIDIA Vera CPU Rack delivers dense, liquid-cooled CPU infrastructure for modern AI factories. As reinforcement learning and agentic AI systems scale, CPUs run the sandbox environments that execute code, use tools, evaluate results, and analyze data that drives results. The NVIDIA Vera CPU Rack features up to 256 interconnected Vera CPUs and provides a fast path to deploying high-density CPU capacity alongside NVIDIA Vera Rubin NVL72 systems, completing workloads up to 80% faster than traditional CPU infrastructure and helping AI factories generate more tokens per dollar.
Breakthroughs
CPU Sandboxes
Environments for Agentic AI at Scale
Reinforcement learning and agentic AI run in continuous feedback loops between models and execution environments. Models generate tokens, code, and queries while CPU-based sandboxes execute actions, evaluate results, and return data for the next step. At scale, thousands to millions of environments run in parallel, often mapped to dedicated CPU cores. Faster per-core performance shortens evaluation cycles, reduces agent wait time, and helps AI factories generate more tokens per dollar.
The NVIDIA Vera CPU Rack is purpose-built to scale these environments across AI factories. A single liquid-cooled rack integrates up to 256 Vera CPUs, supporting more than 22,500 concurrent CPU environments. With dense, deployable rack-scale infrastructure, Vera CPU Rack scales CPU capacity alongside NVIDIA Vera Rubin NVL72 systems, keeping evaluation loops short and AI factories operating at peak throughput.
Performance
Industry-Leading Agentic CPU Performance
Agentic AI is bottlenecked by traditional CPUs. Across an agent's reasoning loop, the CPU compiles generated code, runs Python tool chains, and analyzes software code. NVIDIA Vera accelerates all three workloads by up to 1.8x over leading x86 CPUs, turbocharging the agentic inner loop to maximize AI factory output.
Relative performance based on measured data, and subject to change. NVIDIA Vera CPU with LPDDR5X performance baselined to latest generation x86 CPU.
Features
Explore the Rack-Scale Breakthroughs
Built on NVIDIA MGX, the NVIDIA Vera CPU Rack brings Vera’s agentic AI performance to data-center scale in a dense, liquid-cooled system. With up to 256 Vera CPUs, massive LPDDR5X memory bandwidth, NVIDIA® BlueField®-4 DPUs, and NVIDIA Spectrum-X™ Ethernet networking, Vera CPU Rack gives AI factories a fast path to deploy high-density CPU capacity alongside NVIDIA Vera Rubin NVL72 systems. The result is more concurrent environments, shorter evaluation cycles, and more tokens per dollar.
Technologies
Inside Vera CPU Rack
Specifications1
NVIDIA Vera
| NVIDIA Vera CPU | NVIDIA Vera CPU Rack | |
|---|---|---|
| Configuration | 1 NVIDIA Vera CPU | 256 NVIDIA Vera CPUs |
| Cores | Threads | 88 custom NVIDIA Olympus cores 176 threads |
22,528 custom NVIDIA Olympus cores (88 per CPU) | 45,056 threads (176 per CPU) |
| L2 Cache (per core) | 2 MB | 2 MB |
| Unified L3 Cache | 164 MB | 42 GB (164 MB per CPU) |
| SIMD (per core) | 6x 128bSVE2 FP8 |
6x 128bSVE2 FP8 |
| Memory Capacity | Up to 1.5 TB SOCAMM LPDDR5X |
Up to 400 TB2 SOCAMM LPDDR5X |
| Peak Memory Bandwidth | Up to 1.2 TB/s | Up to 300 TB/s aggregate |
| NVIDIA NVLINK™-C2C Bandwidth | 1.8 TB/s | 1.8 TB/s per CPU |
| PCIe CXL | 88 PCIe Gen 6 (CPU-only) 96 PCIe Gen 6 (Vera Rubin) x16, x8, x4, x2 bifurcation CXL 3.1 |
Up to 22,528 lanes PCIe Gen 6 total; CXL 3.1 |
| NIC | BlueField-4 CX9 Any compatible PCIe NIC |
64x PCIe gen Xx with support for NVIDIA BlueField-4 DPUs |
| Confidential Computing | Yes | Yes |
| Form Factor and Cooling | 1S and 2S Server Air or liquid cooled 250 W to 450 W Configurable TDP |
48U MGX Rack 100% liquid cooled |
1. Preliminary information. All values are up to and subject to change.
2. 200 TB recommended config.
Partners
Meet Our Partners
Get Started
Stay Up to Date on NVIDIA News
Sign up for the latest news, updates, and more from NVIDIA.