Details
- NVIDIA launched the Vera CPU at GTC, purpose-built for agentic AI and reinforcement learning, delivering twice the efficiency and 50% faster performance than traditional rack-scale CPUs.
- Key collaborators include hyperscalers like Alibaba, ByteDance, Meta, Oracle Cloud Infrastructure, CoreWeave, Lambda, Nebius, Nscale; manufacturing partners Dell, HPE, Lenovo, Supermicro, ASUS, others.
- Vera features 88 custom Olympus Armv9.2 cores with Spatial Multithreading for 176 threads, up to 1.2 TB/s LPDDR5X memory bandwidth at half the power of rivals, NVLink-C2C for 1.8 TB/s CPU-GPU bandwidth.
- Builds on Grace CPU success with a new 256-CPU liquid-cooled rack via MGX architecture sustaining 22,500 concurrent environments; pairs with Rubin GPUs in NVL72 platform, doubling prior gen performance.
- Adopters like Cursor, Redpanda report 5.5x lower latency; national labs (TACC, Los Alamos) and clouds (Cloudflare, Together.AI) plan deployments; full production now, available H2 2026.
Impact
NVIDIA Vera CPU sets a new standard for AI factories by prioritizing energy-efficient, high-bandwidth processing critical for agentic AI scaling. It outpaces Grace and rivals like AMD/Intel in per-core performance and Arm ecosystem integration, accelerating adoption among hyperscalers. As Rubin platforms enter production, Vera enables broader access to advanced RL and inference, intensifying competition in data center AI infrastructure.
