Nvidia powers further into the CPU market with new rack systems packing 256 Vera processors

Nvidia powers further into the CPU market with new rack systems packing 256 Vera processors

Summary

Nvidia has unveiled liquid-cooled rack systems that pack up to 256 of its new Vera CPUs, aimed at handling CPU-heavy parts of AI workloads such as agentic frameworks, tool-calling, reinforcement learning and sandboxed code execution. Vera is an Arm-based design featuring 88 Olympus cores, simultaneous multithreading, a much wider memory bus and high-speed chip-to-chip interconnects. Nvidia claims Vera delivers substantially higher memory bandwidth and improved per-core performance compared with contemporary x86 processors.

The vendor will offer Vera in single- and dual-socket configurations through major ODMs and OEMs, and intends to provide Vera as an alternative to Intel and AMD in its HGX systems. Nvidia’s MGX reference packs up to 256 Vera CPUs and 64 BlueField-4 DPUs in a single liquid-cooled rack, delivering more than 22,500 CPU cores and around 400 TB of memory for dense agent workloads. Several large cloud and datacentre operators including Alibaba, ByteDance, Meta, Oracle, CoreWeave, Lambda, Nebius and NScale have committed to deploy Vera.

Key Points

• New Nvidia rack systems can house up to 256 Vera CPUs in a liquid-cooled MGX reference design.
• Vera uses 88 custom Olympus Arm cores with simultaneous multithreading and a 10-wide decode pipeline featuring a ‘neural branch predictor’.
• Each socket supports up to 1.5 TB of LPDDR5X SOCAMM memory with ~1.2 TB/s bandwidth per socket, giving much higher memory throughput than many x86 alternatives.
• Faster NVLink-C2C interconnects enable up to 900 GB/s unidirectional (1.8 TB/s bidirectional) chip-to-chip transfers between CPUs and GPUs.
• Nvidia positions Vera to avoid CPU bottlenecks in agentic AI workloads where CPUs handle tool invocation, SQL queries and code compilation/execution.
• Vera will be available from major OEMs/ODMs (Foxconn, Wistron, Dell Technologies, Lenovo, HPE) and is already slated for deployments by several hyperscalers and cloud providers.

Why should I read this?

Short version: if you care about AI infrastructure, this matters. Nvidia isn’t just chasing GPUs anymore — it’s building CPUs tailored for the real-world chores agents and large AI systems actually need to do. Expect changes to datacentre design, procurement and performance tuning if Vera ships at scale. Read on to see why memory bandwidth, CPU-to-CPU links and dense rack designs could become as important as raw GPU flops.

Source

Source: https://go.theregister.com/feed/www.theregister.com/2026/03/16/nvidia_vera_cpu_rack/