Nvidia Vera CPU is punching hard against the competition according to early reviews. Grace Blackwell architecture maturing.

[GreenEcho]

HPCwire coverage this week highlighted early performance data on Nvidia's Vera CPU showing it competing strongly against AMD EPYC and Intel Xeon in data centre compute workloads. The Grace Blackwell architecture combining the Vera CPU with Blackwell GPU in a unified memory system is the infrastructure that Nvidia is betting will dominate the agentic AI era.

The Vera CPU is not Nvidia's first CPU but it is the first built specifically to complement the Blackwell GPU rather than being a general purpose processor. The NVLink connection between Vera and Blackwell allows the CPU and GPU to share memory directly at 900 GB per second.

https://www.hpcwire.com/

[FridayFeeling]

900 GB per second between CPU and GPU over NVLink versus the PCIe bandwidth of a normal server is the architectural advantage that makes the Grace Blackwell system interesting beyond raw benchmark numbers

[Cass]

Nvidia building a CPU that is specifically designed to work with their GPU rather than being a general purpose processor is the vertical integration play that Intel and AMD should have seen coming

[RogueDepot]

AMD EPYC and Intel Xeon have decades of software optimisation behind them. Vera performing competitively in early benchmarks while being architecturally purpose-built for AI is a meaningful result

[MJF_Fan]

The agentic AI workload has different CPU requirements from traditional HPC. More parallel coordination, more memory bandwidth, less pure floating point compute. Vera is tuned for the actual workload

[KnotKnull]

If Nvidia can capture both the GPU and CPU budget in an AI data centre they move from being the GPU company to being the AI infrastructure company. That changes the addressable market significantly

[FrostBear]

The competitive data from actual deployments rather than synthetic benchmarks is what will determine whether Vera captures real market share from EPYC and Xeon