Nvidia just announced the RTX Spark – this is AI server tech trickling down to the consumer space with a Grace CPU (20 cores), a Blackwell GPU (6,144 CUDA cores) and 128GB of LPDDR5X. Now here’s what’s next for servers – and maybe one day consumer devices too.
The new Vera CPU is the CPU half of the Vera Rubin platform – the other being the Rubin GPU. Vera promises a 1.8x average speedup over “leading x86 CPUs” (Nvidia didn’t name them explicitly).
Vera is massive – it has 88 Olympus cores (based on the ARM instruction set) with Spatial Multithreading for 176 threads per socket. The processor can be paired with up to 1.5TB of LPDDR5X RAM, which can deliver a whopping 1.2TB/s bandwidth, which is crucial for AI inference.
Vera can be used as a standalone CPU for agentic AI workloads, reinforcement learning, data processing and analytics. Nvidia has even designed the Vera CPU Rack, which houses 256 CPUs for 22,528 cores and 45,056 threads (oh boy).
Alternatively, Vera can be a host CPU used in conjunction with Rubin GPUs. For example, the NVIDIA Vera Rubin NVL72 has 36 Vera CPUs and 72 Rubin GPUs. The CPUs and GPUs can talk to each other at 1.8TB/s using the Nvidia NVLink-C2C interconnect.
Nvidia has already secured key customers – Anthropic (Claude), OpenAI (ChatGPT) and SpaceXAI (Grok) will use Vera CPUs and so will hyperscalers like ByteDance, CoreWeave and Oracle Cloud Infrastructure.
Additionally, Dell, HP, Lenovo and Supermicro will be building standalone Vera CPU systems. Also: Asus, Compal, Foxconn, Gigabyte, Pegatron, Quanta Cloud Technology, Wistron and Wiwynn. Even the New York Stock Exchange is interested – NYSE processes 1.1 trillion messages per day, so it is working with Redpanda and HP to build out new infrastructure.
