Nvidia challenges Intel and AMD with its 88-core Vera CPU

What happened?

Nvidia has taken a historic step by unveiling its first fully custom CPU, named Vera, along with its Olympus cores. This launch, detailed in a recent whitepaper, marks Nvidia's direct entry into the data center CPU market, competing head-on with Intel and AMD. Vera is designed to function both as the main node for AI systems (managing GPUs in the upcoming Vera Rubin platforms) and to host AI agents, an emerging market that does not rely on GPUs.

Vera's architecture is notable: a monolithic compute die with 88 Armv9.2 cores, manufactured on TSMC's 3nm process, surrounded by chiplets for memory and I/O. This configuration, similar to Amazon's Graviton 4, aims to maximize bandwidth and reduce latency between cores. Additionally, Vera supports dual-socket configurations (Vera CPU Superchip) with a 1.8 TB/s bidirectional NVLink-C2C interconnect, offering a total of 176 cores and 352 threads.

Why is it important?

This move is important for several reasons. First, Nvidia, primarily known for its GPUs, is now competing in the CPU segment, a market dominated by Intel and AMD for decades. Second, early adoption by major names like Alibaba, ByteDance, Meta, Oracle, CoreWeave, Lambda, Nebius, and NScale indicates strong demand for specialized AI CPUs. Third, the design of the Olympus cores, although based on Arm IP, showcases Nvidia's ability to innovate in processor architecture, with a 10-wide decoder and 8 integer ALUs, focused on eliminating bottlenecks in executing AI and agent tasks.

The historical context is crucial: Nvidia had already dabbled in CPUs with the Grace series, but using standard Arm cores. With Vera, Nvidia demonstrates its intention to control the entire technology stack, from CPU to GPU, offering integrated solutions that could reduce dependence on x86 architectures.

What consequences will it have?

The consequences are multiple. For Intel and AMD, Vera represents a serious threat, especially in the data center segment where AI is a priority. Vera's high energy efficiency and memory bandwidth (2.4 TB/s in Superchip configurations, double that of AMD Turin Epyc) could attract hyperscalers looking to maximize performance per watt. For the market, Nvidia's entry could accelerate the adoption of Arm architectures in servers, eroding x86's dominance.

Furthermore, the focus on AI agents is strategic: as language models become more efficient, agents that execute logic and reasoning do not require GPUs, and Vera could become the preferred CPU for those workloads. This could change the cost dynamics in data centers, as agents are lighter and require fewer resources.

What should readers know?

Readers should know that, although Vera is a promising CPU, its success will depend on the availability of optimized software and the ecosystem. Nvidia has ensured compatibility with Armv9.2, but developers will need tools to leverage its specific features. Additionally, Nvidia's strategy of selling Vera as a standalone platform (without mandatory GPUs) is a significant shift, as previously its CPUs were complements to its GPUs.

Speculation: although not confirmed, it is likely we will see future versions of Vera with more cores and improvements in the manufacturing process. It is also possible that Nvidia will offer integrated bundles with its GPUs to simplify deployment, but this is only a guess based on market trends.

In summary, Vera is a milestone in Nvidia's history and the CPU market. Its impact will be felt in the coming years, and IT professionals and decision-makers should keep an eye on its evolution.