Nvidia has unveiled detailed technical specifications and benchmark results for its new Vera data center Central Processing Unit (CPU). The release, which occurred on Tuesday, provides potential customers with the architectural information needed to evaluate the chip against processors manufactured by key competitors, AMD and Intel.
Performance Benchmarks Against Competitors
The company published a white paper alongside unofficial SPEC CPU 2026 benchmarks comparing Vera directly to AMD’s Epyc 9755. According to reports from Tom’s Hardware, the comparison utilized a dual-socket configuration for both processors.
In the SPECrate integer suite testing, Vera achieved an overall base score of 925. This figure outperformed the dual-socket Epyc 9755 system, which scored 898. The margin represents approximately a 3% advantage for Vera, notable because Nvidia achieved this while operating with a lower thread count than the tested AMD system.
Technical Specifications and Architecture
Vera is engineered around Nvidia’s proprietary Olympus core. Unlike chips prioritizing sheer core density, the design focuses specifically on maximizing memory bandwidth and single-thread performance. The CPU features 88 Olympus cores and supports 176 simultaneous multithreading threads.
The chip’s robust memory subsystem offers up to 1.2 terabytes per second of aggregate bandwidth using SOCAMM2 LPDDR5X modules. Furthermore, the internal architecture includes a 164 MB unified L3 cache and an on-die coherency fabric that delivers up to 3.4 TB/s of internal bandwidth.
Nvidia states that Vera was specifically designed for agentic AI workloads—applications that require strong per-thread progress through complex code, pointer-heavy data structures, and prolonged dependency chains. Independent testing conducted by the cloud platform DeepInfra showed that Vera could support up to 1.6 times more concurrent AI agents and deliver orchestration speeds up to 2.2 times faster compared to alternative CPUs.
Market Context and Deployment
The company confirmed that Vera chips were delivered in June to major clients, including OpenAI, Anthropic, and SpaceX. The available sales options are diverse, ranging from a standalone chip unit and a liquid-cooled 256-chip rack, to dual-chip single-server setups. A bundled offering is also available under the name Vera Rubin platform, pairing the CPU with Nvidia’s Rubin GPUs.
Technically, the chip has a thermal design range of 250 to 450 watts and possesses the capacity to address memory up to 1.5 terabytes.
Challenging the Server Market
Nvidia is entering the server CPU market as a major challenger. Industry data indicates that AMD currently holds approximately 33% of the server CPU market, while Intel maintains a share of around 66.8%. Both companies have established deep relationships with large hyperscale and cloud customers.
At a previous presentation, Nvidia’s hyperscale vice president, Ian Buck, argued that the expansion of agentic AI has made CPUs “much more integral,” specifically highlighting the importance of “how fast a CPU can answer one question.” Furthermore, Nvidia estimates that the total server CPU market could potentially reach $200 billion.
The company had previously confirmed at the Computex conference in June that the Vera CPU was in full production. A separate announcement from the same event featured Hewlett Packard Enterprise unveiling a compatible server, the ProLiant Compute DL394 Gen12. This system is aimed at reinforcement learning and agentic AI orchestration workloads, with expected availability slated for Fall 2026.
