Performance Benchmarks Versus Competitors
Nvidia recently provided detailed technical specifications and benchmark outcomes for its data center CPU, the Vera chip. The company published a white paper alongside unofficial SPEC CPU 2026 results, allowing prospective buyers to compare Vera against processors developed by competitors AMD and Intel.
According to reports covering the release, Nvidia compared Vera in a dual-socket configuration with AMD’s Epyc 9755. When tested using the SPECrate integer suite, the overall base score for Vera reached 925 points. This performance surpassed the 898 recorded by the dual-socket Epyc 9755 system, representing a margin of approximately 3%, despite Vera operating with fewer available threads.
Technical Architecture and Capabilities
The architecture powering Vera is centered on Nvidia’s proprietary Olympus core. Unlike designs focused purely on maximizing raw core count, the company engineered the Olympus core to optimize for memory bandwidth and single-thread performance. The chip configuration includes 88 Olympus cores and supports 176 simultaneous multithreading threads.
The memory subsystem is designed to deliver up to 1.2 terabytes per second of combined aggregate bandwidth, utilizing SOCAMM2 LPDDR5X modules. Furthermore, the internal infrastructure features a unified L3 cache measuring 164 MB and an on-die coherency fabric that provides up to 3.4 TB/s of internal bandwidth.
Targeting Agentic AI Workloads
Nvidia specified that Vera was engineered specifically for agentic AI workloads. These types of tasks require robust single-thread progress through code that is heavily dependent on pointers, long dependency chains, and branch-heavy structures.
Independent testing conducted by the cloud platform DeepInfra demonstrated that Vera supports up to 1.6 times a greater number of concurrent AI agents and achieves orchestration speeds up to 2.2 times faster than alternative CPUs.
Market Deployment and Sales Options
The company confirmed that Vera chips had been supplied to major clients, including SpaceX, OpenAI, and Anthropic, in June.
For deployment, Nvidia offers several options: a standalone chip purchase, a liquid-cooled 256-chip rack setup, a dual-chip single-server arrangement, or a combined package known as the Vera Rubin platform when bundled with Nvidia’s Rubin GPUs. The CPU has a thermal design range spanning from 250 to 450 watts and can manage up to 1.5 terabytes of memory.
The company announced at the Computex conference in June that the Vera CPU had entered full production, confirming its specifications of 88 cores and an LPDDR5X memory subsystem capable of 1.2 TB/s bandwidth. Additionally, Hewlett Packard Enterprise unveiled a complementary server, the ProLiant Compute DL394 Gen12, powered by Vera. This system is intended for reinforcement learning and agentic AI orchestration, with availability anticipated in fall 2026.
Market Context
Nvidia enters the server CPU market as a challenger to established players. Currently, AMD accounts for approximately 33% of the server CPU market share, while Intel holds roughly 66.8%. Both companies maintain strong relationships with major cloud and hyperscale service providers.
Speaking at an earlier presentation, Nvidia’s hyperscale vice president, Ian Buck, suggested that the growth of agentic AI had made CPUs “much more integral,” focusing specifically on metrics such as “how fast a CPU can answer one question.” Nvidia provided an estimate that the total server CPU market could eventually reach $200 billion.
