NVIDIA Debuts Vera Server CPU Amid Backlash Over 'Misleading' Technical Claims
The company's first custom server CPU targets agentic AI, but critics argue its performance claims rely on cherry-picked compiler benchmarks.
NVIDIA has unveiled Vera, its first custom server CPU designed to optimize the compute stack for "agentic AI" systems. While the hardware represents a significant expansion of NVIDIA's silicon ecosystem, the company's technical presentation has sparked a backlash from the hardware community over its depiction of competing architectures.
Detailed in a 45-page whitepaper, Vera is built on the "Olympus" core architecture, featuring an 88-core monolithic compute die. Each Olympus core is a 10-wide Arm v9.2 design. The technical specifications include value prediction, a graph prefetcher, and 2 MB of private L2 cache per core, supported by 164 MB of shared last-level cache. To handle data-heavy AI tasks, the chip utilizes eight LPDDR5X memory interfaces, delivering a total bandwidth of 1.2 TB/s. NVIDIA claims these specifications allow Vera to deliver 1.8x higher per-core performance in workloads representative of agentic AI.
Technical Controversies
Despite the impressive specifications, the whitepaper has drawn sharp criticism from technical analysts. Writing for Chips and Cheese, analysts argued that NVIDIA used misleading diagrams to disparage x86 architectures, specifically by depicting traditional simultaneous multithreading (SMT) as simple time-slicing to make the Vera design appear superior. The author of the Chips and Cheese critique noted that NVIDIA spent a significant portion of the paper attempting to turn its design choices into a "morality play about x86."
Further scrutiny has fallen on NVIDIA's performance metrics. While the company marketed its results as "agentic benchmarks," critics pointed out that the tests used—cppcheck, llvm, cpython, and gcc—are actually standard compiler benchmarks from SPEC CPU2026. This has led to accusations that NVIDIA cherry-picked specific compiler-heavy tasks to simulate the behavior of AI agents without using a diverse or representative set of actual agentic workloads.
Industry Implications
This move signals NVIDIA's intent to move beyond GPUs and interconnects to control the entire processing pipeline. By developing custom CPUs, NVIDIA can better support AI agents that autonomously execute multi-step tasks, such as system administration and coding. This puts the company in direct competition with Arm-based server chips from AWS Graviton and Ampere, as well as the established x86 dominance of Intel and AMD.
What's Next
The controversy highlights a growing tension between NVIDIA's aggressive marketing and the technical reality of its performance gains. The industry will now be watching to see if Vera's real-world performance in non-compiler tasks holds up to the 1.8x claim. Whether "compiler-heavy" benchmarks truly represent the future of AI agent workloads remains a central point of contention for architects and developers.