NVIDIA published a 45-page whitepaper on Vera, its first server CPU built around the Olympus core — an 88-core, 10-wide Arm v9.2 design with value prediction, a graph prefetcher, 2 MB private L2 per core, 164 MB shared LLC, and 1.2 TB/s LPDDR5X bandwidth. The hardware is genuinely impressive and early independent benchmarks confirm Olympus is competitive. However, the whitepaper's competitive narrative is riddled with problems: it misrepresents x86 SMT as leaving resources idle, presents an optional 32-NUMA-node EPYC configuration as the default, relabels standard SPEC CPU workloads as 'agentic benchmarks,' uses undefined cross-ISA performance counters as causal proof, and attributes a memory bandwidth advantage to monolithic die design when the actual gap is explained by memory interface specs. Independent Turin testing shows NVIDIA's EPYC bandwidth figures are significantly understated (~400 GB/s vs ~570 GB/s measured), reducing Vera's bandwidth lead from ~3x to ~1.9x. The RL training result is presented as a pictogram with no methodology. The conclusion: Vera is a strong CPU that doesn't need the misleading marketing wrapped around it.
