NVIDIA于7月21日发布的Vera芯片45页白皮书展现了强大的硬件基础，但其对比宣传存在多个问题。[1]
Vera采用88核心单芯片设计，搭载Olympus 10宽Arm v9.2核心，配备2MB私有L2缓存、164MB共享末级缓存和8个LPDDR5X接口，实现1.2TB/s内存带宽。[1]独立测试表明该芯片性能表现突出——根据Phoronix 5月的测试，Vera几何平均性能比5GHz EPYC 9575F高10%，比Xeon 6980P高1.55倍，比Grace高1.63倍。[1]在SPEC 2026基准测试中，两颗Vera得分925分，相比两颗EPYC 9755的898分仅领先3%；但若按核心数归一化，Vera快50%。[1]
然而，白皮书在与x86处理器对比时采用了误导性方法。[1]在内存带宽对比上，白皮书声称Turin带宽为400GB/s，但实际测试显示Turin可达570GB/s（达到93%的理论值），这使Vera的带宽优势从3倍降低至1.9倍。[1]此外，白皮书采用了选择性数据和不完整的性能指标。[1]值得注意的是，Vera采用的核心技术并非NVIDIA独有——其值预测类似苹果的实现，图形预测器与Intel Data-Dependent Prefetcher功能相同，而神经分支预测器基于2012年AMD Piledriver的感知器设计。[1]
Nvidia released a 45-page whitepaper on July 21 promoting its Vera server CPU, which features competitive hardware specifications including an 88-core single-chip design, Olympus Arm v9.2 cores with 10-wide instruction issue, 2MB private L2 cache per core, 164MB shared last-level cache, eight LPDDR5X memory interfaces, and 1.2TB/s memory bandwidth [1]. Independent testing conducted by Phoronix in May demonstrated strong performance, with Vera achieving approximately 10% higher geometric mean performance than a 5GHz EPYC 9575F processor, 1.55 times the throughput of an Xeon 6980P, and 1.63 times that of a Grace CPU [1]. However, the whitepaper's marketing claims contain multiple problematic comparisons against x86 processors, including misleading charts, selective data presentation, and incomplete performance metrics [1].
A critical discrepancy emerged regarding memory bandwidth comparisons [1]. Independent testing revealed that AMD's Turin processor actually achieves approximately 570GB/s of memory bandwidth—93 percent of its theoretical maximum—rather than the 400GB/s figure cited in Nvidia's whitepaper, which substantially narrows Vera's claimed three-fold bandwidth advantage to approximately 1.9 times [1]. In SPEC 2026 benchmarks, two Vera processors scored 925 points compared to 898 points for two EPYC 9755 processors, representing only a 3 percent advantage in aggregate performance; when normalized per core, however, Vera demonstrated 50 percent faster performance [1]. Furthermore, several core technical features highlighted in the whitepaper are not proprietary to Nvidia: value prediction techniques resemble those found in Apple processors, the graphics prefetcher functions identically to Intel's Data-Dependent Prefetcher, and the neural branch predictor builds upon design principles from AMD's 2012 Piledriver architecture [1].