Zen 6 Venice 20% Faster in Single-Core! AMD Executive: We Expected to Barely Beat NVIDIA's Vera CPU, But Advantage Was Much Larger

AMD recently released estimated SPEC CPU 2026 benchmark comparisons between its upcoming Zen 6 architecture 'Venice' server CPU and NVIDIA's Vera CPU at the Advancing AI 2026 event in San Francisco. Ravi Kuppuswamy, Vice President of Enterprise Solutions at AMD, said they initially expected to lead Vera by about 10%, but real-world testing showed a 20% advantage in both integer throughput and single-core performance—despite Venice not yet being fully optimized.

According to China's 'Kuaikeji' (Fast Tech), AMD's presentation revealed that the dual-socket 256-core, 600W EPYC 9996 outperformed NVIDIA's dual-socket 88-core Vera CPU with 2.2 times higher integer throughput and 1.2 times better single-core performance.

A closer look at the data shows that throughput comparisons are heavily influenced by core count and TDP. EPYC has nearly 2.9 times more cores than Vera and a 150W higher TDP, making the overwhelming performance lead less surprising.

AMD counters that comparing against Vera's single 88-core SKU isn't fair, but NVIDIA maintains a consistent stance: Vera is a specialized CPU designed specifically for AI factories and is not intended to compete head-on with general-purpose 256-core x86 CPUs.

NVIDIA disclosed Vera's total score as 925 (÷88 ≈ 5.3 per core), while AMD's 96-core high-frequency EPYC achieved 1210 (÷96 ≈ 6.3 per core), indicating an ~18.8% per-core lead—close to the claimed '1.2x' figure.

However, both results are labeled 'estimated' and have not been submitted for official SPEC certification. Both use GCC 15.2 but haven't disclosed optimization flags; differences in compiler settings alone could shift rankings. Until official results are published, definitive conclusions remain premature.

More critically, the choice of battlefield matters: so far, only integer performance has been compared, while floating-point and vector (AVX-512) results remain unreleased.

AMD has long held strength in server-side vectorization, but AI training and inference rely heavily on matrix operations, where floating-point and matrix throughput are far more relevant than integer performance. Vera's true capability when integrated with NVIDIA's full software and hardware stack remains unknown. In short, AMD's announcement is a psychological prelude—not a final verdict.