In the race for the most performant server architecture, AMD has responded to Nvidia's recent publications with its own benchmarks for the Zen 6 'Venice' chips. During the Advancing AI 2026 event, Ravi Kuppuswamy, corporate vice president of compute and enterprise solutions at AMD, said he was 'very happy' to see Nvidia's results, as they enabled a direct comparison between the two architectures. According to AMD, Venice processors offer 2.2x higher throughput and 1.2x higher per-core performance compared to Nvidia's Vera.
The direct comparison between Venice and Vera is based on SPEC benchmarks
AMD used the configurations specified in Nvidia's white paper to run its own SPEC CPU 2026 tests, calling it an 'apples-to-apples' comparison. The results show that the 256-core, 600W Epyc 9996 significantly outperforms the 88-core Vera in overall throughput, an expected outcome given the higher thread count. However, the most significant data concerns per-core performance: AMD calculated a per-core score of 6.3 for its high-frequency Epyc 9686F chip, versus 5.3 for Nvidia Vera, translating to an approximately 18.8% advantage. Kuppuswamy stated he expected a 10% margin, but real tests showed a 20% lead for AMD, noting that optimizations are not yet complete. In a context where Qualcomm has announced Snapdragon chip price increases, the rivalry between AMD and Nvidia in the server market intensifies.
Sponsored Protocol
Technical specifications reveal substantial architectural differences
AMD compared its 256-core Epyc 9996 with the 88-core Vera, both in dual-socket configurations. The TDP difference is 150W in favor of AMD (600W vs 450W). For per-core performance, AMD used a 96-core Epyc 9686F model with boost up to 5 GHz, although the chip has not yet been officially announced. This has raised criticism about the fairness of the comparison, but AMD argues that Nvidia offers Vera only as a single 88-core SKU, making the comparison inevitable. SPECrate_int benchmarks show a score of 925 for Vera and 2070 for Epyc 9996, while per-core values stand at 5.3 and 6.3 respectively (normalized).
Sponsored Protocol
Estimated results require caution before drawing definitive conclusions
AMD labeled its results as 'estimated' because they have not yet undergone SPEC's strict reporting guidelines. The same applies to Nvidia's data. Until both companies present official benchmarks, it is impossible to determine a clear winner. Furthermore, the comparison is limited to integer workloads; vectorized (floating-point) performance has not yet been published, an area where AMD historically excels. The field remains open, but the challenge between SPEC and different architectural philosophies promises further developments in the coming months.
Sponsored Protocol
