Overview
- AMD published modeled 100kW rack results that claim its shipping Turin EPYC 9965 is about 2.37 times faster than Nvidia Vera and that a projected 256‑core Zen 6 Venice would deliver about 3.3 times Vera’s rack throughput.
- The company built its estimates from single‑node benchmark runs and a roster of tests that include SPEC CPU 2017 integer rate, SPECjbb2015‑derived Java, NGINX with WRK, redis‑benchmark, memcached with memtier, and TPROC‑C on MySQL.
- Citing June 10’s methodology paper, AMD derived Vera numbers by scaling published Grace measurements by 1.63x and derived Venice 256‑core figures by applying a 1.7x scale to its Turin results plus internal testing, and it calls the outcomes “directional.”
- Independent outlets and benchmarking experts note that rack scaling can break linear gains because of interconnect limits, thermal and power distribution constraints, and packaging or yield issues, so full measured racks and customer deployments are needed to validate the claims.
- Beyond raw numbers, AMD’s paper argues higher core density yields more throughput under a fixed power envelope and highlights supply moves such as a Venice production ramp and Taiwan packaging investment that will affect who can deliver dense AI CPU racks in H2 2026.