All news

First independent EPYC Venice tests show gains of up to 93% over Zen 5, while MEXT replaces some DRAM with flash

Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .

Servers for SI are gaining not only faster processors but also a way to work around the DRAM shortage: access prediction allows inexpensive flash memory to serve as part of RAM without a proportional loss of speed.

Tom's Hardware tested dual-socket servers with AMD EPYC 9996 and EPYC 9686F processors based on Zen 6. In the multithreaded integer SPEC CPU 2026 test, a system with two 256-core EPYC 9996 chips was 93% faster than the previous flagship EPYC 9965, while floating-point tests showed gains of roughly twofold over Zen 5; these are results for specific workloads, not a general improvement across all tasks. The compute chiplets are manufactured on TSMC's 2nm N2P process, and the platform supports up to 256 cores, 16 memory channels and PCIe 6.0. AMD also introduced MEXT: a Linux driver and an SI engine running on one EPYC core predict data accesses, keep active pages in DRAM and move the rest to an NVMe drive. According to AMD, cutting DRAM by 50% in electronic-design workloads preserved 95% of performance, while having 25% of the memory available preserved 80%; in some scenarios, the company promises up to 3.5 times more performance per dollar of memory. AMD is already deploying MEXT in its own server infrastructure.

History

AMD Zen 6: Helion rack and Venice processors

2 events
  1. First independent EPYC Venice tests show gains of up to 93% over Zen 5, while MEXT replaces some DRAM with flashOpen now
  2. AMD shows Helion reference rack for SI agents: up to 32,768 Zen 6 cores in one system