AMD Zen 6 EPYC Venice tested: dual 256-core chips run 1,024 threads

[Joanne_24]

AMD's next generation server chips have arrived with enormous core counts. Tom's Hardware has published a deep dive and benchmarks of Zen 6 EPYC Venice, including a dual socket system with two EPYC 9996 processors

The EPYC 9996 has 256 cores and 512 threads per socket, runs at 2.55GHz base and 4.1GHz boost, has 1,024MB of L3 cache and uses 600W, configurable from 400W. The EPYC 9686F has 96 cores and 192 threads, 3.4GHz base and 5GHz boost, 384MB of L3 and a 500W rating

Venice is built on TSMC's N2P process and uses the SP7 socket, with 16 channel DDR5 memory tested with DDR5-12800 MRDIMMs, up to 8TB per socket, and PCIe 6.0 with 96 lanes in single socket systems or 128 in dual socket systems. In SPEC CPU 2026 integer rate, the dual 9996 was 93 percent faster than a dual EPYC 9965, and the 9686F was 19 percent ahead of the 9755 and 17 percent ahead of Nvidia's 88 core Vera

The 9686F was also 40 percent ahead of Vera in floating point rate, though Vera won some tests such as Octave, C-Ray and SciMark FFT. AMD claims a 27 percent IPC uplift, but the SPEC figures are unofficial estimates, Venice was compiled with GCC 16 against GCC 15 for the others, and the IPC gain does not translate directly to consumer Ryzen chips

A 93 percent generational gain is huge for data centres, where more performance per rack saves space and power. Memory bandwidth from 16 channels at those speeds is remarkable. The compiler difference means some caution is wise until independent tests are done

Which Venice feature impresses you most? And how will chips like this change cloud computing?

AMD Zen 6 EPYC 'Venice' deep dive and benchmarking - 9686F outclasses Nvidia's Vera, dual EPYC 9996 with 1,024 threads delivers up to 93% performance gain gen-on-genwww.tomshardware.com