The video looks at AMD's Zen 6 EPYC launch, using two dual-socket systems: one with EPYC 9996 (256 Zen 6C cores per socket) and one with EPYC 9686F (96 Zen 6 cores per socket). The presenter, Wendell, uses a dual EPYC 9755 Zen 5 system as the main reference. He notes that the parts are largely sold out and that the testing was done in limited time.
On architecture, he describes a new SP7 socket with two I/O dies per socket and eight memory channels each, so 16 per socket. AMD's CCDs hold 12 Zen 6 cores or 32 Zen 6C cores, with 4 MB of L3 per core. He says the Zen 6 core is the same in both, with clocks up to 5.3 GHz and 4.1 GHz respectively. Per-cycle floating-point results were close to Zen 5.
In measurements, he reports the 9686F holding about 5.29 GHz and beating the 9996 on compile work, while the 9996 wins on parallel capacity. A pointer-chase test gave about 152 ns for memory local to an I/O die, about 193 ns for the other die in the socket and about 347 ns across sockets. Blender was 9 to 18 per cent faster than the 9755 across six scenes, and he puts compile gains at roughly 15 to 20 per cent.
His main practical warning concerns NPS modes: he found NPS1 was 9.2 per cent slower than NPS2 in one test and 17 per cent slower on SQLite. He also tried Gemma 3 12B CPU inference, discussed AMD's MEXT memory-tiering feature, and covered the roadmap: Venice, Venice X and Verano in 2027, with Florence and Ravenna later. He concludes the 9996 is a density play and the 9686F a frequency play.
Key takeaways
- The presenter measured the 9686F at about 5.3 GHz on many workloads, which let it match or beat the 256-core part on compile tasks.
- Each socket has two I/O dies, so NPS2 is the natural mode; the video found NPS1 cost roughly 9 to 17 per cent in his tests.
- His pointer-chase latency was about 152 ns local, 193 ns to the other I/O die and 347 ns cross-socket.
- He reports Blender gains of 9 to 18 per cent over the Zen 5 9755, from a two-system comparison.
- CPU inference on Gemma 3 12B decoded faster on one socket than across two, because of limited cross-socket bandwidth.
- Availability is the catch: he says the chips are effectively sold out, though cloud rental is possible.
Our take
For anyone planning server purchases, the useful lesson is the NUMA behaviour: the video suggests BIOS and workload testing matter as much as core count. The results come from one creator, on two systems, with limited time. The presenter says some comparisons, such as Chromium builds, were not perfectly like for like, and AMD claims were checked only in part.
The video fits our report AMD EPYC 'Venice' Zen 6 server chips are shipping, Phoronix reports, which covers the 256-core parts shipping and SMT remaining worthwhile; the video also gives SMT uplift figures but cites AMD's material rather than its own tests. It also pairs with AMD releases AOCC 6.0 compiler and AOCL 6.0 libraries with Zen 6 support, Phoronix reports, given the video's point that software support for new features lags.


