What changed in Zen 6
AMD's Zen 6 architecture builds on Zen 5 and expands it in nearly every direction. The chips use TSMC's N2P process, which offers 18% more speed at the same power or 36% less power at the same speed compared with N3E. Inside the core, a redesigned vector engine natively supports half-precision FP16, and split integer schedulers allow more instructions in flight.
Cores and memory
The single I/O die of previous generations is now split across two IODs, giving the SP7 platform 16 memory channels at the cost of PCIe 6.0 lanes. CCDs grew for the first time since Zen 3: a standard Zen 6 CCD scales to 12 cores, while a Zen 6c CCD packs 32 cores. Shared L3 per core in Zen 6c was doubled.
Test results
- The 256-core EPYC 9996 delivers 1,024 threads in a dual-socket configuration and came out 93% ahead of the previous-generation 768-thread EPYC 9965 in SPEC CPU 2026 integer throughput.
- The frequency-optimized EPYC 9686F stands out in scientific computing and HPC workloads and serves as AMD's competitor to NVIDIA's Vera.
- AMD says Zen 6 brings a 27% average IPC uplift in enterprise, cloud and AI workloads, measured by comparing fixed-frequency Zen 5 with 16-core Zen 6c.
Competitive picture
Intel's Granite Rapids-AP rarely appears in these comparisons. NVIDIA's Vera holds some competitive positions, but Venice is often ahead, sometimes by large margins. Intel's Diamond Rapids and AMD's Venice-X and Verano platforms are expected next year.
