Intel Xeon 6517P on the OpenMetal Large v5

Q: What is the Intel Xeon 6517P (Granite Rapids) and why did OpenMetal pick it for the Large v5?

The Xeon 6517P is Intel’s Granite Rapids server CPU on the Intel 3 process node — OpenMetal selected it for the Large v5 to keep the same 32-core / 64-thread footprint as the Large v4 while gaining 14% higher base clocks, 92% more L3 cache per socket, and a 15% reduction in TDP.

See large v5 specs

The 6517P delivers 16 cores / 32 threads per socket at 3.2 GHz base / 4.2 GHz turbo, with 72 MB of L3 cache per CPU (144 MB total in the dual-socket Large v5 configuration). Granite Rapids moves to the Intel 3 process node, replacing the Intel 7 node used on the Emerald Rapids 6526Y in the Large v4. The smaller node enables the lower 190W TDP per socket while delivering higher sustained clocks — more thermal headroom for turbo and less rack heat to plan around.

OpenMetal’s selection criteria for the Large v5 prioritized per-core performance and cache footprint over raw core count. Customers using the Large v4 had built deployments around the 32C/64T target; jumping to a higher core count would have changed application sizing assumptions. Holding cores steady while nearly doubling L3 cache and stepping memory to DDR5-6400 produces a meaningful per-core uplift on workloads where the v4 was DRAM-bandwidth or cache-bound: PostgreSQL OLTP, ClickHouse and Presto analytics, Proxmox/KVM VM density, and JIT-compiled query engines.

Beyond clock and cache, the 6517P brings Intel AMX for INT8/BF16 matrix operations (CPU-based ML inference), AVX-512 with dual FMA, Intel QAT for cryptographic and compression offload, and full Intel SGX enablement with 128 GB EPC per socket. PCIe 5.0 connectivity is delivered through 88 lanes per processor, feeding the Micron 7500 MAX NVMe drives and LACP-bonded networking directly with bandwidth headroom for additional drives or accelerators.