AMD EPYC 9554 (Genoa, 64-core, 12-channel DDR5 host CPU)
Fields
accelerators/amd-epyc-9554-genoaNotes
CPU-only inference host: the 'memory' that matters is the socket's DDR5 channel bandwidth, not VRAM. Decode is bandwidth-bound, so a fully-populated 12-channel Genoa (~460 GB/s theoretical) measures ~7.1 tok/s on Llama-3.3-70B Q4_K_M (ahelpme.com llama-bench, --numa distribute). The single biggest gotcha is channel population: llama.cpp discussion #3167 measured the same class of box dropping to 1.5 tok/s on a 34B model when only ONE DIMM was installed - one stick wastes ~92% of the bandwidth paid for. vram_gb left null: capacity is a function of how many RDIMMs are fitted (up to ~6 TB on SP5), not a fixed package figure. 2026 DRAM crisis has made the RAM the dominant cost of a Genoa build, not the CPU.
Cross-references
This record references no other records — it is a leaf. It is cited by 18.