2x AMD EPYC 9654 (Genoa, 24-channel DDR5-4800 total)

Compare 2x AMD EPYC 9654 (Genoa, 24-channel DDR5-4800 total) against another machine in the Arena

Class
CPU
Memory
1536 GB DDR5-4800 (24 channel, 12/socket, fully populated)
Runs (Q4_K_M)
Bandwidth
679.2 GB/s
TDP
720 W
Released
2022-11-10
Price (US)
$21600 used · $28000 new as of 2026-08
Sourcing and disambiguation notes

679.2 GB/s is MEASURED (likwid-bench load, both sockets, NPS1) by the source, vs 921.6 GB/s theoretical for 24x DDR5-4800 (only 73.6% of theoretical achieved - Genoa's dual-socket bandwidth utilization was notably worse than the Turin platform tested alongside it in the same writeup). 2x 96-core EPYC 9654, Gigabyte MZ93-FS1 board. TDP summed from AMD's 360W spec per 9654. 2026-08 build-total price (itcreations.com): dual-socket SP5, 24x DIMM full population (12ch/socket). Breakdown: 2x CPU EPYC 9654 (96c) ~$5,650 each = $11,300, Gigabyte MZ73-LM0 board ~$1,100, 24x32GB DDR5-4800 RDIMM ~$650 each = $15,600. Memory is again the dominant line -- see the EPYC 9554 entry's note on the $520-$2,908/module spread; check it first. Used total ($21,600): 2x pulled/OEM-tray CPUs ~$2,450 each = $4,900, board/memory held flat.

Specification sources

Measurements

Decode is token generation — the speed you feel while an answer streams. Prefill is prompt processing — the wait before it starts. Why bandwidth predicts decode speed.

What verified, single-source and estimated mean, and the same rows with every filter and sort in the benchmarks explorer.

Speed over time

What it can run

Fit is arithmetic, not a measurement — how it is computed. Against 1536 GB; models that fit are listed largest first.