Planning an EPYC 9175F workstation

An EPYC workstation configuration

I considered the 5th Gen AMD EPYC 9175F (Turin) for CPU inference of 1T-parameter MoE models. The proposed setup uses its large L3 cache, high clocks, and 12-channel memory.

The candidate was HPCT WCE51-GP, from HPCT (High Performance Computing Technologies). I reviewed memory, expansion, and cooling for LLM infrastructure.

Building or buying

I also considered an EPYC 9005 motherboard such as ASRock Rack GENOAD12M3-2Q. TDP exceeds 300W, with cooling and power requirements for 12-channel DDR5.

The estimate was roughly 520,000 JPY for the CPU and 740,000 JPY for memory. I chose an integrated chassis, power supply, and cooling package rather than assembling these costly components myself. HPCT WCE51-GP uses a Supermicro platform.

Technical Specifications of the HPCT WCE51-GP

The single-socket EPYC configuration uses a full-tower chassis.

CPU and Memory

The planned setup is AMD EPYC 9175F with 768GB DDR5-6400 ECC RDIMM in all 12 slots. The aim is to improve memory bandwidth for LLM decoding.

Expansion and I/O

PCIe 5.0 slots allow later GPU expansion:

• 3 x PCIe 5.0 x16 Slots
• 2 x PCIe 5.0 x8 Slots

Storage includes eight SATA3 ports and three MCIO interfaces (PCIe 5.0 x8) for NVMe SSDs. M.2 supports 22110 drives, including enterprise SSDs with Power Loss Protection (PLP).

Networking and Management

Broadcom BCM57416 provides dual 10GBase-T. A dedicated IPMI LAN port supports remote management, with Supermicro Server Manager (SSM) and SuperDoctor 5 for hardware monitoring.

Ordering and inference plans

The proposed configuration includes 12-channel memory and a 750W - 1500W power supply. Cooling and power management are part of the choice for long inference workloads.

The next step is ordering the system and configuring Ubuntu 24.04 for 1T MoE models such as Kimi-K2.5. This is a configuration review, not an inference benchmark.

Specs Overview