← All machines

CPU + RAM server · DIY build

EPYC 9355P + 768 GB 12-channel DDR5 (CPU only)

A single-socket Turin box whose 768 GB holds the largest open MoE models at 4-bit, generating at speeds set by ~600 GB/s of memory bandwidth. No GPU, so prompt processing is slow.

Memory for models

760 GBassumption

768 GB of RAM less 8 GB assumed for the operating system.

Memory bandwidth

614.4 GB/s

system RAM, theoretical

Parts

~$3,663 + unpriced parts

motherboard, ram and cooling prices not included

Largest at 4-bit

Kimi K2.6

545.7 GB at 8,192 tokens of context (an assumed limit, see Memory for models)

gpt-oss-20b, 4-bit

–

no bandwidth figure to estimate from

Parts

PartQtyEach
CPUAMD EPYC 9355P (32-core, 8 CCDs, 280 W, SP5)1~$2,598
MotherboardSupermicro H13SSL-N rev 2.01 (12 DIMM slots, DDR5-6400 with EPYC 9005)1not priced
RAM12x 64GB DDR5-6400 ECC RDIMM12not priced
PSUCorsair HX1000i (1000 W Platinum)1~$290
CaseCorsair 7000D AIRFLOW full tower (450 mm GPU clearance, 8 slots)1~$290
CoolingSP5 tower cooler1not priced
StorageCorsair MP600 CORE XT 2TB PCIe 4.0 NVMe1~$485
Priced parts only$3,663motherboard, ram and cooling prices not included

New parts are priced at the maker’s own store or list price. Used cards have no price here yet, and a total with a missing GPU would understate the build, so none is shown.

What it runs at 4-bit

Every model in the catalog at its 4-bit file with 8,192 tokens of context. 39 of 41 fit. Speeds are the median of measured runs on the same chip where any exist; otherwise an estimate from memory bandwidth. Try other quants and context lengths.

ModelNeedsFitGeneration speed
Kimi K32.8T, 104B active · UD-Q4_K_XL1406.8 GBcontext memory estimatedDoes not fit
DeepSeek V4 Pro1.6T, 49B active · Q4_K_M890.5 GBcontext memory not countedDoes not fit
Kimi K2.61T, 32B active · UD-Q4_K_XL545.7 GBcontext memory estimatedFitsassumed limit–
DeepSeek V4.1 Flash552B, 16B active · Q4 (DwarfStar)484.5 GBFitsassumed limit–
GLM-5.3753.3B · UD-Q4_K_XL437.4 GBcontext memory estimatedFitsassumed limit–
DeepSeek V3.2671B, 37B active · Q4_K_M379.6 GBcontext memory estimatedFitsassumed limit–
MiniMax M3428B, 23B active · Q4_K_M244.8 GBcontext memory not countedFitsassumed limit–
Qwen3.5 397B-A17B397B, 17B active · Q4_K_M229.1 GBcontext memory estimatedFitsassumed limit–
GLM-5.3 Flash320B, 18B active · UD-Q4_K_XL187.6 GBcontext memory estimatedFitsassumed limit–
Hy3295B, 21B active · Q4_K_M173.6 GBFitsassumed limit–
MiniMax M2.7228.7B · Q4_K_M132.7 GBFitsassumed limit–
Qwen3.8 Flash-Next125B, 6B active · Q4_K_M113.1 GBcontext memory estimatedFitsassumed limit–
Nemotron 3 Super 120B-A12B120B, 12B active · Q4_K_M82.6 GBcontext memory estimatedFitsassumed limit–
Qwen3.5 122B-A10B122B, 10B active · Q4_K_M73.0 GBcontext memory estimatedFitsassumed limit–
Mistral Medium 3.5 128B128B · Q4_K_M74.0 GBFitsassumed limit–
GLM-4.5-Air106B, 12B active · Q4_K_M70.9 GBFitsassumed limit–
Mistral Small 4 119B (2603)119B, 6.5B active · Q4_K_M69.3 GBcontext memory estimatedFitsassumed limit–
Llama 4 Scout 17B-16E Instruct109B, 17B active · Q4_K_M62.4 GBcontext memory not countedFitsassumed limit–
gpt-oss-120b117B, 5.1B active · MXFP4 (native)60.2 GBcontext memory estimatedFitsassumed limit–
Qwen3 Coder Next80B, 3B active · Q4_K_M46.9 GBcontext memory estimatedFitsassumed limit–
Llama 3.3 70B Instruct70B · Q4_K_M43.6 GBFitsassumed limit–
Nemotron 3.5 Lightning 30B-A3B30B, 3B active · Q4_K_M25.3 GBcontext memory estimatedFitsassumed limit–
Qwen3.6 35B-A3B35B, 3B active · Q4_K_M22.4 GBcontext memory estimatedFitsassumed limit–
Ornith 1.5 35B-A3B35B, 3B active · Q4_K_M21.9 GBcontext memory estimatedFitsassumed limit–
Qwen3 Coder 30B-A3B Instruct30.5B, 3.3B active · Q4_K_M19.5 GBFitsassumed limit–
Gemma 4 31B IT30.7B · Q4_K_M18.9 GBcontext memory estimatedFitsassumed limit–
Granite 4.2 30B29.3B · Q4_K_M20.3 GBFitsassumed limit–
Qwen3.8 27B27B · Q4_K_M18.2 GBcontext memory estimatedFitsassumed limit–
Gemma 4 26B-A4B IT25.2B, 3.8B active · Q4_K_M17.4 GBcontext memory estimatedFitsassumed limit–
Qwen3.6 27B27B · Q4_K_M17.7 GBcontext memory estimatedFitsassumed limit–
Devstral Small 2 24B (2512)24B · Q4_K_M16.1 GBFitsassumed limit–
gpt-oss-20b21B, 3.6B active · MXFP4 (native)12.5 GBcontext memory estimatedFitsassumed limit–
Phi-414.7B · Q4_K_M11.3 GBFitsassumed limit–
Ministral 3 14B Reasoning (2512)13.9B · Q4_K_M10.4 GBFitsassumed limit–
Gemma 4 12B IT11.95B · Q4_K_M8.2 GBcontext memory estimatedFitsassumed limit–
Qwen3.5 9B9B · Q4_K_M7.0 GBcontext memory estimatedFitsassumed limit–
Granite 4.2 8B8.8B · Q4_K_M7.9 GBFitsassumed limit–
Gemma 4 E4B IT8B · Q4_K_M6.5 GBcontext memory not countedFitsassumed limit–
Qwen3 8B8.2B · Q4_K_M7.3 GBFitsassumed limit–
Llama 3.1 8B Instruct8B · Q4_K_M7.1 GBFitsassumed limit–
Qwen3.5 4B4B · Q4_K_M4.3 GBcontext memory estimatedFitsassumed limit–

Measured speeds

Nobody has published a run on this chip with a stated model, quant and backend that we could find. The speeds above are estimates.

Specifications

GPUs
none (CPU only)
System RAM
768 GB
RAM bandwidth
12 channels x 6400 MT/s x 8 bytes = 614.4 GB/s (AMD lists 614 GB/s per socket); 576 GB/s if the board runs DDR5-6000
Power supply
1000 W
Estimated load
EPYC 9355P 280 W + 100 W = 380 W

Before you build it

  • Fill all 12 channels and pick a CPU with enough CCDs: owners measured ~200 GB/s on a 2-CCD EPYC because the CCD-to-IO-die links, not the DIMMs, become the limit. The 9355P has 8 CCDs.
  • DDR5-6400 on Turin needs board support: the H13SSL-N runs DDR5-6000 on rev 2.0 and DDR5-6400 on rev 2.01, and AMD says 6400 needs OEM enablement and a BIOS update.
  • Server RDIMMs cost more per GB than desktop kits in the 2026 shortage: about $2,245 per 64 GB DDR5 RDIMM (~$35/GB) in the Oct 2026 retail tracker, so 12 sticks alone run well over $20,000.

Sources

Compared with

Similar machines