← All machines

Multi-GPU · DIY build

2x Radeon RX 7900 XTX

AMD's cheapest new 48 GB: two 24 GB RDNA 3 cards running llama.cpp on ROCm/HIP or Vulkan. Same VRAM as 2x 3090 without CUDA.

Memory for models

46.5 GB

VRAM across all cards

Memory bandwidth

960 GB/s

per card; 2 cards take turns on each token

Parts

~$5,637

sum of parts

Largest at 4-bit

Llama 3.3 70B Instruct

43.6 GB at 8,192 tokens of context

gpt-oss-20b, 4-bit

~120–175 tok/sestimate

no measurement on this hardware yet

Parts

PartQtyEach
GPUAMD Radeon RX 7900 XTX2~$999
CPUAMD Ryzen 9 9950X (16-core, 170 W)1~$649
MotherboardASUS ProArt X870E-CREATOR WIFI (2x PCIe 5.0 x16 slots, x16 or x8/x8)1~$550
RAMCorsair Vengeance 64GB (2x32GB) DDR5-6000 CL301~$1,275
PSUCorsair HX1500i (1500 W at 115-240 V, 1200 W at 100-115 V)1~$390
CaseCorsair 7000D AIRFLOW full tower (450 mm GPU clearance, 8 slots)1~$290
StorageCorsair MP600 CORE XT 2TB PCIe 4.0 NVMe1~$485
Total$5,637

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. 21 of 41 fit. Speeds are the median of measured runs on the same GPUs 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 estimatedDoes not fit
DeepSeek V4.1 Flash552B, 16B active · Q4 (DwarfStar)484.5 GBDoes not fit
GLM-5.3753.3B · UD-Q4_K_XL437.4 GBcontext memory estimatedDoes not fit
DeepSeek V3.2671B, 37B active · Q4_K_M379.6 GBcontext memory estimatedDoes not fit
MiniMax M3428B, 23B active · Q4_K_M244.8 GBcontext memory not countedDoes not fit
Qwen3.5 397B-A17B397B, 17B active · Q4_K_M229.1 GBcontext memory estimatedDoes not fit
GLM-5.3 Flash320B, 18B active · UD-Q4_K_XL187.6 GBcontext memory estimatedDoes not fit
Hy3295B, 21B active · Q4_K_M173.6 GBDoes not fit
MiniMax M2.7228.7B · Q4_K_M132.7 GBDoes not fit
Qwen3.8 Flash-Next125B, 6B active · Q4_K_M113.1 GBcontext memory estimatedDoes not fit
Nemotron 3 Super 120B-A12B120B, 12B active · Q4_K_M82.6 GBcontext memory estimatedDoes not fit
Qwen3.5 122B-A10B122B, 10B active · Q4_K_M73.0 GBcontext memory estimatedDoes not fit
Mistral Medium 3.5 128B128B · Q4_K_M74.0 GBDoes not fit
GLM-4.5-Air106B, 12B active · Q4_K_M70.9 GBDoes not fit
Mistral Small 4 119B (2603)119B, 6.5B active · Q4_K_M69.3 GBcontext memory estimatedDoes not fit
Llama 4 Scout 17B-16E Instruct109B, 17B active · Q4_K_M62.4 GBcontext memory not countedDoes not fit
gpt-oss-120b117B, 5.1B active · MXFP4 (native)60.2 GBcontext memory estimatedDoes not fit
Qwen3 Coder Next80B, 3B active · Q4_K_M46.9 GBcontext memory estimatedDoes not fit
Llama 3.3 70B Instruct70B · Q4_K_M43.6 GBFits~11–17 tok/sestimate
Nemotron 3.5 Lightning 30B-A3B30B, 3B active · Q4_K_M25.3 GBcontext memory estimatedFits~94–137 tok/sestimate
Qwen3.6 35B-A3B35B, 3B active · Q4_K_M22.4 GBcontext memory estimatedFits~125–183 tok/sestimate
Ornith 1.5 35B-A3B35B, 3B active · Q4_K_M21.9 GBcontext memory estimatedFits~128–187 tok/sestimate
Qwen3 Coder 30B-A3B Instruct30.5B, 3.3B active · Q4_K_M19.5 GBFits~119–174 tok/sestimate
Gemma 4 31B IT30.7B · Q4_K_M18.9 GBcontext memory estimatedFits~26–40 tok/sestimate
Granite 4.2 30B29.3B · Q4_K_M20.3 GBFits~26–41 tok/sestimate
Qwen3.8 27B27B · Q4_K_M18.2 GBcontext memory estimatedFits~27–42 tok/sestimate
Gemma 4 26B-A4B IT25.2B, 3.8B active · Q4_K_M17.4 GBcontext memory estimatedFits~93–136 tok/sestimate
Qwen3.6 27B27B · Q4_K_M17.7 GBcontext memory estimatedFits~28–44 tok/sestimate
Devstral Small 2 24B (2512)24B · Q4_K_M16.1 GBFits~33–52 tok/sestimate
gpt-oss-20b21B, 3.6B active · MXFP4 (native)12.5 GBcontext memory estimatedFits~120–175 tok/sestimate
Phi-414.7B · Q4_K_M11.3 GBFits~53–83 tok/sestimate
Ministral 3 14B Reasoning (2512)13.9B · Q4_K_M10.4 GBFits~57–90 tok/sestimate
Gemma 4 12B IT11.95B · Q4_K_M8.2 GBcontext memory estimatedFits~66–104 tok/sestimate
Qwen3.5 9B9B · Q4_K_M7.0 GBcontext memory estimatedFits~83–130 tok/sestimate
Granite 4.2 8B8.8B · Q4_K_M7.9 GBFits~85–133 tok/sestimate
Gemma 4 E4B IT8B · Q4_K_M6.5 GBcontext memory not countedFits~155–243 tok/sestimate
Qwen3 8B8.2B · Q4_K_M7.3 GBFits~94–147 tok/sestimate
Llama 3.1 8B Instruct8B · Q4_K_M7.1 GBFits~96–150 tok/sestimate
Qwen3.5 4B4B · Q4_K_M4.3 GBcontext memory estimatedFits~172–270 tok/sestimate

Measured speeds

2 published runs with one user, plain decoding and at most 4,096 tokens already in context, on 2× Radeon RX 7900 XTX.

Plain single-user runs
ModelQuantBackendGenerationPromptContextMeasured onSource
Gemma 4 31B ITQ8_0llama.cpp (build not stated) Vulkan24 tok/s852 tok/s512-token prompt02× Radeon RX 7900 XTXllama.cpp benchmark discussion2026-08
Gemma 4 31B ITQ8_0llama.cpp (build not stated) ROCm23 tok/s921 tok/s512-token prompt02× Radeon RX 7900 XTXllama.cpp benchmark discussion2026-08

Cards working in parallel

Tensor parallel splits every layer across the cards so they read their share of the weights at the same time, instead of taking turns. It can beat one card’s bandwidth limit, needs fast links between cards, and is kept out of the medians above.

Runs with tensor parallel
ModelQuantBackendGenerationPromptContextMeasured onSource
Qwen3.8 27BQ8_0llama.cpp b10998 (HIP, internal AllReduce, RCCL OFF; PR #27825)37 tok/s1,357 tok/s8,192-token prompt0this machinellama.cpp benchmark discussion2026-09
Qwen3.8 27BQ8_0llama.cpp b10998 (HIP, internal AllReduce, RCCL OFF; PR #27825)37 tok/s1,446 tok/s512-token prompt0this machinellama.cpp benchmark discussion2026-09
Qwen3.8 27BQ8_0llama.cpp b10998 (HIP, RCCL ON)36 tok/s1,580 tok/s512-token prompt0this machinellama.cpp benchmark discussion2026-09
Qwen3.8 27BQ8_0llama.cpp b10998 (HIP, RCCL ON)36 tok/s1,479 tok/s8,192-token prompt0this machinellama.cpp benchmark discussion2026-09
Qwen3.8 27BQ8_0llama.cpp b10998 (HIP, internal AllReduce, RCCL OFF; PR #27825)32 tok/s792 tok/s512-token prompt100,000this machinellama.cpp benchmark discussion2026-09
Qwen3.8 27BQ8_0llama.cpp b10998 (HIP, RCCL ON)31 tok/s823 tok/s512-token prompt100,000this machinellama.cpp benchmark discussion2026-09

Specifications

GPUs
2 × AMD Radeon RX 7900 XTX
VRAM
48 GB
System RAM
64 GB
RAM bandwidth
2 channels x 6000 MT/s x 8 bytes = 96.0 GB/s (dual-channel AM5)
Power supply
1500 W
Estimated load
2x RX 7900 XTX 355 W + Ryzen 9 9950X 170 W + 100 W = 980 W
Software
ROCm, Vulkan

Before you build it

  • ROCm supports the Radeon RX cards (7900 XTX gfx1100, 9070 XT gfx1201) only on Ubuntu; Radeon PRO and Instinct cards cover more distros.
  • On the ProArt X870E-CREATOR both CPU-attached slots share 16 lanes: x16 with one card, x8/x8 with two.
  • Reference 7900 XTX is 2.5 slots and 287 mm, so two need slots three positions apart plus airflow between them.
  • Corsair rates the HX1500i at 1,500 W only on 115-240 V input; at 100-115 V it is a 1,200 W supply.
  • DDR5 prices are at shortage highs: Corsair lists a 64 GB (2x32) DDR5-6000 kit at $1,274.99, so RAM can cost more than a used GPU.

Sources

  • amd.com: price of AMD Radeon RX 7900 XTX
  • shop-us-en.amd.com: price of AMD Ryzen 9 9950X (16-core, 170 W)
  • eshop.asus.com: price of ASUS ProArt X870E-CREATOR WIFI (2x PCIe 5.0 x16 slots, x16 or x8/x8)
  • corsair.com: price of Corsair Vengeance 64GB (2x32GB) DDR5-6000 CL30
  • corsair.com: price of Corsair HX1500i (1500 W at 115-240 V, 1200 W at 100-115 V)
  • corsair.com: price of Corsair 7000D AIRFLOW full tower (450 mm GPU clearance, 8 slots)
  • corsair.com: price of Corsair MP600 CORE XT 2TB PCIe 4.0 NVMe
  • amd.com: 355 W, 287 mm, 2.5 slots, 24 GB
  • rocm.docs.amd.com: gotcha source
  • asus.com: gotcha source

Compared with

Similar machines