← All GPUs
AMD · Datacenter (used) · Vega 20 (GCN 5, gfx906)
AMD Instinct MI50 (32GB)
gfx906 is not in the current ROCm support list; owners run llama.cpp on Vulkan or community ROCm builds. No fan and no display output.
VRAM
32 GB
HBM2 ECC
Memory bandwidth
1,024 GB/s
FP16 compute
not published
Price
Used only
used price not tracked yet
What fits at 4-bit
The 4-bit file plus 8,192 tokens of context and 1.5 GB of overhead, against the VRAM of 1, 2 and 4 cards. Speeds are measured medians where someone has published a run at that card count; a blank speed means no run, not a slow one. Machine pages carry estimates for the gaps.
| Model | Needs | 1 card (32 GB) | 2 cards (64 GB) | 4 cards (128 GB) |
|---|---|---|---|---|
| Qwen3.5 4B4B · Q4_K_M | 4.3 GB | fits | fits | fits |
| Llama 3.1 8B Instruct8B · Q4_K_M | 7.1 GB | fits | fits | fits |
| Qwen3 8B8.2B · Q4_K_M | 7.3 GB | fits | fits | fits |
| Gemma 4 E4B IT8B · Q4_K_M | 6.5 GB | fits | fits | fits |
| Granite 4.2 8B8.8B · Q4_K_M | 7.9 GB | fits | fits | fits |
| Qwen3.5 9B9B · Q4_K_M | 7.0 GB | fits | fits | fits |
| Gemma 4 12B IT11.95B · Q4_K_M | 8.2 GB | fits | fits | fits |
| Ministral 3 14B Reasoning (2512)13.9B · Q4_K_M | 10.4 GB | fits | fits | fits |
| Phi-414.7B · Q4_K_M | 11.3 GB | fits | fits | fits |
| gpt-oss-20b21B, 3.6B active · MXFP4 (native) | 12.5 GB | fits | 78 tok/smedian of 1 run | fits |
| Devstral Small 2 24B (2512)24B · Q4_K_M | 16.1 GB | fits | fits | fits |
| Qwen3.6 27B27B · Q4_K_M | 17.7 GB | fits | fits | fits |
| Gemma 4 26B-A4B IT25.2B, 3.8B active · Q4_K_M | 17.4 GB | fits | fits | fits |
| Qwen3.8 27B27B · Q4_K_M | 18.2 GB | fits | fits | fits |
| Granite 4.2 30B29.3B · Q4_K_M | 20.3 GB | fits | fits | fits |
| Gemma 4 31B IT30.7B · Q4_K_M | 18.9 GB | 25 tok/smedian of 2 runs | fits | fits |
| Qwen3 Coder 30B-A3B Instruct30.5B, 3.3B active · Q4_K_M | 19.5 GB | fits | fits | fits |
| Ornith 1.5 35B-A3B35B, 3B active · Q4_K_M | 21.9 GB | fits | fits | fits |
| Qwen3.6 35B-A3B35B, 3B active · Q4_K_M | 22.4 GB | 90 tok/smedian of 2 runs | fits | fits |
| Nemotron 3.5 Lightning 30B-A3B30B, 3B active · Q4_K_M | 25.3 GB | fits | fits | fits |
| Llama 3.3 70B Instruct70B · Q4_K_M | 43.6 GB | too large | fits | fits |
| Qwen3 Coder Next80B, 3B active · Q4_K_M | 46.9 GB | too large | fits | fits |
| gpt-oss-120b117B, 5.1B active · MXFP4 (native) | 60.2 GB | too large | fits | fits |
| Llama 4 Scout 17B-16E Instruct109B, 17B active · Q4_K_M | 62.4 GB | too large | fits | fits |
| Mistral Small 4 119B (2603)119B, 6.5B active · Q4_K_M | 69.3 GB | too large | too large | fits |
| GLM-4.5-Air106B, 12B active · Q4_K_M | 70.9 GB | too large | too large | fits |
| Mistral Medium 3.5 128B128B · Q4_K_M | 74.0 GB | too large | too large | fits |
| Qwen3.5 122B-A10B122B, 10B active · Q4_K_M | 73.0 GB | too large | too large | fits |
| Nemotron 3 Super 120B-A12B120B, 12B active · Q4_K_M | 82.6 GB | too large | too large | fits |
| Qwen3.8 Flash-Next125B, 6B active · Q4_K_M | 113.1 GB | too large | too large | 49 tok/smedian of 1 run |
Machines built on it
- 4x Instinct MI50 32GB (used) on EPYC 7003~$1,998 + unpriced parts
Measured speeds
One card
| Model | Quant | Backend | Generation | Prompt | Context | Source |
|---|---|---|---|---|---|---|
| Gemma 4 31B IT | UD-Q4_K_XL | llama.cpp gfx906-furnace fork | 28 tok/s | 220 tok/s512-token prompt | 0 | independent benchmark2026-10 |
| Gemma 4 31B IT | UD-Q4_K_XL | llama.cpp (stock master base of fork, build not stated, ROCm) | 22 tok/s | 187 tok/s512-token prompt | 0 | independent benchmark2026-10 |
| Qwen3.5 9B | Q8_0 | llama.cpp (build not stated, HIP) with MMQ nwarps patch | – | 860 tok/s512-token prompt | 0 | llama.cpp benchmark discussion2026-06 |
| Qwen3.6 35B-A3B | UD-Q4_K_XL | llama.cpp (stock master base of fork, build not stated, ROCm) | 90 tok/s | 864 tok/s512-token prompt | 0 | independent benchmark2026-10 |
| Qwen3.6 35B-A3B | UD-Q4_K_XL | llama.cpp gfx906-furnace fork | 90 tok/s | 968 tok/s512-token prompt | 0 | independent benchmark2026-10 |
2 cards
| Model | Quant | Backend | Generation | Prompt | Context | Source |
|---|---|---|---|---|---|---|
| gpt-oss-20b | MXFP4 MoE | llama.cpp b6698 ROCm | 78 tok/s | 1,307 tok/s2,048-token prompt | 0 | llama.cpp benchmark discussion2025-10 |
4 cards
| Model | Quant | Backend | Generation | Prompt | Context | Source |
|---|---|---|---|---|---|---|
| Qwen3.8 Flash-Next | UD-Q4_K_XL | llama.cpp gfx906-furnace fork (ROCm 7.1) | 49 tok/s | 1,977 tok/s8,192-token prompt | 0 | independent benchmark2026-10 |
Speculative decoding, several users, or long context
Kept apart because each measures a different workload from one person chatting near the start of a conversation.
| Model | Quant | Backend | Generation | Prompt | Context | Measured on | Source |
|---|---|---|---|---|---|---|---|
| Qwen3.8 Flash-Next | UD-Q4_K_XL | llama-server, gfx906-furnace fork | 62 tok/s | 1,561 tok/s32,768-token prompt | 0 | 4× Instinct MI50 (32GB) | independent benchmark2026-10 |
| Qwen3.8 Flash-Next | UD-Q4_K_XL | llama.cpp gfx906-furnace fork (ROCm 7.1) | 44 tok/s | – | 32,768 | 4× Instinct MI50 (32GB) | independent benchmark2026-10 |
Specifications
- Memory
- 32 GB HBM2 ECC
- FP16 basis
- No matrix cores. AMD lists 26.5 TFLOPS peak FP16 (vector) and 53 TOPS INT8.
- Board power
- 300 W
- Slots
- 2
- Length
- 267 mm
- PCIe
- Gen 4 x16
- NVLink
- No
- Cooling
- Passive (needs server airflow)
- Software
- Vulkan
- Released
- 2018
- Availability
- used only
- Price basis
- Not sold new.
- Maker page
- amd.com
Sources
- amd.com: 32 GB HBM2 4096-bit, peak bandwidth 1 TB/s, 300 W, PCIe 4.0/3.0 x16, passive, 267 mm double slot, 26.5 TFLOPS FP16, 2 Infinity Fabric links
- rocm.docs.amd.com: ROCm 7.14 supported-GPU list: 7900 XTX/W7900 (gfx1100), 9070 XT/R9700 (gfx1201), MI100 (gfx908) supported; MI50 (gfx906) not supported; Radeon RX cards Ubuntu-only
- localaimaster.com: 1,024 GB/s; no fan; gfx906 dropped from official ROCm, Vulkan and community builds; ROCm vs Vulkan comparison