AMD MI50 - 16GB usable
16GB of HBM2 at 1 TB/s, bandwidth that still embarrasses newer cards. Vega is off ROCm's supported list, so expect to pin old versions.
Specifications
| Brand | AMD |
|---|---|
| Model | MI50 |
| Usable VRAM | 16GB |
| Architecture | Vega |
| CUDA / Stream Processors | 3,840 |
| Memory Bandwidth | 1024 GB/s |
| TDP | 300W |
| FP32 TFLOPS | 13.3 |
Current Offers
No fresh used offer today · recently €175 on eBay
GPUDojo is reader-supported. When you buy through links on our site, we may earn an affiliate commission.
Price History
For AI / LLM Use
Good for 14B models. 30B requires aggressive quantization. Older architecture may have limited software support (check ROCm compatibility). Datacenter card with no display output, may need aftermarket cooling.
What Models Can It Run?
- 14B Q6_K, 30B Q3_K (tight)
- 14B Q4_K_M, 7B full precision
- 7B Q6_K, 14B Q3_K (tight)
- 7B Q4_K_M only
Estimated Performance
Generation: ~77 tokens/sec
Prefill: ~238 tokens/sec
Recommended Quantisations
- Q4_K_M for 14B models
- Q6_K for 7B-8B models
- Q8 for 7B if VRAM allows
Pros & Cons
Pros
- High memory bandwidth for fast generation
Cons
- 16GB usable VRAM: may need quantization for 30B+ models
- 300W TDP: high power draw
- Older Vega architecture: verify current ROCm support
- No display output: headless only
- May need aftermarket cooling solution
Community Verdict
No community reviews yet for the MI50. Know a good review? Let us know.