Powering MI50 for AI Training | Generated by AI
Question: Can a Corsair CX650F / 750W PSU power an AMD Radeon Instinct MI50 16GB for AI training?
Answer:
Yes, a Corsair CX650F 650W or 750W 80+ PSU can power an AMD MI50, but there are a few important details.
AMD Radeon Instinct MI50 power requirements:
- TDP: ~300W
- Power connector: 1 × 8-pin PCIe power connector + 1 × 6-pin PCIe power connector (not the same as RTX 4070)
- Recommended PSU: usually 650W+ depending on the rest of the system
The MI50 is a datacenter card, so it does not use the newer RTX 40-series style 12VHPWR connector.
Your PSU:
Corsair CX650F RGB
or the 750W version:
- PCIe 8-pin cables: yes
- Enough wattage: yes
- Quality: acceptable for a single MI50
A typical system:
CPU (Ryzen 7 / Intel i7) 80-150W
Motherboard + RAM 50W
SSD + fans 20W
AMD MI50 ~300W
----------------------------------
Total ~450-520W
So a good 650W PSU has enough headroom.
However, check the PCIe power cables:
MI50 needs:
PSU PCIe cable
|
+---- 8-pin PCIe -> MI50
|
+---- 6-pin PCIe -> MI50
Do not use:
SATA -> PCIe adapter ❌
Molex -> PCIe adapter ❌
because the card can pull close to 300W continuously during AI workloads.
Your Xianyu MI50 listing:
16G HBM2 ECC, performance similar to RTX 3060, AI training, temperature 90°C
The description is mostly reasonable, but some notes:
Performance
MI50:
- 16GB HBM2 ECC
- 10240 stream processors
- ~13.3 TFLOPS FP32
- ~1TB/s memory bandwidth
Compared with RTX 3060:
| GPU | VRAM | Memory bandwidth | FP32 |
|---|---|---|---|
| RTX 3060 12GB | GDDR6 12GB | 360 GB/s | 12.7 TFLOPS |
| MI50 | HBM2 16GB ECC | ~1 TB/s | 13.3 TFLOPS |
For large model inference / training, MI50 can actually feel better because:
- 16GB VRAM
- huge bandwidth
- ECC
- cheap used price
But software is the problem:
- CUDA ecosystem ❌
- PyTorch ROCm support ⚠️
- many AI packages assume NVIDIA
You need ROCm-compatible versions.
At 630 CNY, the MI50 is interesting:
RTX 4070:
12GB GDDR6X
CUDA
~200W
MI50:
16GB HBM2 ECC
~300W
ROCm
For your use case (local LLM experiments), the biggest advantage is:
16GB VRAM + ECC at 630 CNY is extremely cheap.
The main thing I would verify before buying:
lspci | grep VGA
rocm-smi
python -c "import torch; print(torch.cuda.is_available())"
after installing ROCm.
Also check cooling. MI50 cards often run hot because they were designed for server airflow. A desktop case with poor airflow can easily hit 90°C. For long training runs, add a strong side/front airflow or a blower setup.