NVIDIA CMP 50HX for AI: A Step-by-Step Guide

I’m gradually diving into the AI world. To gain hands-on experience, I bought a GPU specifically for image generation with ComfyUI. Wanting an affordable yet efficient solution, I chose the NVIDIA CMP 50HX - a card originally designed for mining. It has 10 GB of VRAM and 446 second-generation Tensor Cores, and cost me only $75.

Since I couldn’t find a complete guide for using mining GPUs in AI, I’ll share my experience here.

What about drivers?

❌ GeForce Game Ready Driver (v560.94 in my case) didn’t work.
❌ NVIDIA’s website doesn’t offer special drivers for this card.
❌ NVIDIA A100 Data Center driver (v573.39 for Windows) also failed.
✅ The solution: Install CUDA Toolkit (v12.9.1 for me) and NVIDIA cuDNN.

After this, Windows 10 recognized the CMP 50HX without issues.

What about CUDA?

Mining GPUs like the CMP do support CUDA, but the version depends on the model (check "NVIDIA dedicated GPU for professional mining comparison" page). My 50HX supports CUDA 7.5.

What about performance?

Based on specs, the CMP 50HX should be comparable to an RTX 2080 or RTX 3060 Ti.
I also own a GTX 1060 6GB, so I used a Python script (via CuPy) to compare them. Results:

	NVIDIA GTX 1060 6GB	NVIDIA CMP 50HX
Matrix multiplication 4096x4096	0.083 sec	2.032 sec
Memory bandwidth	77.9 GB/s	181.4 GB/s
Vector operations on 100000000 elements	0.540 sec	0.483 sec

Note: The CMP 50HX is slower for matrix math but faster in memory/vector tasks - likely due to its mining-optimized architecture.

I also found a detailed study on using mining GPUs for AI (focused on the CMP 170HX):
Exploration of Cryptocurrency Mining-Specific GPUs in AI

What about image generation?

I tested ComfyUI Desktop v0.4.60.
ComfyUI detected the CMP 50HX, but the first workflow failed with error «CheckpointLoaderSimple. CUDA error: operation not supported. CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1. Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.».

✅ The fix: Disable «Use CUDA malloc for memory allocation» in ComfyUI’s settings "Server-Config".
After this, it ran smoothly! Now I can generate images for free on my own hardware.

Speed test (using "v1-5-pruned-emaonly-fp16.safetensors" base model):

• 512 x 512: ~22 seconds
• 1024 x 1024: ~1 minute 32 seconds
• 2048 x 2048: ~17 minutes 55 seconds

Limitations to know

⚠️ No Vulkan, Ray Tracing, or DirectML support
⚠️ Limited PCIe 1.1 x4 interface, that reduce bandwidth
⚠️ Not detected by some tools (Windows Task Manager, HWMonitor)
⚠️ Old CUDA 7.5 version (may cause compatibility issues with newer AI frameworks)

Conclusion

Old mining cards can be useful in the field of AI!