I bought a datacenter GPU that doesn't fit in a normal motherboard, macgyvered the fan with jumper wires, and now I'm running a model that ties with Claude Sonnet 4.6 on benchmarks, all for £200.
Not an AI guy, but I do like using niche hardware wrong to get results cheap. Can anyone tell me what this would be like for gaming or general computing? My 1660 super was a budget pick when I got it back in '18.
Cooling wouldn’t be too hard, I’m currently overengineering an Xbox 360 for that so I’ve done a good deal of learning. Getting of to display video is outside my skillset though.
A cursory search says somewhere between a 3060 and 4060, which seems about right. Games that parallelize well across the GPU cores will benefit, though that benefit will be niche if it exists at all. HBM2 memory is weird for gaming.
You’d see some benefit for sure, but this really is better suited to parallel computing, given the emphasis on CUDA core counts and memory bandwidth. You may also find you run into some latency, since it requires selecting the V100 as your primary GPU but a secondary card as the video output. Integrated GPUs work well for this, since your motherboard may already have this ability. Otherwise, you could use it in tandem with a 1030 or similar to get display out.
Not an AI guy, but I do like using niche hardware wrong to get results cheap. Can anyone tell me what this would be like for gaming or general computing? My 1660 super was a budget pick when I got it back in '18.
A lot of data center GPUs don’t have display outputs or cooling. So you would need to figure out how to cool them.
Cooling wouldn’t be too hard, I’m currently overengineering an Xbox 360 for that so I’ve done a good deal of learning. Getting of to display video is outside my skillset though.
A cursory search says somewhere between a 3060 and 4060, which seems about right. Games that parallelize well across the GPU cores will benefit, though that benefit will be niche if it exists at all. HBM2 memory is weird for gaming.
You’d see some benefit for sure, but this really is better suited to parallel computing, given the emphasis on CUDA core counts and memory bandwidth. You may also find you run into some latency, since it requires selecting the V100 as your primary GPU but a secondary card as the video output. Integrated GPUs work well for this, since your motherboard may already have this ability. Otherwise, you could use it in tandem with a 1030 or similar to get display out.
So I’m picking up that it’s not a viable alternative for the desperate?
Probably not your best option, no