Just finished testing a bit on an Instinct MI100 bare metal server, wanted to purchase some AMD hardware to have the possibility testing more easily locally the needed specific setup, before returning on more powerful cloud servers for real work.
Due even for 6b models 8GB memory is often not enough was interested to the RX 6700 XT as cheapest 12GB VRAM option, but RDNA 3 seems to have introduced some specific Ai acceleration hardware.
Will make a huge difference choosing one or the other?
The 7700XT is much faster in compute but that’s perhaps not as important as it having somewhat greater memory bandwidth (430 vs 380GB/s). That is going to help.
And as you say there’s also the Wave Matrix Multiply Accumulate (WMMA) support which may give a nice boost too.
The RDNA3 part would seem to be the way to go unless pricing is very different.
Here the 7700 XT is priced +30% respect 6700 XT. For comparison the much more loved 7800 XT is priced +60% always respect 6700 XT. Can I expect at least +30% performances with the 7700 XT? I suppose will be also better supported due ROCm officially support at least the 7900.
A more difficult question: I have tested the MI100 so a CDNA1 architecture and to me performs equal or better than the V100 32GB I’m also using for testing. Any way to understand how will perform a 7700 XT respect a MI100?
There’s not much in the way of comparative ML benchmarks out there so hard to say really. There is this test which shows the 7700XT being 300-426% faster than the 6700XT in Stable Diffusion (depending on image size) which is a significant difference.