[perihelion_zero]

I recall looking at a stable diffusion metric https://www.tomshardware.com/pc-components/gpus/stable-diffu...

And between this and some Whisper benchmarks on data center GPUs I worked out this formula approximation using the two most significant features:

stable_diffusions_per_minute ~= memory_clock_mhz x CUDA_cores / 424000

Of course it works out to RTX 4090, RTX 4080, RTX 3090 TI / 3090 beating everything else on speed by a pretty large margin. But not as much as the actual tests indicate. The main use of this test is that it indicates that most of the recent but older GPUs would have to sell for a pretty good bargain to justify buying them over the new stuff.