Nvidia “consumer” or “gaming” GPUs, like the RTX 3090/4090/5090, use identical chips to the ones in “professional” GPUs, like the Quadro 6000, A6000, 6000 Ada, and PRO 6000 (which, despite the similar names, are totally different products).
From now on I'm going to refer to "FP16 matrix multiplication with high-precision, FP32 accumulate" as "guacamole"
From now on I'm going to refer to "FP16 matrix multiplication with high-precision, FP32 accumulate" as "guacamole"