PinnedBhaiya Aditya Sinha·Aug 30How I went from 5.6% to 33% of GPU peak: building a CUDA matmul from scratchMatrix multiplication is the dominant operation in neural network inference. Every linear layer, every attention head, every feedforward…