This is where the market is headed. But kernel optimisation has limitations. Decode silicon is the next step @fastinference