Our research and announcements.
A network running on a chip is a neural simulation, and its speed follows the shape of each operation. We found that one axis, not size, sets the rate. Shaping that axis made a layer pair 3.40× faster and cut speech transcription time by 21.5%.
A tensor's values are fixed. How they are stored and read is a choice. We bound what that choice costs, prove it, and measure it. A kernel that reads the file as it is ran 4.9× faster without moving a value.