2 ms·
You can get away with 8-bit floats on many neural network type applications so that might be the idea here.
by astrodust 10y ago
You can get away with 8-bit floats on many neural network type applications so that might be the idea here.
- bitfiddler0 10y agoIt's an 8-bit int type though. Perhaps even that's sufficient for inference?
- astrodust 10y agoFor a lot of network types that will do the job. 2x the performance is often better than 2x the fidelity. You can get higher accuracy with more nodes.