4 ms·
Anyone knows what is this new INT8 instruction does?
by p1esk 10y ago
Anyone knows what is this new INT8 instruction does?
- svennek 10y agoIt is most likely just to show, that it is even faster working with 64-bit ints than floats... So my guess, all integer ops...
- rys 10y agoIt's not all integer ops. The peak throughput comes from a 4-wide accumulated dot product. dp4a if you want to look it up.
- nabla9 10y agoNVIDIA says: 44 TOPS INT8 (new deep learning inferencing instruction) I think it's a related to storing float arrays as arrays of 8-bit integers in memory and converting them into floats just before using. It's 2x more space efficient than fp16 http://www.kanadas.com/program-e/2015/11/method_for_packing_int8_arrays.html http://www.kanadas.com/program-e/2015/11/method_for_packing_...
- astrodust 10y agoYou 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.