4 ms·
Do 1080tis have fp16 support? Seems like a waste if the model can be fp16 trained and you're using full 32bit. Similarly you should probably try a bunch of oth
by bwasti 9y ago
Do 1080tis have fp16 support? Seems like a waste if the model can be fp16 trained and you're using full 32bit.
Similarly you should probably try a bunch of other frameworks (caffe2, cntk, mxnet) as they might be better at handling this non standard configuration.
- dharma1 9y agono double speed fp16 on 1080ti
- shaklee3 9y agoIt does, however, have int8 support.
- gwern 9y agoTheir test case is a GAN. I'm not sure I've ever seen someone train a GAN on int8. It'd probably work...
- dharma1 9y agoyep 4x int8 (44 TOPS)on 1080ti. Is the framework support for there for inference at 4x speed int8 on 1080ti? How about training - I thought you need fp16 minimum for training. I've seen some research into lower precision training (XNOR) but unsure how mature it is. Being able to use 44 TOPS for training on a single 1080ti would be pretty awesome.
- dgacmu 9y agoYes - here's a doc about doing quantized inference in TensorFlow, for example: https://www.tensorflow.org/performance/quantization https://www.tensorflow.org/performance/quantization AFAIK, there's still a bit of a performance gap between just using TF and using the specialized gemmlowp library on Android, but that part's getting cleaned up. Haven't seen much in generalized results on training using lower precision.
- dharma1 9y agoDoes that work with Pascal CUDA8 INT8 out of the box?
- dgacmu 9y agoI'm not sure - I believe it depends on getting cuDNN6 working, and from this bug, I can't quite tell if it works or not (but it's probably not officially supported yet): https://github.com/tensorflow/tensorflow/issues/8828 https://github.com/tensorflow/tensorflow/issues/8828