PCIe Risers

[Joel]

So I'm considering making the jump from crypto stuff into deep learning. Since I already have gear optimized for mining it seems like a sensible thing to do, plus I'm interested in some really neat projects using deep learning: algo trading, self driving cars, and even machine writing.

A Neural Network Wrote the Next 'Game of Thrones' Book Because George R.R. Martin Hasn't


Anyway, to my question: my rigs have a bunch of 1x -> 16x PCIe risers, so the 7 GPUs in a rig operate with a single PCIe lane. Would this be a handicap for deep learning applications?

 

[big dog]

The received wisdom in GPU computing is that you need 16 PCIe lanes per GPU, and the debate is generally x16 vs. x8, never mind x16 vs. x1, but actual benchmarks (e.g. PCIe X16 vs X8 for GPUs when running cuDNN and Caffe) indicate that using a mining rig for deep learning could be within reason. For training on multiple GPUs at once, the trick will be to focus on deep learning tools optimized for GPU clusters, not those designed for multiple GPUs on the same machine, which will assume the traditional multilane interconnects. PCIe 3.0 x1 is 8 Gbit/s and PCIe 2.0 x1 is 5 Gbit/s, both of which are faster than gigabit ethernet, and neither of which are that much slower than 10GbE. Any GPU cluster deep learning framework should work on a mining rig just fine. For example, here is a tutorial for using a package called Rescale with Tensorflow that should work on a mining rig : Deep Learning with Multiple GPUs on Rescale: TensorFlow Tutorial | Rescale Here is another one called MapR: Deploy Distributed Deep Learning QSS on MapR GPU Cluster, Part 1 | MapR