4 ms·
Slide: First algorithm for training deep neural nets faster on CPUs than GPUs
- m0zg 7y agoIt's a first algorithm to train _recommenders_ faster on CPUs, not neural nets in general. And the "CPU" they use are actually 2 CPUs which, even after the price drop, sell for $4K+ each.
- jimsmart 7y agoYes, but they're benchmarking those 2x $4k CPUs (which you're complaining about) against a v100 GPU, which costs $7-11k — so where's the problem here? It's not like they're comparing their $8k/worth of CPU against a consumer GTX 1080. Furthermore, there's nothing in their paper that seems to indicate that their new algorithm cannot be generalised to tasks other than recommendation. (Although admittedly I may have missed something)
- m0zg 7y agoNo problem whatsoever. It's just that it's a super narrow problem and it doesn't make GPUs obsolete, which is how people misinterpreting this. In fact it doesn't even save that much money, even if you use a V100. For fp32 it has about the same throughput as a consumer GPU at 7-10x the cost.
- jimsmart 7y agoThe primary discussion thread on this now seems to be here. HTH. https://news.ycombinator.com/item?id=22502131 https://news.ycombinator.com/item?id=22502131
- yayr 7y agosource: https://github.com/keroro824/HashingDeepLearning https://github.com/keroro824/HashingDeepLearning paper: https://arxiv.org/abs/1903.03129 https://arxiv.org/abs/1903.03129