It'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.
Yes, 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)
No 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.
5 comments
[ 3.0 ms ] story [ 21.5 ms ] threadpaper: https://arxiv.org/abs/1903.03129
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)
https://news.ycombinator.com/item?id=22502131