Ask HN: Why do all cloud providers have Tesla K80 vs. cheaper Titan X or 1080ti?

For Tesla: K80 seems to have higher double precision flops

For Titan X/1080ti: > In terms of raw performance the titan x and 1080 ti have more flops per GPU. > They are significantly cheaper (~1000$ vs 4000$)

There are numbers which show that the common nets like AlexNet are faster on Titan X (https://plot.ly/~JianminSun/4/nvidia-titan-x-pascal-vs-nvidia-tesla-k80/). However, most cloud providers like Amazon, Azure and Google recently introduced gpu instances with K80's. Wouldn't it be better to just have cloud instances with 1080ti s/Titan X s ?

2 comments

[ 2.1 ms ] story [ 7.2 ms ] thread

webmaven 9y ago ↗

Consider that the K80 gives you more FLOPS-per-$ (just the chip cost, ie. CapEx), and also more FLOPS-per-watt (ie. OpEx).

foo80 9y ago ↗

K80 has ECC memory, while enthusiast card like 1080 do not. K80 is dual gpu so it takes less space (more gpu per node)