Recently, I want to buy some 30xx series GPUs which are used to do deep learning jobs, so does anyone know whether Mathematica's NetTrain supports 30xx series GPUs?
$\begingroup$
$\endgroup$
6
-
$\begingroup$ Do you mean the latest Nvidia cards, like the RTX 3080 and 3090? I don't have one and can't say for sure but I would be very surprised if they weren't supported.MManny users have the 20 series cards and they seem to work fine. My 1650 is also well supported. It seems likely the newer cards are also supported. Perhaps someone who has one can chime in with an answer for sure, but I would probably buy one on the assumption that it works well. $\endgroup$– Carl LangeCommented Jan 31, 2021 at 9:48
-
1$\begingroup$ Yeah, RTX 3080 or 3090. I have saw the Mathematica's help of TargetDevice, NetTrain only support GPUs as follow: Capability 3.7 Kepler architecture Tesla K80. Capability 5.0 Maxwell architecture GTX 750 etc. Capability 5.2 Maxwell architecture GTX 980 etc. Capability 6.0 Pascal architecture Tesla P100 etc. Capability 6.1 Pascal architecture GTX 1080 etc. Capability 7.0 Volta architecture Tesla V100 etc. Capability 7.5 Turing architecture RTX 2080 etc. $\endgroup$– ChuanNan LiCommented Jan 31, 2021 at 12:11
-
$\begingroup$ As I know the Mathematica's neural network is based on MXNet, my classmates have used RTX 3090 with MXNet, so I guess it's reasonable to use 30's GPU with Mathematica. $\endgroup$– ChuanNan LiCommented Jan 31, 2021 at 12:19
-
1$\begingroup$ Yes; I would expect support for the 30 series to be coming soon, if it doesn't already exist. $\endgroup$– Carl LangeCommented Jan 31, 2021 at 12:21
-
1$\begingroup$ I just discovered that a) Mathematica can't read my 3090 Core Count (returns Indeterminate), and it defaults batch size to 64!!! I tried to create a batch size equal to the number of cores, but the max, apparently is 10k, which might be the reason why true Core Count is broken as it's 10496 cores. Either way, with 10k batch size, the demo NetTrain runtime went down from 24s to 16s. Still sluggish, but much faster than before. $\endgroup$– Gregory KlopperCommented Apr 20, 2021 at 3:51
|
Show 1 more comment
2 Answers
$\begingroup$
$\endgroup$
At the time of writing this (Feb. 6, 2021), Mathematica 12.2 doesn't support these cards, see for example:
Extremely Long GPU Initialization Times on Mathematica 12.2 and RTX3090
and
https://community.wolfram.com/groups/-/m/t/2141352
The reason seems to be that the MXNet paclet hasn't been updated to a compatible version yet, so one should presumable expect to see this issue solved sometime in the (near?) future...
answered Feb 6, 2021 at 22:54
$\begingroup$
$\endgroup$
1
2021-03-12
Yes, I can run NetTrain with RTX3060 on Window 10
But, I'd waited for a long time the first time without any tips.
answered Mar 12, 2021 at 7:06
-
$\begingroup$ Mine works too, but evaluation times are not at all what I would expect from 3090. When I run the example given in the help file under TargetDevice option, my 24-core Threadripper 3960x finishes in about 33 seconds, but GPU finishes in only about 23-24 seconds. I would expect a MUCH faster execution from 3090 on such a simple net. $\endgroup$ Commented Apr 20, 2021 at 1:00