Best driver version for Tesla M2050 to maximise performance

Hello,

anybody have any idea which driver version would result in the best performance from a Tesla M2050?

kind regards,

Andrew

The M2050 is a device with compute capability 2.0 and as such supported by the currently shipping version of CUDA, i.e. CUDA 8.0, and the latest driver package from NVIDIA. Why not use that?

Support for M2050 is dropped from CUDA 9, by the way. CUDA 9 is currently in a EA status, meaning public releases are expected to follow, “soon”.

So IMO you’re in the twilight of reasonable usability of this ~7 year old GPU. I don’t know what you paid for the GPU or the fancy 3D printed fan assembly you made for it, but for not a lot of US$ (say, less than $200) you can get a fairly modern GPU that might outperform that GPU, depending on code specifics.

Hopefully you got it for, like, free.

In terms of DP FLOPS per dollar an M2050 might be a pretty sweet deal compared to newer GPUs, even taking into account the cost of constructing a custom shroud/fan assembly. The energy efficiency is horrible, of course, and it is nearing the end of its serviceable life fast, as txbob points out.

It was like $70 US shipped to Australia. 3d printing was done by a friend and I designed it all. It was a fun little project and the temperatures are very reasonable.

The reason I asked about the drivers is I know the gaming community has found that with older GTX cards that updating to the latest driver can sometimes result in considerable performance reduction. I experienced this when updating the drier on the previous GPU (560 ti) and the performance dropped considerably until I rolled back to version 314.22.

The code I’m running is a 2 dimensional hydraulic model that uses double precision floating point. It was more of a proof of concept, that for not too much $ we can reduce model run times significantly. Ideally I’d probably want to get a Titan Z or one of the higher DP FLOPS Tesla K or P series cards. Our business doesn’t do much GPU accelerated work yet, so trying to get budget for expensive GPUs is difficult.