Tesla P100 and Quadro GP100

Hello,

I would like to understand the main difference between latest Quadro and Tesla product lines, related to G/P100 and G/V100.

Since I guess Tesla is more suitable for server GPGPU computing, I don’t understand why the Quadro GP100 have more FP64 performance than the Tesla P100 (5.2 TFLOPS vs 4.7 TFLOPS), but Quadro is also cheaper.

Comparing with latest Volta architecture, I observe the same performance interval: Tesla V100 have 7 TFLOPS and Quadro GV100 have 7.4 TFLOPS for FP64 perfs.

I need to buy a GPGPU server for CUDA computing, so I want to understand.

So, Quadro shows more performance than Tesla, and are much cheaper: why ?

Thanks,

Likely a minor difference in clock speeds. It may be that the Quadro line allows for Boost clocks / dynamic clocking because in essence it’s a graphics centric device that can experience bursty peak loads. Also it’s actively cooled, which allows for the fan to react more quickly to variable heat emission.

About the “much cheaper”: check if the terms and conditions of the Quadro drivers allow use in a datacenter (servers). Maybe they don’t, which could explain some price difference.

For me personally the Quadro GV100 is interesting because it has 32 GB of HBM2 RAM and is available right now, whereas I haven’t seen any Tesla V100 GPUs with 32GB RAM in my market yet (Germany).

The Quadro will not have ECC on caches and registers. The Quadro will have active cooling.

Thanks for replies.

I’m already aware about the cooling differences, but I was not aware of the ECC differences. I don’t know if it’s a mandatory point for our needs. How to decide if we need ECC on internal memory ?

Regarding the CLUF, I don’t plan to use it in a datacenter but in a workstation, or a tower-format supermicro server, so the Quadro would be more suitable for that thanks to the active cooling.

Do you have some prices for the Quadro GV100 ? I can’t find it anywhere I search (France). Regarding the Pascal architecture, I plan for ~6 K€ for a Quadro GP100, and 9 K€ for Tesla P100. So I’m highly interested in Quadro option. Maybe if GV100 is no hugely much expensive I could envisage it.

about ECC: would a single calculation error cause you large risk of damages? like… are you designing bridges or airplanes with your computations?

GV100 32GB is starting at EUR 8700 in Germany currently.

https://www.idealo.de/preisvergleich/OffersOfProduct/6152268_-quadro-gv100-32gb-hbm2-pny.html

Ok, I see. I guess my computations are less critical than the examples provided, and we always run many times the same algorithms with “noise” added to get more precision.

Anyway, thanks for the link, I’ll check with our buy center if we can get a Volta GPU instead of Pascal one.