About the performance of nVidia Quadro K2100M for computing in double precision

[jgpallero]

Hello,

I will probably buy in the next months a mobile workstation (probably a Thinkpad W540) equipped witha nVidia Quadro K2100M. I need a good GPU in order to do some computing in double precision, but I'm a bit confused about the performance of this GPU for computing in double. Apparently, the GPU core is a GK106 (http://en.wikipedia.org/wiki/Nvidia_Quadro), and also apparently the ratio double/single precision for the GK106 chip is 1/24 (http://www.anandtech.com/show/6276/nvidia-geforce-gtx-660-review-gk106-rounds-out-the-kepler-family), so the double precision performance is so poor (worshe than my actual desktop GeForce GTX 550 Ti; yes my 550 Ti is not a mobile GPU, but the K2100M is a Quadro!).

Knows anyone if these data are correct? What is the DP/SP ratio for the Quadro K2100M?

I've seen also that in most of the GPU reviews in www.anandtech.com the ratio FP64/FP32 is presented? I surfed the net and the nVidia webpage and I've not found this information. What is the source of annandtech? Anyone knows a list if GPUs and its FP64/FP32 ratios?

Thanks

 

[ShintaiDK]

Its 1/24th. CPU wise you can outperform it. Its not a card designed for DP. You need GK110(Titan or Quadro/Tesla) with 1/3rd ratio or AMDs Tahiti chips for FP64. Even Hawaii chips are slower than Tahiti i FP64 with 1/8th ratio vs 1/4th for Tahiti.

 

[jj109]

http://en.wikipedia.org/wiki/Comparison_of_Nvidia_graphics_processing_units

This is a pretty handy reference for looking at SP/DP numbers. I think all the uncrippled DP NV cards are around $3k...

Edit: Cripes never mind. Half the charts you want are empty.

 

[Lonyo]

http://www.nvidia.com/content/PDF/l...raphicssolutions-linecard-july13-final-lr.pdf

This implies that some don't support DP at all... although it would seem more reasonable that they simply don't list it because it's so slow (1/24)

 

[jpiniero]

Looks like nVidia only sells mobile GPUs that are crippled in DP.

 

[ShintaiDK]

I am sure they all support FP64. But again, GTX780 and GTX780ti is also 1/24th. This is why the Titan still easily sells for 999$ with its 1/3rd ratio.

 

[ShintaiDK]

Looks like nVidia only sells mobile GPUs that are crippled in DP.

I think all AMDs mobile are "crippled" too. Since you need Tahiti chips before you get any real performance. But you could get away with 1/8th ratio maybe on the mobile front with AMD.

Maybe the best solution laptop wise is simply Haswell with AVX2

 

[jgpallero]

Thanks a lot for your answers!

So, could anyone recommend a mobile workstation with a GPU (nVidia or ATI/AMD) with decent (at least 1/8 FP32) double precision performance?

 

[ShintaiDK]

Thanks a lot for your answers!

So, could anyone recommend a mobile workstation with a GPU (nVidia or ATI/AMD) with decent (at least 1/8 FP32) double precision performance?

You would need AMD HD7970M or HD8970M. AMD got no FirePro cards with 1/8th it seems for mobile.

 

[jgpallero]

You would need AMD HD7970M or HD8970M. AMD got no FirePro cards with 1/8th it seems for mobile.

I've found only the EUROCOM Neptune 3.0 (www.eurocom.com), which can be configured with a HD8970M. This is the first time I see these laptops. Anybody knows something about its quality? Any problem running Linux on them?

Anyway, the HD8970M has a FP64/FP32 of 1/16, which is so poor...

 

[BallaTheFeared]

AMDs Linux drivers (at least for gaming) leave a lot to be desired in my experience.

 

[Anarchist420]

wtf i want to know is how nvidia cripples the DP performance on GTX780 and 780 Ti since they use GK110 like GTX titan does.

so do they divide the core clock speed by a factor of 8 in the bios or in the driver or do they physically damage some part(s) of the chip?

 

[futurefields]

wtf i want to know is how nvidia cripples the DP performance on GTX780 and 780 Ti since they use GK110 like GTX titan does.

so do they divide the core clock speed by a factor of 8 in the bios or in the driver or do they physically damage some part(s) of the chip?

They take a hammer to the chip. Lol.. Oh Anarchist420, what would Anandtech be without you? A lot more sane, that's what!

 

[Bubbleawsome]

Really, how do they do it? I was looking at full compute cards, but $$$.

 

[ViRGE]

Anyway, the HD8970M has a FP64/FP32 of 1/16, which is so poor...

The only GPUs from AMD and NVIDIA with at least a 1/8 FP64 rate are:

GK110
Tahiti
Hawaii

Those are the respective companies' top-tier GPUs. Simply put, they aren't offered in mobile form factors due to their size and power/heat requirements. You won't find a laptop using a standard mobile GPU with the FP64 performance you want. Sorry.

 

[jgpallero]

The only GPUs from AMD and NVIDIA with at least a 1/8 FP64 rate are:

GK110
Tahiti
Hawaii

Are you sure? Here (http://en.wikipedia.org/wiki/Comparison_of_Nvidia_graphics_processing_units) I can see some GK110 (GeForce GTX 780 and 780 Ti) with double performance worst that 1/8

Thanks for your answer

 

[Zodiark1593]

Would laptops using the older Fermi architecture have better fp64? I believe the GTX 580M was based on a similar chip as the desktop GTX 560 (TI), but I'm not sure of the compute capabilities of that chip.

 

[Fox5]

Thanks a lot for your answers!

So, could anyone recommend a mobile workstation with a GPU (nVidia or ATI/AMD) with decent (at least 1/8 FP32) double precision performance?

Sager makes some powerful options. Many of them ship with the Radeon HD 8970M, but the NP9570 can actually run a 6 core Ivy Bridge cpu, which could be faster depending on the workload. 4 core Haswell should also be quite strong in double precision.
http://www.sagernotebook.com/index.php?page=category_browse&selected_cat=13

FYI, a 6 core ivy bridge cpu will do about 200GFlops. A 4 core Haswell chip will do about 250GFlops if the software is recompiled to use AVX2.
The Radeon 8970M does about 140Gflops DP. The Geforce 780M is around 100Gflops DP.

So I guess the questions need to be, what software are you using and are you sure you require double precision performance and does it have to be in a laptop?

If you're compiling the software yourself, there's no reason not to go with a Haswell laptop and recompile it to use AVX2.
If you're using something pre built that scales well with cpu cores, go with a 6 core ivy bridge laptop.
If you're building it yourself, take into account both CPU and GPU usage, or pick a platform.
If you're using something that requires OpenCL, see how it performs on a cpu anyway.
If you're using something that requires CUDA, it also has a cpu backend, plus are you sure it requires fp64?

 

[Anarchist420]

Are you sure? Here (http://en.wikipedia.org/wiki/Compari...ocessing_units) I can see some GK110 (GeForce GTX 780 and 780 Ti) with double performance worst that 1/8 Thanks for your answer

well, the richardheads at nvidia put fuses on every GK110 die that read the programmed configuration to cripple gtx 780s and 780 Tis. so i guess it is kind of like how the asic rating is done.

and because of that i hope the IP system collapses all over nvidia's sorry ass almost as much as i wish it would collapse on intel, microsoft, and AMD.

they've tricked people into thinking that hardware blending is always better overall than software blending under any possible circumstances (as did intel's sorry asses when they didnt make larrabee radically parallel), they've basically said that they're the only one that is good enough to make drivers, they've insulted end-users by telling them they're too dumb to be able to control what optimizations and options are turned on or off making games they didnt create nor are the sole users of look a lot worse, and they've tricked people into thinking that double precision can never be of any possible benefit whatsoever in games just so they could make an extra profit.

in short, nvidia treat gamers like we're noobs and that the imagination of nvidia is better than everyone else's.

but then maybe i will do the same thing some day, who the hell knows what hypocrite i may turn out to be.

 

[Anarchist420]

FYI, a 6 core ivy bridge cpu will do about 200GFlops. A 4 core Haswell chip will do about 250GFlops if the software is recompiled to use AVX2. The Radeon 8970M does about 140Gflops DP. The Geforce 780M is around 100Gflops DP. So I guess the questions need to be, what software are you using and are you sure you require double precision performance and does it have to be in a laptop? If you're compiling the software yourself, there's no reason not to go with a Haswell laptop and recompile it to use AVX2. If you're using something pre built that scales well with cpu cores, go with a 6 core ivy bridge laptop. If you're building it yourself, take into account both CPU and GPU usage, or pick a platform. If you're using something that requires OpenCL, see how it performs on a cpu anyway. If you're using something that requires CUDA, it also has a cpu backend, plus are you sure it requires fp64?

excellent advice