MATLAB Answers

0

Matlab GPU processing: Nvidia Quadro vs Geforce

Asked by Alex R. on 25 Jan 2014
Latest activity Commented on by Ben Tordoff on 6 Feb 2014
Hi,
I'll be buying a laptop soon and could get either a Quadro K1100M or a Geforce 750M as graphics card. I'll be using it with Matlab and the parallel computing toolbox.
Can anyone confirm if the Geforce 750M is crippled to max FP32 (single precision) and whether the Quadro K1100M can do FP64 (double precision)? Or is it that the Geforce 750M is capable of FP64 but with much lower performance (capped)?
Also, assuming I do FP32 operations on both in Matlab, does anyone have any info on which is faster? They are both using the same Kepler GK107 core and 2GB of GDDR5 memory, however:
  • The 750M is clocked at 967 MHz, GDDR5 clocked at 5 Ghz
  • The K1100M is clocked at 700 Mhz, GDDR5 clocked at 2.8 GHz
How about FP64 on Quadro vs FP32 on Geforce?
Any Matlab-related comparison between Geforce and Quadro (for the above two models, or in general) would also be highly appreciated.
Thanks in advance, Alex.

  0 Comments

Sign in to comment.

2 Answers

Answer by Ben Tordoff on 28 Jan 2014

Hi Alex,
I don't have access to either of these cards to test, but here's some information that might help you:
  • The "Compute Capability" of the card tells you whether it supports true IEEE double-precision maths. Any card with compute capability 1.3 or higher supports IEEE-compliant double-precision maths. Both of these cards are Compute Capability 3.0 and so support doubles.
  • Neither card is optimised for double-precision maths and both will be much better for single-precision than double. They might be up to 10x faster for single-precision depending on the type of computation you are doing.
  • Both cards have the same architecture (2x GK107 SMs) and the same total number of CUDA cores (384).
  • Both have the same amount of memory.
  • From what I've read, neither card has error-correcting memory (standard on the compute-optimized cards and all CPU memory), so you should treat numerical results with caution.
You can get an idea of the difference between single and double precision performance for (non mobile) GeForce and Quadro cards by looking at GPUBench: http://www.mathworks.co.uk/matlabcentral/fileexchange/34080-gpubench.
The cards will be tuned by the OEM to trade-off power consumption (battery life) against performance. One or both may also have "boost" mode enabled, which can change the clock speed dynamically. Note also that for DDR memory, a transfer can happen twice per clock cycle (one up, one down). Sometimes the clock cycle rate is quoted (as per the Quadro), sometimes the transfer rate (as per the GeForce). I suspect, from your figures, the GeForce might be slightly faster in terms of computation but the Quadro looks to have better memory bandwidth. Usually GeForce cards are targetted at gamers, Quadro at higher-end graphics and modelling work, however in this case there doesn't seem to be much to choose between them.
I would look at the other features in each of the laptops in order to choose between them.
Ben

  0 Comments

Sign in to comment.


Answer by Alex R. on 28 Jan 2014

Hi Ben,
Thanks for replying. In the meantime I found out that the 750M does indeed support FP64 but it's performance is crippled by a factor of 1/8, whereas the Quadro is not. This is why the Quadro is faster for FP64 apps, e.g. CAD apps. The Geforce beats it (by a good margin) in everything else. I do care about FP64 performance though.
The Geforce has a faster clock than the Quadro for the GDDR5 too (5 GHz vs 2.8 GHz). You said "the Quadro looks to have better memory bandwidth" ... how come?
Sadly I don't have access to either of the cards (as I said, I'm looking to buy one), so I can't run GPUBench. Do you know if there are any benchmarks already available for me to look at?
Thanks

  1 Comment

You don't need a GPU to look at the results in GPUBench. Download it and then run gpuBenchReport to see the pre-stored data.
As regards the clocks, there can be a big difference between the number of clock cycles per second and the number of memory transfers per second (e.g. 2x the number of clock cycles). Unfortunately, sometimes people quote one and sometimes the other. I was assuming that the massive difference between the Quadro and GeForce (using the same memory technology) was probably down to the two specs quoting different things. Maybe they are genuinely a factor of 2 different? You'd have to ask NVIDIA (or find someone with both cards) to be sure.
There is some info on GDDR5 speeds and bandwidths here: http://en.wikipedia.org/wiki/List_of_device_bandwidths#Video_RAM. I'm not sure it really clarifies things much though.
Ben

Sign in to comment.