5

I am currently working with neural nets and I recently made my code much faster by configuring a framework correctly to make use of a GPU. The GPU that is used is a Tesla C2075, which is part of the Nvidia Tesla series.

As enabling GPU computation gave such a huge performance boost, I was curious what the latest GPU of Nvidia is. It is called Titan Z and part of the GeForce 700 series.

However, I am a little bit confused. A "Tesla K40 Workstation Card 12GB DDR5 " currently costs 5500 Euro on Amazon, whereas the Titan Z costs 2750 Euro.

It seems to me that the Titan Z is a much better deal (see specs). I've talked with a friend about this and he thinks that "scientific GPUs" and "gaming GPUs" are different. Could somebody else please elaborate on what is different on those "scientific GPUs" and "gaming GPUs"? Do they have different instruction sets?

(How can one see what might better work for neural net training?)

Specs

The following specs are from http://www.nvidia.com/gtx-700-graphics-cards/gtx-titan-z/ and http://www.nvidia.com/object/tesla-servers.html. The first number is the Titan Z, the second one the Tesla K40

  • Name: Titan Z ↔ Tesla K40
  • Number and Type of GPU: ↔ 1 Kepler GK110B
  • Peak double precision floating point performance: 2.66 Tflps (see link) ↔ 1.43 Tflops
  • Peak single precision floating point performance: ? ↔ 4.29 Tflops
  • Memory bandwidth: 672 GB/sec ↔ 288 GB/sec (ECC off)
  • Memory size (GDDR5): 12 GB ↔ 12 GB
  • CUDA cores: 5760 ↔ 2880

Update

1
  • 2
    One big, very fundamental difference might be the lack of video out on the "Scientific" Teslas and that exists on the Titan Z
    – Journeyman Geek
    Aug 30, 2014 at 23:51

1 Answer 1

2

Scientific GPUs have

  • ECC memory: This seems to make a big difference. According to my adivser, one can expect an error every 26 hours or so.
  • Performance information is available over nvidia-smi.

Gaming GPUs are able to use a monitor. Scientifc GPUs don't necesarily have a video out.

2
  • And for recent graphic card maybe all simply the ability to perform more generic computations superuser.com/q/1598712. Or at least effeciently. Oct 31, 2020 at 14:43
  • And maybe they don t support api like Direct3D or OpenGl but only Cuda and OpenCl. Oct 31, 2020 at 15:02

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .