10
$\begingroup$

Im trying to understand the difference between these two graphics cards for academic computing, specifically for the DGEMM component.

If we look at the raw statistics, both have the same GK110 chip, have comparable statistics in virtually every category, and, I believe, have the same core architecture. Before any discounts the K20X is roughly 4x the cost of a Titan. From a efficiency perspective it seems to make much sense to use Titan's over the K20X's.

I am having a difficult time understanding the difference here, can anyone illuminate the situation?

As a note I am looking at purchasing these cards for a rack server and run at full tilt essentially until they die; however, I do not view the efficiency of using multiple GPU's for a single job to be particularly paramount.

$\endgroup$
12
$\begingroup$

There are some differences, however they aren't necessarily in hardware or specs. Note that this is all information I have gained from forums or news releases, so take it all with a grain of salt.

The first is the "scalability and reliability" (source). The K20 was designed to sit in a cluster system and run at full tilt 24/7. The Titan is more designed for gaming, so it will run at this duty cycle, but it may suffer long term lifetime issues if used this way.

The drivers are also different, however I am not sure of the major differences. The difference in focus of the cards' design likely leads to relatively small performance gains for the Tesla cards on this front.

"Some Tesla-exclusive features include:

  • NVIDIA GPUDirect RDMA for InfiniBand performance
  • Hyper-Q for MPI (Hyper-Q for CUDA Streams is supported on GeForce GTX TITAN)
  • ECC protection for all internal and external registers and memories
  • Supported tools for GPU and cluster management, such as Bright Computing, Ganglia." (source)

This points to the fact that the main differnce is their scalability. If you are looking to run on a desktop in your office, it would be hard to argue against a Titan over the K20 for the price difference. If you need the extra performance of multiple K20's, find yourself a HPC center and buy time with their servers.

Edit:

After looking a bit more into ECC, I am updating this answer to point out the implications of having it on the K20 and not on the Titan. The following information is a paraphrase of info found here.

ECC is error checking on the DRAM and registers for the GPU. Soft errors are when a bit is incorrectly transfered/stored. The faster and closer together the circuits, the higher the probility of a soft error. If you are solving a set of coupled ODE's or solving a linear system, a single number being off by one bit could significantly change the results in a non-reproducable way. Most standard RAM and caches in the CPU are error checked for these errors using ECC.

GPU's on the other hand, do not, in general, have ECC even though their memory bus are much faster than those on the CPU. This is because if a pixel on the screen is off by a bit for one frame, the quality of the program is not diminished. These errors also don't propogate. Therefore a lot of chip real estate (and cost) can be saved by skipping this feature. This extra complexity likely causes a large portion of the extra cost of the Tesla line.

$\endgroup$
  • 3
    $\begingroup$ Great answer +1! It is hard to believe those features are that costly. I think the line "Develop with GeForce, Deploy with Tesla" from the linked Nvidia site sums up the important issues. Looks like the best solution for now is to buy several GeForces and run them hard until they give up the blue smoke so to speak. $\endgroup$ – Ophion Aug 15 '13 at 18:58
  • 3
    $\begingroup$ "This however did not stop them from being used at Oakridge." OLCF's Cray XK7 named "Titan" uses Tesla K20 GPUs, not the GTX Titan. NVidia says that the GTX Titan has "the technology of" OLCF Titan, which is the same vocabulary used when saying that an economy car has "the technology of" a Formula-1 car. (GTX Titan performance is pretty good, but it doesn't have ECC and isn't used in any major installations that I'm aware of.) $\endgroup$ – Jed Brown Aug 15 '13 at 20:28
  • 1
    $\begingroup$ My mistake, I misinterpreted the article. I will update the answer to not be misleading. $\endgroup$ – Godric Seer Aug 15 '13 at 20:35
2
$\begingroup$

In my opinion the difference seems to be mostly market segmentation. If you are a scientist then NVidia wants you to be afraid that your paper will be rejected because you are using a GPGPU without as much error correcting RAM as would be available with K20X. Similarly if you are a corporation then you might want to pay 4x if it means you are less likely for being sued on suspicion that your calculations are not as error-corrected as possible. Individual gamers or hobbyist GPGPU'ers are sold Titan because they have less money and they are harder to persuade in these ways.

$\endgroup$
  • 2
    $\begingroup$ I have only run calculations with ECC, do you happen to have a good article demonstrating the failures of non ECC systems and logical breakpoints where its beneficial to have? $\endgroup$ – Ophion Aug 15 '13 at 18:54
  • 2
    $\begingroup$ @Ophion An Investigation of the Effects of Error- Correcting Code on GPU-accelerated Molecular Dynamics Simulations ---> this may be of some interest to you. $\endgroup$ – BenC Mar 26 '14 at 16:22
  • $\begingroup$ For those who want the executive summary of the very excellent link by BenC: soft errors that ECC would fix are exceedingly rare, and the paper goes as far as recommending turning off ECC on Tesla's for increased speed. Caveat: this was not actually tested with consumer GPUs. $\endgroup$ – semi-extrinsic Jan 27 '15 at 10:43
0
$\begingroup$

It really depends on the application you run. GPUGRID.net runs on machines that do not have ECC on and everything is fine. The results are as good as they are on any other platform. Acellera also sells hardware with GeForce cards and in only very few cases the GPUs have failed. GeForce is all you need.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.