-1

I am using NVIDIA Quadro K2000 GPU. I ran deviceQuery.exe, the results of which are below. It says that I have only 2 SM units. I am not sure if I really have only 2 SMs or some of my SMs are disabled as mentioned at the third comment at this SO question.

I also saw that the number of SP are 192 per SM. May be there is some way to enable more SM and then the number of SP per SM will decreased.

image

Update: The reason I am asking is the fact that I want to make my GPU processing efficient. I have an image of 1080 x 1920 which I have divided into three segments. I am transferring these segments H2D, processing and transferring D2H asynchronously. Therefore, I want to choose such a dimension of blocks and number of threads per block that can efficiently utilize the hardware configuration of my GPU. Furthermore, I am still confused that a GPU having more than 2 SMs (let's say 8 SMs) but less number of SPs per SM (384/8 = 48 SP per SM) would perform same to a GPU having 2 SMs and 192 SPs per SM ? I mean, is it the total number of available SPs which matters?

  • It is entirely possible that the extra units are disabled by lasering certain important links on the physical chip die and cannot be re-enabled in any way via software or hardware. This is quite likely, to prevent the kind of thing you are wanting. While some of those SMs might actually be functional aside from being disabled it is also possible that they are disabled because they malfunctioned while under test. It's a lot simpler and easier to buy what you actually need rather than the cheapest and hoping that you can re-enable disabled parts. – Mokubai Jul 20 '15 at 17:37
  • @Mokubai: NVIDIA Quadro K2000 has capability 3.0. It's nearly a new generation GPU. Do you really think that is should have only 2 SMs. Whereever I read about GPUs, they say that GPUs usually have 2-30 SMs. And, I can't believe that this latest GPU has only 2 SMs. – skm Jul 20 '15 at 18:26
  • 1
    Given that the K2000 is at the "bargain bucket" side of the Quadro line nothing would surprise me. See en.wikipedia.org/wiki/… for confirmation that you have 2 functional SMX units. You do, in all likelihood, have a chip that has many more SMX units, but they are likely to be physically disabled, either due to them being faulty or by having the tracks laser cut. – Mokubai Jul 20 '15 at 18:54
  • Your "latest" GK107 is, in fact, over three years old, is a bottom line part, and has exactly 2 SMX, each with 192 cores: tomshardware.com/reviews/geforce-gt-640-review,3214.html (your K2000 is just a Quadro version of this card). And there is no magic software commands that will tell the transistors on the silicon to reorganize themselves. Are all these questions just buyers regret or do you have an underlying reason to be asking this? – talonmies Jul 21 '15 at 6:47
  • @talonmies: I did not write the reason of asking this question because somebody at SO told not to ask hardware questions there. So, here I just asked about the hardware problem. Anyways, I have updated my question to clarify the reason of asking this question. – skm Jul 21 '15 at 7:12
1

In order:

  1. I am not sure if I really have only 2 SMs or some of my SMs are disabled

    Your GPU is GK107. It has two SMX, each with 192 SP. There are no disabled processors.

  2. May be [sic] there is some way to enable more SM and then the number of SP per SM will decreased.

    There is no magic way of changing the silicon in software.

  3. I am still confused that a GPU having more than 2 SMs (let's say 8 SMs) but less number of SPs per SM (384/8 = 48 SP per SM) would perform same to a GPU having 2 SMs and 192 SPs per SM ?

    That is a huge question and one I'm not going to try and answer here. NVIDIA has released a lot of very detailed whitepapers on the architecture and performance of the various generations of their flagship compute cards as each has been released. If you are so inclined, I suggest reading them for yourself. Having said that, the rule of thumb is that within the same architecture generation the more SM or SMX a device has, the better it will perform at arithmetic bound operations. Comparing different generations of GPUs is futile. For example, the very first Telsa compute card NVIDIA released has 16 SM (the C860). The most recent has 15 SMX (the K40). There is about 10 times difference in their peak single precision performance. Note also that many real world applications are memory bandwidth bound, and that can radically change the performance difference between different cards. While core count and peak FLOP have more or less followed Moore's Law, memory bandwdith has not.

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.