Nvidia's CUDA: The End of the CPU?


However we did decide to measure the processing time to see if there was any advantage to using CUDA even with our crude implementation, or on the other hand if was going to take long, exhaustive practice to get any real control over the use of the GPU. The test machine was our development box – a laptop computer with a Core 2 Duo T5450 and a GeForce 8600M GT, operating under Vista. It’s far from being a supercomputer, but the results are interesting since our test is not all that favorable to the GPU. It’s fine for Nvidia to show us huge accelerations on systems equipped with monster GPUs and enormous bandwidth, but in practice many of the 70 million CUDA GPUs existing on current PCs are much less powerful, and so our test is quite germane.

The results we got are as follows for processing a 2048x2048 image:

  • CPU 1 thread: 1419 ms
  • CPU 2 threads: 749 ms
  • CPU 4 threads: 593 ms
  • GPU (8600M GT) blocks of 256 pixels: 109 ms
  • GPU (8600M GT) blocks of 128 pixels: 94 ms
  • GPU (8800 GTX) blocks of 128 pixels / 256 pixels: 31 ms

Several observations can be made about these results. First of all you’ll notice that despite our crack about programmers’ laziness, we did modify the initial CPU implementation by threading it. As we said, the code is ideal for this situation – all you do is break down the initial image into as many zones as there are threads. Note that we got an almost linear acceleration going from one to two threads on our dual-core CPU, which shows the strongly parallel nature of our test program. Fairly unexpectedly, the four-thread version proved faster, whereas we were expecting to see no difference at all on our processor, or even – and more logically – a slight loss of efficiency due to the additional cost generated by the creation of the additional threads. What explains that result? It’s hard to say, but it may be that the Windows thread scheduler has something to do with it; but in any case the result was reproducible. With a texture with smaller dimensions (512x512), the gain achieved by threading was a lot less marked (approximately 35% as opposed to 100%) and the behavior of the four-thread version was more logical, showing no gain over the two-thread version. The GPU was still faster, but less markedly so (the 8600M GT was three times faster than the two-thread version).

nvidia CUDA

The second notable observation is that even the slowest GPU implementation was nearly six times faster than the best-performing CPU version. For a first program and a trivial version of the algorithm, that’s very encouraging. Notice also that we got significantly better results using smaller blocks, whereas intuitively you might think that the reverse would be true. The explanation is simple – our program uses 14 registers per thread, and with 256-thread blocks it would need 3,584 registers per block, and to saturate a multiprocessor it takes 768 threads, as we saw. In our case, that’s three blocks or 10,572 registers. But a multiprocessor has only 8,192 registers, so it can only keep two blocks active. Conversely, with blocks of 128 pixels, we need 1,792 registers per block; 8,192 divided by 1,792 and rounded to the nearest integer works out to four blocks being processed. In practice, the number of threads are the same (512 per multiprocessor, whereas theoretically it takes 768 to saturate it), but having more blocks gives the GPU additional flexibility with memory access – when an operation with a long latency is executed, it can launch execution of the instructions on another block while waiting for the results to be available. Four blocks would certainly mask the latency better, especially since our program makes several memory accesses.

Create a new thread in the US Reviews comments forum about this subject
This thread is closed for comments
Comment from the forums
    Your comment
  • Anonymous
    CUDA software enables GPUs to do tasks normally reserved for CPUs. We look at how it works and its real and potential performance advantages.

    Nvidia's CUDA: The End of the CPU? : Read more
  • pulasky
  • Anonymous
    Well if the technology was used just to play games yes, it would be crap tech, spending billions just so we can play quake doesnt make much sense ;)
  • MTLance
    Wow a gaming GFX into a serious work horse LMAO.
  • dariushro
    The Best thing that could happen is for M$ to release an API similar to DirextX for developers. That way both ATI and NVidia can support the API.
  • dmuir
    And no mention of OpenCL? I guess there's not a lot of details about it yet, but I find it surprising that you look to M$ for a unified API (who have no plans to do so that we know of), when Apple has already announced that they'll be releasing one next year. (unless I've totally misunderstood things...)
  • neodude007
    Im not gonna bother reading this article, I just thought the title was funny seeing as how Nvidia claims CUDA in NO way replaces the CPU and that is simply not their goal.
  • LazyGarfield
    I´d like it better if DirectX wouldnt be used.

    Anyways, NV wants to sell cuda, so why would they change to DX ,-)
  • Anonymous
    I think the best way to go for MS is announce to support OpenCL like Apple. That way it will make things a lot easier for the developers and it makes MS look good to support the oen standard.
  • Shadow703793
    Mr RobotoVery interesting. I'm anxiously awaiting the RapiHD video encoder. Everyone knows how long it takes to encode a standard definition video, let alone an HD or multiple HD videos. If a 10x speedup can materialize from the CUDA API, lets just say it's more than welcome.I understand from the launch if the GTX280 and GTX260 that Nvidia has a broader outlook for the use of these GPU's. However I don't buy it fully especially when they cost so much to manufacture and use so much power. The GTX http://en.wikipedia.org/wiki/Gore-Tex 280 has been reported as using upwards of 300w. That doesn't translate to that much money in electrical bills over a span of a year but never the less it's still moving backwards. Also don't expect the GTX series to come down in price anytime soon. The 8800GTX and it's 384 Bit bus is a prime example of how much these devices cost to make. Unless CUDA becomes standardized it's just another niche product fighting against other niche products from ATI and Intel.On the other hand though, I was reading on Anand Tech that Nvidia is sticking 4 of these cards (each with 4GB RAM) in a 1U formfactor using CUDA to create ultra cheap Super Computers. For the scientific community this may be just what they're looking for. Maybe I was misled into believing that these cards were for gaming and anything else would be an added benefit. With the price and power consumption this makes much more sense now.

    Agreed. Also I predict in a few years we will have a Linux distro that will run mostly on a GPU.
  • kelfen
    Well this is a huge step, hope to see it successful.
  • LogicalError
    FYI: Apple has been working with the Khronos group (the people behind OpenGL at the moment) to make an API called OpenCL which should do all the things that Cuda et al can do. Since it's not just Apple that's behind it, but also the Khronos group, it should be cross platform. So who knows.. maybe this is going to be the unifying API for this.. well, until Microsoft comes up with 'DirectC' ofcourse
  • Anonymous
    the last page comments on how MS could come in and create a common API, this common API is already in process, its just that MS isn't part of it ;)
  • Anonymous
    I know that this is not too close to the article, but i hope that it is still not too OFF topic.
    I just have a question, and someone might answer it (the TH is full with smart guys). My problem is that there are too many misconceptions floating around in the net regarding CUDA and overall the whole GPGU businnes.
    I have seen somewhere, that these GPU's are able to do Double Precision floating point calculations, but personally i find this unlikely.
    Others say that you can take directly your parallel code writen in C or Fortran90, and adopt it to CUDA, because the standard stuff can run serial on the CPU and the most computationally expensive part parallel on the GPU. On top of that you can 'adress' or cummunicate with your GPU directly from a Fortran code with sort of system calls (i think this is BS).
    Quiet frankly, i have not found a site on which i can really rely on, where they show an example (source code and explanation) of how something like this could be done.
  • bf2gameplaya
    I wish Intel and NVidia would get over themselves and co-operate and finally give total system performance that big ass boost it needs.

    Intel is wasting time ray-tracing on a CPU and NVidia is wasting frames by folding proteins on their GPU.

    "You're doing it wrong!"
  • Anonymous
    No, the best would be if we got an open API, like OpenGL. I seriously do not want another DirectX locking me to MS >_
  • thr3ddy
    @dariushro: That would quite possibly be the worst thing that could happen to GPGPU. Microsoft equals Windows and GPGPU and super computing is not Windows' strongest point (understatement).

    It would be better for a neutral party composed of GPGPU experts from different IHVs to initiate something like what you propose, more like what the OpenGL ARB creates, a specification.

    IHVs and other companies could then implement this standard on their own hardware, thus decentralizing development from the ISV. If you leave development of this type of technology up to Microsoft (or any other single developer) you'll end up with vendor lock-in, which is a Bad Thing, for all of us.

    Anyway, CUDA is great but not cross-platform compatible (Intel, AMD/ATI, etc.) which makes it impossible to implement in commercial software, unless a CPU-bound alternative is provided, which would defeat the purpose of the architecture.

    On a similar note: think of the choice between the PhysX SDK and Havok Physics. Do you want partial GPU accelerated physics supported by one brand (PhysX, NVIDIA G80+) or do you want to stay CPU-bound but have the same feature set regardless of the hardware (Havok)?
  • magnesious
    If you had the patience to read this entire thing, I'd recommend you look at the CUDA programming guide(link) It's the same information, but less terse.

    Tom's also forgot to point out that development is possible via emulation (emuDebug build setting, I think, with the .vcproj they give you), so anyone can get their hands dirty with the API. You don't get the satisfaction of seeing cool speedups, but it's just as educational, and easier to debug. No screen flickers :)
  • MxM
    I wonder if a PC can be build today without processor at all? It probably requires different BIOS for mobo and some kind of x86 emulator for NVIDIA card, but is it possible in principle without any modifications in hardware?
  • godmodder
    The end of the CPU is nowhere near. To think the GPU could be used for every task is just absurd. The GPU is only good for tasks which can be massively parallellized. Unfortunately, not that many tasks, apart from graphical processing, can be divided into smaller, completely independent parts.