5 ms·

Above, we’re looking at parallel performance of the NASA USM3D CFD solver as it computes flow over a classic NACA 0012 airfoil section at low speed conditions.

• bee_rider 5y ago
  Just for anyone as out of touch with the MacOS ecosystem as me: Accelerate includes a BLAS implementation, so at least seems plausible (depending on how this library was compiled) that their special instructions might have been used.
• torginus 5y ago
  But that makes me think, what prevents people from running these calculations on the GPU? Even the memory is shared - the few 100 gflops they get out of the M1 ultra is pocket change in GPU terms.
  • the_svd_doctor 5y ago
    Well that particular NASA code is from the 90’s and in Fortran. Maybe with MPI. That’s probably part of the reason why.
    • willis936 5y ago
      Don't all high performance math libraries have the option of LAPACK interfaces?
  • CamperBob2 5y ago
    If a CPU doesn't run the OS and software I need to run, it might as well be a GPU or a Cray 9 or something dug out of the wreckage at Area 51. All of which are good descriptions of a CPU available only on the increasingly-proprietary Mac platform. So this entire thread is kind of pointless with regard to many real-world use cases.
    • madengr 5y ago
      Came here to say the same. I don’t know of any computational heavy engineering software available for the Mac. Unfortunately I’m stuck on Windows for all my software.
      • _ph_ 5y ago
        Well, the Apple Silicon Macs just arrived. If they are a compelling architecture for engineering software, then sooner or later a company will support them. Also, most Unix-based software should be rather easy portable, especially programs for computations which do not depend on a GUI. There is a lot of engineering software which is only available on Unix and these days mostly on Linux, as it comes from times when Windows and PCs where not even remotely a consideration as a platform.
    • alimov 5y ago
      Saying nothing in so many words. > “If a CPU doesn't run the OS and software I need to run, it might as well be a GPU or …” > So this entire thread is kind of pointless with regard to many real-world use cases. Does this kind of reasoning apply to everything that doesn’t suit your unique personal needs? Personally I think this is an interesting thread, but I don’t use the qualifier of “is it useful to me” to determine whether something is interesting.
    • lupire 5y ago
      Alternatively, your OS of choice may as well be a Cray 9 OS if it doesn't run on modern high-end hardware. Philosophical purity rarely provides top performance. And that's OK, slow and steady works for me, because I'm not doing max performance computing.
    • 5y ago
  • my123 5y ago
    Apple GPUs don't support FP64 at all - and Apple only provides Metal as a (supported and not deprecated) API. NVIDIA GPUs have much more ubiquitous programming models, from CUDA C++ to OpenMP (both C++ and Fortran), allowing them to be much more useful. AMD ROCm (a CUDA clone, but not managed too well...) isn't in a great state today, but it definitely provides an OpenMP compiler, for C++ only though, Fortran is not covered. Intel GPUs have a much better SW story on all fronts for GPGPU than AMD's, and OpenMP is supported for both C++ and Fortran there. tldr: if you want to run this code on GPUs, the two vendors to look for are NVIDIA and Intel. The others just don't have the software stack to do so. (without more heavyweight porting) Of course, optimising for GPUs also adds work. That said, an OpenMP version is a very good starting point.
    • paulmd 5y ago
      You can always emulate FP64 - it is not performant to do so, but since CFD is essentially memory-bandwidth-bottlenecked rather than limited by shader performance, this probably is not a big concern. You've got shaders sitting around waiting anyway. The M1 architecture only allows the CPU to access a fraction of the memory bus, it is around 25% of the total bandwidth or so iirc. The rest has to go to the GPU or NPU or one of the other units, so even if your FP64 code was super shitty emulated with no hardware support, on paper you should be able to quadruple performance by utilizing it. If you can "pre-digest" it at all, and then push it back from the GPU to the CPU, that might be another viable approach. Of course none of this will exist out of the box in some 90s FORTRAN code!
      • my123 5y ago
        > You can always emulate FP64 That's what's done by Intel for their newer consumer GPUs AFAIK, zero FP64 units present on die, but an emulation mode can be enabled to expose that support. (which is then done by software on the GPU) > it is around 25% 50%, ~ 200GB/sec on an M1 Max out of ~ 400GB/sec
  • whatever1 5y ago
    Because GPUs calculations are the same type of calculations. If you have a problem where you need to add plus 1 in all of the elements of a matrix, then GPUs are great. Games are a great example because you can take one pixel and calculate its values over time without much worry of what happens around it. Similar story for neural nets. For CFD it is a different problem. In each time-step you essentially solving a linear system Ax=b, so even if you are looking at element x(100,100,timeperiod2) you need to also know the value of the element x(1,1,timeperiod1). There are some algorithms that decouple the problem by introducing residuals for each element and then trying to iteratively reduce them, but as you can see, it is not a linear increase in speed as someone would had expected by looking at the GPU specs. TLDR: Yes you can do CFD with GPUs, but dont expect miracles.
    • dekhn 5y ago
      using GPUs for CFD is pretty standard these days and has been for a few years. GPUs are fantastic for solving linear systems. Nobody uses GPUs for "linear speedups by looking at GPU specs". That's not how accelerators are measured.
      • whatever1 5y ago
        "using GPUs for CFD is pretty standard these days and has been for a few years. " It is not standard in OpenFoam (only though rapidCFD) or Ansys (only this year they made an announcement for full acceleration). "GPUs are fantastic for solving linear systems." *Some* linear systems. They are horrible for sparse algebra.
• stephencanon 5y ago
  If it were taking advantage of Accelerate, the performance would be much higher, but also the scaling would be quite different. Look at the scaling in the tweet you linked--it's anything but linear in the number of cores used.
• snapcore 5y ago
  The scaling curve is different from the other cpus which implies the performance bottleneck is not in matrix multiplication (which would be nearly linear on the older cpus because it is embarrassingly parallel).
  • paulmd 5y ago
    CFD is highly memory-bandwidth-bottlenecked, it is in fact pretty much the prototypical memory-bandwidth-bottlenecked task. The performance scaling you see between systems pretty much corresponds to the memory bandwidth in those configurations. Note that on the M1, the CPU can only access a fraction (about 25% iirc) of the total memory bandwidth, you have to use the GPU to really get the full performance of the M1 here.
    • sliken 5y ago
      I saw this quantified (I think at anandtech), something like 220GB/sec out of 400GB/sec on the M1 max. So something north of 50%. Also keep in mind that normal x86-64's, even without an IGP only get about 60-65% of peak, even with nothing else sharing the memory bus. I often see this quantified with McCalpin's stream benchmark. So the M1 Ultra likely has a pretty impressive memory bandwidth of around 440GB/sec, which isn't a large fraction of 800GB/sec, but it still more than any other desktop or server chip I know of. The AMD Epcy maxes out at 8 channels of DDR-3200, which is in the neighborhood of 208GB/sec peak, with an observed bandwidth of 110-120GB/sec.
      • AceJohnny2 5y ago
        > of around 440GB/sec, which isn't a large fraction of 800GB/sec Where's the 800GB/s from?
        • sliken 5y ago
          Peak = never observed, but calculated from clock speed * bus width. Much like the speed of light, you'll never see it. That number for the M1 Ultra (from the OP's post) = 800GB/sec. McCalpin's stream benchmark is often cited as a practical/useful number for usable bandwidth using a straight forward implementation in C or Fortran without trying to play games, much like the vast majority of codes out there. Also note that the x86-64's in the world use a strict memory model that results in a lower fraction of observed bandwidth vs peak. Arm has a looser memory model which achieves a higher fraction of peak.
• lupire 5y ago
  That's good info but omg why a tweet of a crop of screenshot of a document.
• staticfloat 5y ago
  This linear scaling per-core doesn't match my experience with using the AMX co-processor. In my experience, on the M1 and the M1 Pro, there are a limited number of AMX co-processors that is independent of the number of cores within the M1 chip. I wrote an SO answer exploring some of the performance implications of this [0], and since I wrote that another more knowledgable poster has added more information [1]. One of the key takeaways is that appears to be one AMX coprocessor per "complex", leading us to hypothesize that the M1 Pro contains 2 AMX co-processors. This is supported by taking the code found in the gist [2] linked from my SO answer and running it on my M1 Pro. Compiling it, we get `dgesv_accelerate` which uses Accelerate to solve a medium-size linear algebra problem, that typically takes ~8s to finish on my M1 Pro. While running, `htop` reports that the process is pegging two cores (analogous to the result in my original SO answer on the M1 pegging one core; this supports the idea that the M1 Pro contains two AMX co-processors). If we run two `dgesv_accelerate` processes in parallel, we see that they take ~15 seconds to finish. So there is some small speedup, but it's very small. And if we run four processes in parallel, we see that they take ~32 seconds to finish. All in all, the kind of linear scaling shown in the article doesn't map well to the limited number of AMX co-processors available in Apple hardware, as we would expect the M1 Max to contain maybe 8 co-processors at most. This means we should see parallelism step up in 8 steps, rather than 20 steps as was shown in the graph. Everything I just said is true assuming that a single processor, running well-optimized code can completely saturate an AMX co-processor. That is consistent with the tests that I've run, and I'm assuming that the CFD solver he's running is well-written and making good use of the hardware (it does seem to be doing so from the shape of his graphs!). If this were not the case, one could argue that increasing the number of threads could allow multiple threads to more effectively share the underlying AMX coprocessor and we could get the kind of scaling seen in the article. However, in my experiments, I have found that Accelerate very nicely saturates the AMX resources and there is none left over for future sharing (as shown in the dgesv examples). Finally, as a last note on performance, we have found that using OpenBLAS to run numerical workloads directly on the Performance cores (and not using the AMX instructions at all) is competitive on larger linear algebra workloads. So it's not too crazy to assume that these results are independent of the AMX's abilities! [0] https://stackoverflow.com/a/67590869/230778 https://stackoverflow.com/a/67590869/230778 [1] https://stackoverflow.com/a/69459361/230778 https://stackoverflow.com/a/69459361/230778 [2] https://gist.github.com/staticfloat/2ca67593a92f77b1568c03ea47933e36 https://gist.github.com/staticfloat/2ca67593a92f77b1568c03ea...
• JohnTHaller 5y ago
  The 5900X can do 24 threads, so why artificially limit it to 16 threads? Unless that's because the M1 Ultra has 16 performance cores. Why include the 3 year old 3700X? Why not include the 32 thread 5950X which can be had in a complete system for 1/3 the price of the Mac Studio M1 Ultra?
  • threeseed 5y ago
    They clearly just picked 16 threads because it is enough to demonstrate the point. M1 only has 4 performance cores and they still ran it with 16 threads. So it's not like they are trying to skew the benchmark by not showing the potential of the 5900x. Also the 5950x is 2 years old and still is an active product.
  • hedgehog 5y ago
    The 5900X is a fine reference as a common chip. The peak performance should be at 12 threads, it has 12 cores, and that's what the table appears to show.
    • sharken 5y ago
      It would make more sense to compare systems with similar price points. In this case that would mean ThreadRipper systems and not Ryzen.
      • bigiain 5y ago
        For the guy who, you know, actually did the comparison here... Of _course_ it makes most sense to compare the new system he has on his desk with the previous systems he's already done relevant performance tests as comparison points. If you want a comparison between a Threadripper system and an M1 Ultra system, I'd be fascinated to read your blog post with your results.
      • microtonal 5y ago
        Sure, if I had a ThreadRipper, I’d be happy to add the results. I don’t have an M1 Ultra either, but someone was kind enough to do the benchmark for me.