[HN Gopher] Intel Extension for TensorFlow
___________________________________________________________________
Intel Extension for TensorFlow
Author : hochmartinez
Score : 75 points
Date : 2022-10-28 18:29 UTC (4 hours ago)
(HTM) web link (github.com)
(TXT) w3m dump (github.com)
| echelon wrote:
| I love the hardware Nvidia makes, but for the good of our field
| and for progress, we need our ML stacks to work with all other
| vendors. CUDA is the wrong platform to build our future on.
| _aavaa_ wrote:
| Disappointed by the lack of benchmark comparison in the README
| minimaxir wrote:
| For context, in other Python ML packages like scikit-learn
| running the Intel extension on a compatable CPU can result in a
| _massive_ performance increase.
|
| Although since TensorFlow models should be trained on a GPU
| unlike sklearn, that's less useful, and there are better tools
| for CPU inference. (e.g. SavedModels or ONNX)
| benreesman wrote:
| I don't know much about Intel's toolchain: people say good
| stuff about TBB and things like that.
|
| Historically doing inference on Intel gear was mostly about
| whether or not to target AVX2 or AVX512 when building Eigen or
| whatever. A few years ago the net win was AVX2 because the de-
| clock and re-clock just killed you.
|
| What's the game these days? Long term I doubt inference will be
| done on x86, but I think a lot of people still do it.
| make3 wrote:
| yes, but Intel now has mainstream GPUs that are reasonably (~
| RTX 3060) performant, which is what I assume this is for
| Roark66 wrote:
| Are you talking about Intel Arc? I'm yet to see any ML
| relevant benchmarks on Intel Arc. If you are aware of any
| please let me know.
| throwaway1851 wrote:
| Same here. At 16GB of VRAM and only $349, it could fill a
| really nice slot for DL.
| dweekly wrote:
| For the curious, Apple has an analogous Tensorflow Metal plugin
| to allow for Apple Silicon (and AMD GPU) acceleration using the
| same plugin architecture.
|
| https://developer.apple.com/metal/tensorflow-plugin/
| muxamilian wrote:
| Which is basically unusably buggy.
|
| For example, tf.sort only sorts up to 16 values and overwrites
| the rest with -0. Apparently not fixed for over one year:
| https://developer.apple.com/forums/thread/689299
|
| Also, tf.random always returns the same random numbers:
| https://developer.apple.com/forums/thread/69705
|
| Although I guess these bugs are not the fault of Tensorflow's
| plugin architecture but rather Apple's implementation.
| Roark66 wrote:
| Intel did a great thing for people interested in ML and numeric
| research by making their MKL library and compiler free and cross
| platform compatible. Even today on my AMD zen3 Ryzen machine
| intel's MKL linked numpy and pytorch are in some operations 10*
| (yes that is really ten times) faster in comparison with the next
| best alternative (openBlas etc). I was shocked to discover how
| much of a difference MKL makes for cpu workloads. This is mostly
| because it makes use of AVX2 cpu extensions which make certain
| matrix operations a lot faster.
| galangalalgol wrote:
| Compiling eigen uses avx512 just fine. I don't think it is
| quite the discriminator it once was. However, I amp happy they
| made it open and it would be great of oneAPI became a real
| alternative to cuda for more tasks. I tried getting it to work
| with flux.jl a while back and ran into some difficulty.
| ysleepy wrote:
| They crippled the performance on non-Intel CPUs on purpose
| until recently.
|
| Intel's anti competitive behaviour follows them throughout
| their history.
| pantalaimon wrote:
| They crippled their own consumer CPUs by retroactively
| disabling AVX512
| Roark66 wrote:
| Yes, this is true, but even back then (during the time they
| crippled it) it was possible to pretend to run on Intel
| hardware. I haven't done this, but I read it was possible.
| ysleepy wrote:
| I just felt your first sentence, without a qualifier, gave
| too much credit looking at the context.
| chrchang523 wrote:
| That was my previous experience, but have you tried linking to
| AMD AOCL recently? I would not expect the performance gap
| between Intel MKL and AMD AOCL to still be as large as you
| describe on a Zen 3.
| Scene_Cast2 wrote:
| Would you know if AOCL is supported by numpy? And are there
| any benchmarks out there? (Especially interested in Zen4)
___________________________________________________________________
(page generated 2022-10-28 23:00 UTC)