Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Intel Announces Knights Mill: A Xeon Phi for Deep Learning

anandtech.com

Post reply on HN
Print view View original

41–50 of 57 posts

Re: Intel Announces Knights Mill: A Xeon Phi for Deep Learning

#41

Post by rbanffy »

stonogo wrote: post #40
rbanffy wrote: post #39

Earlier quoted context omitted.

I think it's fair to assume any significant HPC solution will have to be recompiled to target its runtime environment. Squeezing out every GFLOPS of a part is normal when you get an expensive specialized computer. In that context, x86 compatibility is less of an issue than the quality of the compilers. As long as you are easier to program than a GPU, you are good.

I wish this were true, but it's not. Multi-million-dollar contracts have been scuttled because a given customer runs codes that they receive as binaries from a vendor, closed-source, and Company A bid Processor X, but the codes only run on Processor Y...

That's interesting. I'd assume a market so performance driven would compete such vendors out of existence.

Re: Intel Announces Knights Mill: A Xeon Phi for Deep Learning

#42

Post by nl »

visarga wrote: post #36
frozenport wrote: post #28

Intel has thousands of employees and a compiler. Get 200 of them in a room and implement CUDA.

They don't even have to implement the latest flavor of "Inception module", they only need to implement matrix vector operations and some math primitives like exponential, log, tangent and such. Why is it so hard to port to Intel? I would have liked to make use of my Macbook's Intel Iris GPU for deep learning, but it's not supported by anything.

OpenCL supports Iris.

Re: Intel Announces Knights Mill: A Xeon Phi for Deep Learning

#43

Post by jlouis »

jlebar wrote: post #38
cs702 wrote: post #2

We badly need an alternative to Nvidia/CUDA for deep learning... but realistically, if Intel wants to make headway in the deep learning market, it must offer hardware that can not only compete on performance with Nvidia, but also work out-of-the-box (that is, without requiring lots of one-off tinkering and tweaking) with popular deep/machine learning frameworks like TensorFlow, Caffe, Torch, and Theano. There is a lo…

Tensorflow has a mode that lets it run on CPUs. I'm sure other frameworks are the same. Isn't the whole point of Xeon Phi that it looks basically like an x86 CPU with a ton of cores? If so there is almost nothing to port, just the kernel launching. Granted, you do have to bother to make a fast x86 / AVX-512 port of your code. But because the shape of GPUs is so different than CPUs -- GPUs have a more complicated memo…

To elaborate: Tensorflow is clever because it takes your flow graph and cuts it into pieces which it then executes on computational units. Any computational unit will do, really as long as you have a backend for it. This choice makes adaptation to new system much easier.

I believe Caffe and Theano uses the same model, but I didn't study it. There are also some similarity in the model to what OCaml does with the incremental library, though it is not for machine learning.

Re: Intel Announces Knights Mill: A Xeon Phi for Deep Learning

#44

Post by pjmlp »

visarga wrote: post #35
joe_the_user wrote: post #22

Earlier quoted context omitted.

I'd love an alternative to CUDA. The problem is that as far as I can see, OpenCL is in no way that. Basically, OpenCL gives me the impression that the oceans of boiler plate required both make development hard and effectively locks you into a specific vendor also since the boiler-plate is going to be setting things up for one's specific vendor.

Can anyone tell me what exactly is missing from OpenCL to be able to run the primitives of deep learning frameworks? Like, does it not have some kind of operation that is essential for matrix manipulation?

Up to OpenCL 2.1 was support for C++ and Fortran, forcing everyone to either use C or generate C code from their compilers.

Re: Intel Announces Knights Mill: A Xeon Phi for Deep Learning

#45

Post by lmm »

xadhominemx wrote: post #18
vonnik wrote: post #17

Earlier quoted context omitted.

Why use it, if it locks you in?

If you're a hobbyist, why would you care about lock in?

I care about being able to carry on doing my hobby stuff 5 or 10 years from now. With NVidia I can at least be confident that as long as my graphics card keeps working (which feels like something under my control, unlike Google shutting down their products) I can keep running my code on it.

Re: Intel Announces Knights Mill: A Xeon Phi for Deep Learning

#46

Post by dagss »

marmaduke wrote: post #31
joe_the_user wrote: post #22

Earlier quoted context omitted.

I'd love an alternative to CUDA. The problem is that as far as I can see, OpenCL is in no way that. Basically, OpenCL gives me the impression that the oceans of boiler plate required both make development hard and effectively locks you into a specific vendor also since the boiler-plate is going to be setting things up for one's specific vendor.

Nope. I have some substantial simulation code written against OpenCL that runs on Intel OpenCL and NVIDIA without modifications, and rather performant on both of them. The only part of the code specific to vendor is the platform selection, which is one line of code. OpenCL falls down in terms of standard libraries such as cu{dnn,sparse,blas} but if you're writing everything from scratch it's fine.

Interesting...so do you use the vector subset designed for CPUs or the wavefront subset designed for GPUs?

Re: Intel Announces Knights Mill: A Xeon Phi for Deep Learning

#47

Post by scottlegrand »

cs702 wrote: post #2

We badly need an alternative to Nvidia/CUDA for deep learning... but realistically, if Intel wants to make headway in the deep learning market, it must offer hardware that can not only compete on performance with Nvidia, but also work out-of-the-box (that is, without requiring lots of one-off tinkering and tweaking) with popular deep/machine learning frameworks like TensorFlow, Caffe, Torch, and Theano. There is a lo…

Except that up to now at least, CUDA IMO remains the best abstraction for programming multi-core: subsuming away multiple threads, SIMD width, and multiple cores into the language definition. AMBER (http://www.ambermd.org) literally "recompiled and ran" with each succeeding GPU generation since GTX 280 in 2009. 3-5 days of subsequent refactoring then unlocked 80% of the attainable performance gains of each of the subsequent GPUs. DSSTNE (https://github.com/amznlabs/amazon-dsstne) just ran as well, but it's only targeting Kepler and up because the code relies heavily on the __shfl instruction.

So I honestly don't get the Google clang CUDA compiler right now. It's really really cool work, but I don't get why they didn't just lobby NVDA heavily to improve nvcc. With the number of GPUs they buy, I suspect they could have anything they want from the CUDA software teams.

However, if it could compile CUDA for other architectures, sign me up, you'd be my heroes.

For I'd love to see CUDA on Xeon Phi and on AMD GPUs (I know, they're trying). And if Intel poured the same amount of passion and budgeting into building that as they are pouring into fake^H^H^H^Hdeceptive benchmark data and magical powerpoint processors we won't see for at least a year or two (and which IMO will probably disappoint just like the first two), they'd be quite the competitor to NVIDIA, no?

That said, the Intel marketing machine seems to have succeeded in punching NVDA stock in the nose the past few days and in grabbing coverage in Forbes (http://www.forbes.com/sites/aarontilley/2016/08/17/intel-tak...) so maybe they know a thing or two I don't.

Re: Intel Announces Knights Mill: A Xeon Phi for Deep Learning

#48

Post by joe_the_user »

marmaduke wrote: post #31
joe_the_user wrote: post #22

Earlier quoted context omitted.

I'd love an alternative to CUDA. The problem is that as far as I can see, OpenCL is in no way that. Basically, OpenCL gives me the impression that the oceans of boiler plate required both make development hard and effectively locks you into a specific vendor also since the boiler-plate is going to be setting things up for one's specific vendor.

Nope. I have some substantial simulation code written against OpenCL that runs on Intel OpenCL and NVIDIA without modifications, and rather performant on both of them. The only part of the code specific to vendor is the platform selection, which is one line of code. OpenCL falls down in terms of standard libraries such as cu{dnn,sparse,blas} but if you're writing everything from scratch it's fine.

I'm an independent developer in the process of choosing a GPGPU library.

I can see simple, comprehensible 20-50 sample code for cuda that does most simple tasks. With OpenCL, I get references to version, boiler-plate, mode with nothing that boils down to simple code.

If you have a simple sample, you should post it here or blog about it.

Re: Intel Announces Knights Mill: A Xeon Phi for Deep Learning

#49

Post by AIMunchkin »

jcoffland wrote: post #27
cs702 wrote: post #2

We badly need an alternative to Nvidia/CUDA for deep learning... but realistically, if Intel wants to make headway in the deep learning market, it must offer hardware that can not only compete on performance with Nvidia, but also work out-of-the-box (that is, without requiring lots of one-off tinkering and tweaking) with popular deep/machine learning frameworks like TensorFlow, Caffe, Torch, and Theano. There is a lo…

Since Intel's product is just a bunch of CPUs it should work with OpenCL out of the box.

It does... Badly... And you used to have to pay for it...

Re: Intel Announces Knights Mill: A Xeon Phi for Deep Learning

#50

Post by duaneb »

joe_the_user wrote: post #48
marmaduke wrote: post #31

Earlier quoted context omitted.

Nope. I have some substantial simulation code written against OpenCL that runs on Intel OpenCL and NVIDIA without modifications, and rather performant on both of them. The only part of the code specific to vendor is the platform selection, which is one line of code. OpenCL falls down in terms of standard libraries such as cu{dnn,sparse,blas} but if you're writing everything from scratch it's fine.

I'm an independent developer in the process of choosing a GPGPU library. I can see simple, comprehensible 20-50 sample code for cuda that does most simple tasks. With OpenCL, I get references to version, boiler-plate, mode with nothing that boils down to simple code. If you have a simple sample, you should post it here or blog about it.

Boilerplate is one thing—this is what transpilers are for, among other solutions—but can OpenCL match CUDA on a performance level?
Post reply on HN