Intel Opens Up Compute Library for Deep Neural Networks

[phoronix]

Phoronix: Intel Opens Up Compute Library for Deep Neural Networks

Intel's 01.org has open-sourced a new Compute Library for Deep Neural Networks (clDNN)...

http://www.phoronix.com/scan.php?pag...n-Source-clDNN

[microcode]

I wonder if it's a good general-purpose OpenCL DNN library, once the intel extensions are factored out (or ported to other GPUs). If it could run on all three major desktop GPU manufacturers' hardware (and be competitive in performance with each platform's vendor's library) it would be pretty valuable.

[boffo]

Every company is like
Code on my system!
Code on my system!
And only for my system!

Is it that hard to optimize Tensorflow/Theano/Pytorch for other systems? So hard that we need other libraries. Why no optimize OpenCL for Tensorflow/Theano/Pytorch?????? How many libraries do we need to define a convolution, a matrix operation, a softmax,....

[Guest]

Every company is like
Code on my system!
Code on my system!
And only for my system!

Is it that hard to optimize Tensorflow/Theano/Pytorch for other systems? So hard that we need other libraries. Why no optimize OpenCL for Tensorflow/Theano/Pytorch?????? How many libraries do we need to define a convolution, a matrix operation, a softmax,....

Fair point, however in the end, it's not consumers who end up using those, it's TF(, ...) developers who integrate these things as compute backends into their own stuff.

Considering terrible fp performance of intel graphics cards I can't even see these things being used, but on the other hand if it were to work with these (https://www-ssl.intel.com/content/ww...ct-detail.html) then there might be an upside here.

[samdraz]

finally intel starting positioning itself against amd's apu,

[boffo]

Fair point, however in the end, it's not consumers who end up using those, it's TF(, ...) developers who integrate these things as compute backends into their own stuff.

Considering terrible fp performance of intel graphics cards I can't even see these things being used, but on the other hand if it were to work with these (https://www-ssl.intel.com/content/ww...ct-detail.html) then there might be an upside here.

At least AMD is working on ROCm, so that caffe will work. Funny how on the GPU side NVidia is monopolizing linear algebra (Matlab & Co)

[gsedej]

Is it possible to use intel gpu (opencl) while using diacrete card (e.g. opensoure radeon)

[juno]

Is it possible to use intel gpu (opencl) while using diacrete card (e.g. opensoure radeon)

Depends.
Normally, you can select that option in your BIOS/UEFI.

[starshipeleven]

Is it possible to use intel gpu (opencl) while using diacrete card (e.g. opensoure radeon)

If you are using Intel GPU as main display, you can set the primary GPU from BIOS/UEFI as iGPU or Integrated.

I'm not so sure you can use a dedicated card as primary GPU while also keeping Intel iGPU enabled without some external hack, you'd have to set the Intel GPU as primary in BIOS/UEFI and then hope that Linux comes up fine as quite a few BIOS disable both outputs if primary output does not have a screen connected.

If that's the case, you can trick it with a dummy screen adapter, these devices are very liked by cryptocoin miners and the like.
https://www.headlessghost.com/ (hdmi)
http://www.geeks3d.com/20091230/vga-...ga-dummy-plug/ (vga/DVI-I)

[nevion]

I thought AMD would be releasing this. They've got something coming on with MIOpen... probably better product is my estimation given they got much more business riding on it. Who knows what languages that will be for, it might just be HIP and not opencl. hipDNN.