Intel and AMD competing with NVIDIA for deep learning dominance?

[calisnow]

Prior to the news bomb of AP 2.0 last week, the horse race had seemed to be between Mobileye's annotated image, curated approach, and NVIDIA's supercomputing unsupervised learning approach (see Amnon Shashua's public poo-poo'ing of NVIDIA's PX2 last year).

But Elon's conference call last week during the AP 2.0 reveal seemed to indicate that there are other players besides NVIDIA building deep learning supercomputers for autonomous cars - namely AMD and Intel. He told press on the call that Tesla had also considered hardware build by AMD and Intel to run Tesla's neural network in Autopilot 2.0. So clearly, even though Tesla went with NVIDIA, there are other choices available.

I have not seen nearly as much press surrounding the efforts of these two companies to develop deep learning platforms to sell to OEM's for autonomous driving. Anyone know more?

 

[mblakele]

Intel likes to have a presence in any application that could increase demand for processors. They made an announcement in August.

I can't find any specific announcements from AMD, but I'm not sure Tesla needs a platform specifically designed for driving. I think they could develop for any GPU platform that meets their specs for performance and can operate in a mobile environment. Tesla has ex-AMD chip architect Jim Keller, and he would know exactly which AMD products might work out.

Anyway we don't know how seriously Tesla evaluated each of these GPU vendors, do we? Maybe some of them were along the lines of "What about X? No good? Ok, moving right along...."

 

[bob_p]

Competition is good. Though today, NVidia appears to be the market leader - their established as the best GPU vendor and have a lot of experience with high performance parallel image processing needed for 3D and VR - so adapting that technology to also work for self driving isn't as a large a leap as what Intel and AMD would need to do.

 

[sandpiper]

I struggle to see why it would be a challenge to move a given network from any one general purpose GPU to another? The various neural network platforms that are out there, Caffe, Theano, Torch, etc... seem to favour Nvidia only because Nvidia has been most aggressive about supporting the various DNN researchers with libraries, technical support and so on.

I can't imagine that you won't end up with equally good Intel/AMD competitors to the CUDA libraries at some point. That should??? allow a given network to be GPU agnostic?? OpenCL, as I understand it, is partway there and I'd be shocked if AMD wasn't in panic mode trying to close the gap.

I imagine that Tesla is being very careful about wrapping themselves up too tightly with any vendor. They may have a good relationship with Nvidia right now, but this is too important to them to be reliant on any outside party.