[pjreddie]

If Intel is smart they will invest heavily into library development. Researchers use whatever is fast and right now cuda is fast, not just because nvidia has the best GPUs (it's about even) but because the primitives in cuda are so much faster. Matrix multiplication on the same hardware is like 3x faster in cuda than opencl or competing libraries and using the neutral networking primitives is even faster. Intel needs to invest in good, low level libraries so researchers can hack on their platform, build new things, etc. Ultimately I think researchers drive what platform gets widely used since training takes so much longer than inference.

[deepnotderp]

No, researchers use whatever is fastest for development and flexible. Otherwise we'd all be using neon.

[auvi]

i agree. development speed is more important. is there any benchmark that compares neon to other systems?

[deepnotderp]

IIRC, because of Winograd kernels, 3x3 convolutions are amazingly fast and almost all the nets now have switched to 3x3 convolutions.

[rdlecler1]

I bought both NVIDIA and AMD. AMD has a big GPU patent portfolio as well. AMD trades at about 1/5th NVIDIA and could be a takeover target for a bigger player trying to move into AI.

In my opinion NVIDIA needs a moat. CUDA is a good start but they need proprietary data that's hard to create. I thought they should buy Yahoo which would give them a large and unique data set that NVIDIA users could tie into through an API.

[visarga]

Won't get anywhere with a dataset. You need dozens or hundreds of them, and they are usually created using Amazon Turk, so anyone with some money can create one. And they need to be commissioned by the AI researchers, because they know best what they want to study.

NVIDIA needs to become more efficient and build deep learning hardware instead of graphics cards that do faster computation than CPU. If only they could create a small form AI processor for robotics that could do vision at 500+ frames per second. Take a look at this video from 7 years ago which does 1Khz vision: https://youtu.be/-KxjVlaLBmk?t=158

We need that.

[rsp1984]

This. Intel has a lot of catchup to do there. NVidia's drivers and libs are simply on a different level. For example I've seen the same OpenGL code execute faster on an ancient Tegra 4 (mobile) platform than on an Intel Haswell with HD Graphics (which is supposed to beat the Tegra 4 by miles according to raw specs).

[mikepurvis]

See OpenCV and TBB for examples (from Intel) of this strategy at work.