Intel Lands Support For Vector Neural Network Instructions In LLVM

[phoronix]

Phoronix: Intel Lands Support For Vector Neural Network Instructions In LLVM

Intel continues bringing up support for the 2019 Icelake processors within the open-source compiler toolchains...

http://www.phoronix.com/scan.php?page=news_item&px=LLVM-Intel-Icelake-VNNI-Support

[Cape]

Wtf is LLVM?

[bridgman]

http://llvm.org/

[coder]

Dear Intel,
It's nice to see your support for popular computing trends, but nobody serious is going to be either training or inferencing neural nets on x86, in 2019.

Sincerely,
Reality

[Zan Lynx]

Dear Intel,
It's nice to see your support for popular computing trends, but nobody serious is going to be either training or inferencing neural nets on x86, in 2019.

Sincerely,
Reality

So, so wrong. What, do you think the entire world wants to shove all their data into "the cloud" to do machine learning? I for one, hope that we move back to keeping our data local. Which means we need to be able to train neural nets on a laptop CPU.

Sure it won't be as good as using a Nvidia V100 but it will be mine, all mine.

[coder]

So, so wrong. What, do you think the entire world wants to shove all their data into "the cloud" to do machine learning? I for one, hope that we move back to keeping our data local. Which means we need to be able to train neural nets on a laptop CPU.

Who said anything about cloud? My point was orthogonal to that.

In the short term, use GPUs. Even a mid-range desktop GPU is way faster than their Skylake server CPUs. High-end desktop GPUs are ~2x as fast as Knight's Landing, and Knight's Mill is unlikely to completely close that gap.

Even if you're on a laptop with only the integrated GPU, you'd do much better to use it than the CPU cores. Starting with Broadwell, their HD Graphics GPUs have double-rate processing of fp16, meaning close to 1 TFLOPS on the upper-end models.

In the longer term, dedicated neural processing chips and potentially FPGAs will likely outclass even GPUs.