Intel Graphics Compiler Merges New Vector Compute Backend

[phoronix]

Phoronix: Intel Graphics Compiler Merges New Vector Compute Backend

While Intel on the hardware manufacturing side continues facing stiff challenges, on the open-source software side the company continues making legendary progress. Out in today's Intel Graphics Compiler and in turn Intel Compute Runtime releases as part of their GPGPU toolchain is the recent open-sourcing and integration of their Vector Compute back-end...

http://www.phoronix.com/scan.php?page=news_item&px=Intel-IGC-Vector-Backend

[vegabook]

You'd think that AMD would have learned their lesson with CUDA. Now we have Intel making all this progress everywhere GPU compute, and open source to boot, and AMD's about to be lapped again.

I run a Radeon VII, the world's fastest consumer FP64 compute card by a mile, but I'll gladly give up some performance for a good, unified, software stack, especially if it's open source. Intel is looking might tempting. While I appreciate that AMD is making big efforts here, it still feels like they're lacking leadership on this internally. Rocm is fragmented and feels like it's reacting all the time to external forces, rather than having a vision. If I were a betting man I'd wager that compute software people at AMD are distant second-class citizens to hardware and graphics guys, unable to lobby management effectively for resources, or risk taking.

[lucasbekker]

I run a Radeon VII, the world's fastest consumer FP64 compute card by a mile, but I'll gladly give up some performance for a good, unified, software stack,

Same here, I basically gave up hope that ROCm will ever be meaningful. At the moment, I am hoping that SYCL/DPC++ will work well on the Radeon VII.

[ms178]

You'd think that AMD would have learned their lesson with CUDA. Now we have Intel making all this progress everywhere GPU compute, and open source to boot, and AMD's about to be lapped again.

I run a Radeon VII, the world's fastest consumer FP64 compute card by a mile, but I'll gladly give up some performance for a good, unified, software stack, especially if it's open source. Intel is looking might tempting. While I appreciate that AMD is making big efforts here, it still feels like they're lacking leadership on this internally. Rocm is fragmented and feels like it's reacting all the time to external forces, rather than having a vision. If I were a betting man I'd wager that compute software people at AMD are distant second-class citizens to hardware and graphics guys, unable to lobby management effectively for resources, or risk taking.

The irony is that Intel did hire Gregory Stoner, the former ROCm lead, to help in furthering their new efforts.

[tildearrow]

Typo:

today's Intel Compute Runtime 20.29.17408 releasethat also makes

[zexelon]

Could Intel finally be a reasonable competitor to Nvidia (specifically in GPU compute)? I have no experience running GPU compute processes with Intel GPU hardware, due to the nature of the market everything so far in production has been CUDA and then someone at the very back of the room wearing a red shirt yelling ROCm ROCm until kicked out by security.

[angstorms]

The irony is that Intel did hire Gregory Stoner, the former ROCm lead, to help in furthering their new efforts.

Yes, it will have been two years since I left AMD, Oct 2018 is when I left.

It has been a great experience working with the Intel graphics software engineering team and Intel's Management. They are doing a lot of great things. I just moved on to a new project recently at Intel. So just like you get to see fruits of number project through Phoronix and my friends in the team.

Some day I may write memoirs of ROCm experience. There is a lot more to High-Performance GPU Compute, especially when you are starting from ground zero-like we did back then.

I have some ideas around Level 0 and AMDGPU driver + SPIR-V support AMDGPU LLVM compiler. that would get you closer to getting SYCL running on this stack cleanly on that stack.

On Vega VII ( Vega20) it was a fun part to work on, especially after the experience working on Vega64 ( Vega10) . Fiji Nano, I still have soft spot for.