4 ms·

> but I don't see what he direction he should have taken the company instead. Software and design support. As pointed out in the article. Intel is geneticall

by bsder 2y ago

> but I don't see what he direction he should have taken the company instead. Software and design support. As pointed out in the article. Intel is genetically incapable of valuing anything other than "The Fab(tm)". VLSI designer ... second class citizen. Software ... second class citizen. etc. The ARC graphics card series is an exemplar of the problem. It was always going to be a market disaster. Fine, accept it. However, Intel should have pushed a couple of cards for free to every single HPC/AI grad department in the US. And perhaps even tossed in a $50K grant to boot. That's, what, $50K*100=$5 million max and a thousand graphics cards? Sure, the top 10-20 departments are well-funded and still going to use Nvidia--the rest will chew on your graphics cards. In return you get a legion of grad students developing software for your hardware. And you will quickly get feedback as to what you need to fix for v2. However, you have to value software for that idea to penetrate. And Intel just ... doesn't.

almostgotcaught 2y ago
> In return you get a legion of grad students developing software for your hardware. That would be worthless. It's so weird that people think academia is contributing in any meaningful way to the top lines or bottom lines here. You're free to look at the commit history of PyTorch/TF/Triton/HF/llama/whatever and judge for yourself.
- alephnerd 2y ago
  > It's so weird that people think academia is contributing in any meaningful way > PyTorch/TF/Triton/HF/llama/whatever and judge for yourself. Where do you think the creators of most of these current models came from and started their careers? More critically, the ecosystem the arose for these to exist was subsidized by research for decades.
- physicsguy 2y ago
  That’s so misguided it’s unreal. I’ve worked at several companies where there have been decent HPC/GPU codebases and at all of them the staff working on that have largely come out of academia in one way or another.
  - almostgotcaught 2y ago
    > come out of academia isn't the same as in academia.
alephnerd 2y ago
> However, Intel should have pushed a couple of cards for free to every single HPC/AI grad department in the US. And perhaps even tossed in a $50K grant to boot. That's, what, $50K*100=$5 million max and a thousand graphics cards Nvidia used to do this in the late 2000s and early 2010s. It did wonders to drive CUDA uptake versus Vulkan, and Nvidia would sponsor plenty of researchers and students to leverage CUDA and Nvidia GPUs in Parallel Programming classes. > you have to value software for that idea to penetrate I'm not sure it's software per say and more so Developer Experience. If a vendor fails to truly interact with their implementers, they lose critical feedback. Intel used to have a very active Technical Marketing Engineer org (the old word for DevEx), but I'm not sure how active or relevant it is anymore.
- rincebrain 2y ago
  > not software per [se] and more so Developer Experience This aligns with my experiences trying to use Intel software. The software was often technically very good, but the user experience trying to use it was horrible, and god help you if you got an error. Larrabee all over, really. I suppose linking to [1] and [2] is probably germane for this discussion. I swear there was another blogpost explaining the tradeoffs made in the Larrabee/Xeon Phi development process in exacting detail, and confirming that it ran gaming workloads pretty well before being shelved, but I can't find it again. [1] - https://pharr.org/matt/blog/2018/04/18/ispc-origins https://pharr.org/matt/blog/2018/04/18/ispc-origins [2] - http://tomforsyth1000.github.io/blog.wiki.html#%5B%5BWhy%20didn%27t%20Larrabee%20fail%3F%5D%5D http://tomforsyth1000.github.io/blog.wiki.html#%5B%5BWhy%20d...
  - alephnerd 2y ago
    Thanks for the links! I absolutely should follow Matt Pharr and Tom Forsyth's blog tbh - I've been out of the HPC world for sometime now but it's getting relevant again.
- physicsguy 2y ago
  Vulkan didn’t even exist when I got a free card from NVidia. For cross platform programming on GPUs people in academia/industry just gave up because performance was worse, the tooling was terrible and library support was non existent. OpenCL, OpenACC, OpenMP GPU offloading have all been disappointing.
beebeebeeber 2y ago
No one wanted to write code for the Knight’s product line. That was the main problem. Followed by p1272,1274,1276 disasters.
- Pet_Ant 2y ago
  Had to look it up, but the magical numbers as the end of the parents quote is the Intel naming of processes. From https://en.wikichip.org/wiki/intel/process#Timeline https://en.wikichip.org/wiki/intel/process#Timeline those are the 14nm, 10nm, and 7nm respectively.

by bsder 2y ago

almostgotcaught 2y ago

> In return you get a legion of grad students developing software for your hardware. That would be worthless. It's so weird that people think academia is contributing in any meaningful way to the top lines or bottom lines here. You're free to look at the commit history of PyTorch/TF/Triton/HF/llama/whatever and judge for yourself.

alephnerd 2y ago
> It's so weird that people think academia is contributing in any meaningful way > PyTorch/TF/Triton/HF/llama/whatever and judge for yourself. Where do you think the creators of most of these current models came from and started their careers? More critically, the ecosystem the arose for these to exist was subsidized by research for decades.
physicsguy 2y ago
That’s so misguided it’s unreal. I’ve worked at several companies where there have been decent HPC/GPU codebases and at all of them the staff working on that have largely come out of academia in one way or another.
- almostgotcaught 2y ago
  > come out of academia isn't the same as in academia.

alephnerd 2y ago

> However, Intel should have pushed a couple of cards for free to every single HPC/AI grad department in the US. And perhaps even tossed in a $50K grant to boot. That's, what, $50K*100=$5 million max and a thousand graphics cards Nvidia used to do this in the late 2000s and early 2010s. It did wonders to drive CUDA uptake versus Vulkan, and Nvidia would sponsor plenty of researchers and students to leverage CUDA and Nvidia GPUs in Parallel Programming classes. > you have to value software for that idea to penetrate I'm not sure it's software per say and more so Developer Experience. If a vendor fails to truly interact with their implementers, they lose critical feedback. Intel used to have a very active Technical Marketing Engineer org (the old word for DevEx), but I'm not sure how active or relevant it is anymore.

rincebrain 2y ago
> not software per [se] and more so Developer Experience This aligns with my experiences trying to use Intel software. The software was often technically very good, but the user experience trying to use it was horrible, and god help you if you got an error. Larrabee all over, really. I suppose linking to [1] and [2] is probably germane for this discussion. I swear there was another blogpost explaining the tradeoffs made in the Larrabee/Xeon Phi development process in exacting detail, and confirming that it ran gaming workloads pretty well before being shelved, but I can't find it again. [1] - https://pharr.org/matt/blog/2018/04/18/ispc-origins https://pharr.org/matt/blog/2018/04/18/ispc-origins [2] - http://tomforsyth1000.github.io/blog.wiki.html#%5B%5BWhy%20didn%27t%20Larrabee%20fail%3F%5D%5D http://tomforsyth1000.github.io/blog.wiki.html#%5B%5BWhy%20d...
- alephnerd 2y ago
  Thanks for the links! I absolutely should follow Matt Pharr and Tom Forsyth's blog tbh - I've been out of the HPC world for sometime now but it's getting relevant again.
physicsguy 2y ago
Vulkan didn’t even exist when I got a free card from NVidia. For cross platform programming on GPUs people in academia/industry just gave up because performance was worse, the tooling was terrible and library support was non existent. OpenCL, OpenACC, OpenMP GPU offloading have all been disappointing.

beebeebeeber 2y ago

No one wanted to write code for the Knight’s product line. That was the main problem. Followed by p1272,1274,1276 disasters.

Pet_Ant 2y ago
Had to look it up, but the magical numbers as the end of the parents quote is the Intel naming of processes. From https://en.wikichip.org/wiki/intel/process#Timeline https://en.wikichip.org/wiki/intel/process#Timeline those are the 14nm, 10nm, and 7nm respectively.