I feel like I'm seeing a lot about how Gelsinger's decisions were the wrong move for Intel, but I don't see what he direction he should have taken the company instead. What is (or was when Gelsinger became CEO) the move for Intel to get back on top?
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.
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.
> 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.
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.
isn't the same as in academia.
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.
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
[2] - http://tomforsyth1000.github.io/blog.wiki.html#%5B%5BWhy%20d...
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.
From https://en.wikichip.org/wiki/intel/process#Timeline those are the 14nm, 10nm, and 7nm respectively.
Larabee was a mess....anybody who as actually been involved with designing a chip which breaks new ground (and is not just repackaging of somebody else's IP) knows that designing and taping out ANY AND EVERY chip is a big mess. It is not easy to tell heaven from hell on this one--and arguably, Intel's management at the time made the wrong call. Pretty much every decision they were making at the time was the wrong decision, which is why Intel is in the state it is in.
When I worked at Intel, Robert Noyce had unfortunately died, but we still had Andy Grove on the business end and Gordon Moore on the technical end. All three of those people are not just good, not even merely great. They were the kind of transcendent geniuses who rewrite the world, and which come along once every 2 or 3 hundred years. Impossible to replace.
AFAIAK, If Gelsinger couldn't fix Intel, it was irreparably broken by the time he got there. We should celebrate him for having the courage to try to fix what 4 previous CEOs couldn't fix.
I understand that it's painful when one of your suppliers has financial difficulty and has to exit a business you depend on. I feel ya. I hope that nobody will be writing blog posts about the OP's decision to partner with Intel in the first place, pointing to it as one of the fatal mistakes made by the company and their leaders.
* Signal Intel is in it for the long term, specifically kill the dividend.
* Be more customer centric/rebuild customer trust, specifically publicly acknowledge the 10nm slip.
Though it seems customer trust in Intel has eroded to the point where only actions will change customers' perception of Intel — I'm not sure what Intel could have communicated to Tofino/Oxide that they hadn't already.
There is no reason why TSMC should be leading Intel on fabrication by so much. Intel has had a crazy amount of time to restructure and refocus when we consider how long their x86 intellectual property lead kept them afloat with their subpar fabrication and delayed manufacturing process updates. They spent many years having faster chips than anyone else despite having subpar fabrication. But now, they can't really escape the fact that ARM is exploding and AMD has the most flexible x86 platform where as each year goes by the number of use cases where Intel is the best choice diminishes.
If you play games, the x3D chips are basically the only sensible choice. If you are looking for servers, AMD is consistently 10% cheaper in all the cloud platforms. If you want a mobile system, anything involving ARM and/or TSMC will get you better battery life.
The fact that fabless players in the market like AMD and Apple can outpace Intel in so many product lines is, frankly, embarrassing. While I think that other processor design shops selling off fabrication divisions made sense, I think that Intel as a company basically makes no sense if they can't get their fabrication shit together. Because if Intel can't fabricate bleeding edge chips what's the point of using their IP?
The biggest problem Intel has at the moment seems to be they can’t split up and they can’t work together.
If you keep both the fab and the design, you have problems. The thing that makes Intel all the money is their CPUs, and because of the state of the fab many of them are made at TSMC.
But TSMC isn’t too keen on giving a good deal to someone who wants usurp them. So Intel is paying more per chip than they would if they just spun off the fab. If they did that, they could have higher margins/lower prices.
But if you spin off the fab, can they actually get enough business to be able to survive? They’re not on the leading process. They need a massive investment to catch up. Despite their geopolitical importance it seems like they would basically need a real continuing subsidy just to exist, which is effectively being provided by the CPU design side of the business right now right?
It’s a real damned if you do damned if you don’t kind of thing.
I don’t know what the right thing to do is. Splitting them up kind of seems like the better idea but I don’t know if Intel’s ego could take that. They were the fab company. Their advantage let them survive multiple large mistakes that could’ve killed other companies.
Now they’ve all piled up.
PS: the CHIPS act money says it goes to Intel. It sounds like if they try to split they may lose it, so would first have to convince the government to let the fab part keep it
Intel needs to step back, enjoy the runway the CHIPs act got them, and really think hard about the top to bottom design and fab of an absolutely killer 2027 / 2028 chip would look like. They aren’t as good at design as Apple, and they aren’t as good at fab as TSMC, but they are within striking distance of both and they have that vertical integration opportunity.
If they can catch up, but not exceeded, then they can’t make a better product than they do right now. I guess it should cut their costs some amount (don’t know how much) which would certainly be a big plus.
But without taking the lead again… would that be enough? I suppose if they caught up they could also fab other people‘s chips as Pat wanted to. If they’re tied for best (or at least close) they could probably pick up accounts that TSMC doesn’t have space for or are perhaps too small to help fill out their capacity.
While I was listening to the podcast I was thinking about how it almost seems like sort of a historical accident that they’re both design and fabrication. Does anyone else do that at this point? I know it was common early on but maybe it’s just not that sustainable by now.
I don’t know. I guess I see splitting the company is more likely to succeed, simply because the design side works and would lose the anchor that it’s tied to.
Successfully pulling off a fix while keeping the company intact may be a better outcome if possible though.
But I don’t think they have to be better than TSMC, or even quite as good. Intel’s strength is the design + fab integration. If they are 90% as good as fab as TSMC and 95% as good as Apple or Nvidia at design… maybe that’s sufficient to be competitive.
The narrative machine for Intel is in full force given how much monies is in play, I wouldn't take anything that seriously.
>What is (or was when Gelsinger became CEO) the move for Intel to get back on top?
Whatever it takes to drive stonks up up up, obviously. If there was one thing Gelsinger could legitimately be blamed for, it is that he wiped out Intel's market valuation. Nothing else matters, because his "interim" replacements are both beancounters whose literal job is to drive stonks up up up.
because his "interim" replacements are both beancounters whose literal job is to drive stonks up up up.
Bean counters can bring stock value up temporarily. The trick is to get out before shit hits the fan. For a comparison, see Boeing. Where shit is hitting the (turbo) fan right now as well.
I have no doubt Nvidia will start losing its way when Huang retires, like almost all other founder-less companies such as Intel.
Being a founder doesn't magically make you the best person for the job of CEO of a company. Huang happens to be a very good leader, but a lot of founders become a drag on their companies later on.
It's just an anecdote, I worked for one place where the company truly outgrew the founder's abilities. He was just some guy with an art degree who was at the right place at the right time with his dorm room business idea. I would argue that his lack of ability and experience squandered a lot of the later stage growth opportunities that could have made the company into a billion dollar public corporation.
I think about a company like WeWork, that's a company that would not exist today if the founder was not kicked out. Some companies like McDonald's would never have expanded to be a household name if their founders were still in charge.
Their decisions may be no more prescient than anyone else's, but their decisions are reliably in different directions than anyone else's.
Their goal is different, the direction they always aim is different, and that consistent aim has an effect over time regardless what mistakes and misfortunes happen along the way.
If you have a specific and clear goal on the horizon and are always aiming for that even though you have to deal with obstacles closer in front of you, over time you still get somewhere.
If your goal is just chasing whatever random butterflies float by, you get nowhere.
How do you get a CEO to believe innovation is the best thing for them and the company? Even if you get them to believe that how do they get their company to innovate?
Innovation is hard work. It involves failures. Just buy back your shares and make money. Then someone else comes and eats your lunch. That is how the world works.
bean counters are for industries like beans - where the market is stable and the only way to liberate shareholder value is to catch the stray beans. they are exactly the wrong answer when "disruptive innovation" is afoot.