A very one-sided article. Blames everything on Intel without even considering that maybe Habana Labs was just better at getting the founders rich from the acquisition then at making AI processors. And maybe Intel management noticed that all wasn't going as planned and tried to intervene, which may have made things worse.
Sure, Intel have quite a history of failures, especially for the last decade or so. But I would like to think that even Intel can be smart enough not to mess with things that work, so maybe things didn't work and Intel management (unsuccessfully) tried to save it. It is actually a very common scenario: branch doesn't perform as well as expected, upper management tries to do something, it doesn't work, people blame management but don't mention that without intervention it would have also failed. Of course management is to blame, but it is not just because of them.
I worked for another ML-related startup Intel acquired around the same time, and nothing in this article sounds biased or exaggerated to me. "Pursuing multiple competing AI strategies without fully committing to any" is a reasonable description of what was going on. We found it very difficult to get anything done, and everyone quit after the retention bonus period was over.
The article is -- if anything -- generous to Intel. The unvarnished truth (which I have heard from several different Intel insiders) is quite a bit less flattering, and shouldn't be surprising: Intel's failures at this point are fractal.
(If the article has a failing, it's that it's somehow trying to hang the failure of Habana inside of Intel on Avigdor, which is absurd.)
Back in 2019 I looked at what Habana was doing and had a conversation with their CTO. My impression was they had competent technical execution and a good understanding of some of the areas where other chip efforts have stumbled (software stack especially). I expressed at the time that I hoped they would remain independent, and if they sold not to Intel, but of course I don't begrudge the team cashing out for the right price.
Given Intel's string of failures with nearly everything outside their core competency (and in the past decade stumbling even there), I don't understand giving them the benefit of the doubt in this. What at Intel would have changed between all their failures, Habana Labs and the following failures that would make Habana Labs not their fault?
They bought it for $15bn in 2017 and sold out some stake in it at a valuation of $17bn in 2021. Without looking at the overall economy of the time, that doesn't look wildly successful?
They didn't kill it, which is a win I suppose, but it sounds like it just carried on as it's own pace?
IIRC, Nvidia was already ahead in software support 5 years ago. Image recognition was the big thing and the only way I could train them was on Nvidia GPU or Google tpu.
With this context, the articles clain that
>Intel correctly identified AI as the future
seems tenuous at best.
It also claims that Nvidia was just a graphics company in 2019-20. This was clearly not the case for anyone training cv models at the time.
IDK what Gaudi is or what Habana Labs was developing. Does anyone know what the status of pytorch/tensorflow on Intel GPU hardware is? Because add far as research is concerned one of the framework has to be supported.
I'd say this is true - they just failed to execute meaningfully on it. They were adding stuff like VNNI back in 2018, buying Nervana and Habana, but then just...didn't do much with it. At the same time, they made Ponte Vecchio.
Software support for all of these is not great, in my personal opinion.
Nvidia GPUs have been used to train DNNs since around 2011 (the paper on how to do it was written in 2010 I think?). Nvidia was also used heavily to mine crypto currency until recently.
2012? If we are using Alexnet as the "starting" mark for DNN training on GPUs. Though probably there were efforts before that, Alexnet is the most well known for sure.
> As the speed of GPUs increased rapidly, it was soon possible to train deep networks such as convolutional networks without the help of pretraining as demonstrated by Ciresan and colleagues in 2011 and 2012 who won character recognition, traffic sign, and medical imaging competitions with their convolutional network architecture. Krizhevsky, Sutskever, and Hinton used a similar architecture in 2012 that also features rectified linear activation functions and dropout for regularization. They received outstanding results in the ILSVRC-2012 ImageNet competition, which marked the abandonment of feature engineering and the adoption of feature learning in the form of deep learning. Google, Facebook, and Microsoft noticed this trend and made major acquisitions of deep learning startups and research teams between 2012 and 2014. From here, research in deep learning accelerated rapidly.
…
> AlexNet is a convolutional network architecture named after Alex Krizhevsky, who along with Ilya Sutskever under the supervision of Geoffrey Hinton applied this architecture to the ILSVRC-2012 competition that featured the ImageNet dataset. They improved the convolutional network architecture developed by Ciresan and colleagues, which won multiple international competitions in 2011 and 2012 by using rectified linear units for enhanced speed and dropout for improved generalization. Their results stood in stark contrast to feature engineering methods, which immediately created a great rift between deep learning and feature engineering methods for computer vision. From here it was apparent that deep learning would take over computer vision and that other methods would not be able to catch up. AlexNet heralded the mainstream usage and the hype of deep learning.
What's baffling is the hardware is actually pretty good (from direct experience), and the software stack wouldn't take much effort to bring up to par either, at least for inference. Perf/$ is much better than NVIDIA. More HBM than NVIDIA, too. And somehow nobody gives a shit.
Once loaded, weights just stay in HBM. And PCIe throughput is not at all an issue - the cards communicate over a large number of very high speed ROCE links both inside and outside the server. HBM throughput is not as much of an issue either, since during inference you are mostly just reading weights, not also writing out intermediate results and gradients. HBM2e is actually a smart decision on Intel's part IMO - it's not as supply constrained, so they can have more of it. Gaudi2 has 96GB of HBM2e. Gaudi3 has 144GB IIRC, as well as about 4x more compute and even faster interconnect. This means you can do longer context and/or larger models without splitting it across hosts where bandwidth is at somewhat of a premium. Though Gaudi3 is also less price competitive, I must add. No relation to Intel - I have never worked for them and hold no position in their stock. I just had pretty good experience working with Gaudi2.
Once, Nortel Networks was a $100 billion telecommunications giant and could have bought Cisco for $20 billion. It was too late for Cisco when they realized they were awfully behind in IP networking race. So they bought Bay Networks for $4 billion. It didn't work and soon after, the company of 96,000 people dropped to 30,000 in a few months, and then silently disappeared from the history.
back then, Nortel bought some Israeli startup. Deal was mostly paid by in nortel stock. Everybody in startup became millionaire over night. So they went to Israeli IRS to negotiate a "deal" for reduced taxes/etc. While they were negotiating nortel stock dived and everybody lost everything.
edit. startup was chromatis. but it was lucent who did buying and not nortel.
Intel is #1 example of why you don’t let management consultants come inside your company. They are simply the worst. Why do you let a bunch of PE/VC-rejects dictate your corporate strategy?
Honest question: are there any happy acquisition stories? Like, a big company acquired a startup and with startup's vision and big company's resources it achieved success to the glory of the acquirer?
google has botched or killed any number of acquisitions, but google earth and google docs are spectacular success stories that involved acquiring practically unknown startups and then scaling them way up. (I do realise this is likely a simplification of the actual stories behind them.)
Google was a originally search company, but from what I can see their major success was figuring out how to run computer systems at reliably and at scale before anyone else. Everything else you see now - Ads, Maps, Android, Waymo, Youtube was a weed translated onto that infrastructure that went onto thrive. Yeah, there were many, many more weeds that didn't survive the transplant. But since they didn't cost Google much they don't matter.
Given what you see from Google now is dominated by the "everything else" (I doubt even losing search would be fatal), Google is a poster boy for growing a technology company by acquiring new products. Given the rumours about their ongoing cultural shift (I'm sure Intel was like Google at one stage), this may not continue. But for new Google is the exemplar for now to do this right.
I think many pure businesses have been acquired successfully (ie YouTube). But the tech acquisitions require very careful management. The hard part is after the acquisition, when tech and culture need to ge. Leadership likes to declare victory when the acquisition closes.
How often have you seen acquisitions of technology work out? What often happens is new company gets tacked onto an existing business unit and there are pretty massive culture clashes, the tech was oversold, and at best there’s an aquihire of decent engineers.
I’ve more often seen it’s something seen as easy and very visual thing that looks like you’re solving a problem instead of looking internally and making tough choices. Or having to take a critical look at leadership.
I attended an Intel "AI day" in London back in about 2018, before they bought Habana and had all of their focus on Nervana (mentioned in the article) and Saffron - which I've basically never heard mention of again since that day. seems a shame; while the AI accelerators have obvious utility, it was Saffron which sounded like it might be a game-changer.
Intel specifically sought Israeli startups, so I'm a little confused about how this is framed. It's also hard to understand how Intel buying your startup for 2 Billion is ruining it, as most people on this site would see that as a huge win. I detect bias around this topic.
>[...] Falcon Shores, had received negative feedback from customers and would therefore not be marketed commercially. [...] As a result, Intel decided not to develop another generation of Gaudi beyond the existing Gaudi 3. This effectively marks the failure of Habana Labs as another entry in Intel’s long history of unsuccessful acquisitions. [...] Willenz, considered a genius in the semiconductor world, has a long history of successful exits.
The article is saying that Intel failed to lead the startup to developing a successful product by taking steps, not detailed, that the founder may have taken if he had still owned it.
It doesn't work that way. As a rule, most large acquisitions die a few years later, irrespective of the acquiring company, unless they are maintained as separate subsidiaries. And there's no way the Intel mafia would allow it to be separate in the long run. Old boys network eventually moves in, and everything goes to shit. I have seen this happen half a dozen times in my career.
My understanding is that the Intel management, especially former CEO Pat Geislinger, are ideological supporters of Israel [1], and that they have repeatedly made decisions to prop up bad investments in Israel because of that ideology.
One of those decisions was to put an Intel factory on the land of a Palestinian village that had been ethnically cleansed by Israel during the 1948 Nakba.
Another was to announce in Dec 2023, at a time when Israel was losing international support due to what was becoming internationally recognized as a genocide, that it was about to invest $25B in a new factory in Israel [2]. Intel investors were not happy and Intel was made to reverse this decision.
> At the time, Intel and Amazon were the dominant players in the market, while Nvidia was primarily known to gamers who admired the graphics capabilities of its processors.
I couldn't believe that either. The RTX 3090 comes out sept 2020. More importantly, the final form of the GA102, the A6000/A40 (with 48GB of RAM, every single cuda core enable and $5k+ MSRP) comes out a few days later. We were well past the point of Nvidia market segmenting out its best parts for server. The DGX was on its third generation at that point and cost $200k. And I think cuda came out a little before then, right? It was... oh yeah... Nov 2006.
This was also the height of the global GPU shortage... I suppose anything can fit your narrative, just as long as your spin has enough RPMs.
It should be discussed somewhere but maybe not this thread. I would prefer if every HN thread doesn't turn into a referendum or tutorial on tangential topics.
This isn’t some random GitHub discussion thread. This is a discussion about Intels business interests in Israel and why they keep investing in Israel.
You won’t find a more appropriate place to discuss their overall ideology. And it’d really telling that administration here keeps flagging those comments.
It shows a pattern of Intel investing in Israel due to ideology versus business reasons. Zionists may not like it, but it’s directly related to their Jewish supremacist beliefs that Intel keeps investing so much money in the middle of a war zone.
Do we really want to avoid discussion about ideology in business investments? The fact that this site flags any mention of Palestine is irresponsible. This site really needs to figure out a way to not let Zionists control the business discussion.
It’s Ok to allow discussion of the negative consequences of Zionists ideology on technology businesses.
>Nvidia has spent the last two decades refining its AI processors, while Intel has only recently recognized the opportunity—arriving late, like a sluggish hippopotamus at a sprinting contest.
This isnt strongest argument
Copying and improving things takes less time.
For example how many years Nvidia needed to develop 4060 4080 and how many years Intel needed to reach similar capabilities
Sure, Intel have quite a history of failures, especially for the last decade or so. But I would like to think that even Intel can be smart enough not to mess with things that work, so maybe things didn't work and Intel management (unsuccessfully) tried to save it. It is actually a very common scenario: branch doesn't perform as well as expected, upper management tries to do something, it doesn't work, people blame management but don't mention that without intervention it would have also failed. Of course management is to blame, but it is not just because of them.
Another point of view would be interesting.