News TSMC founder says Intel should focus on AI, not advanced process technologies

[Admin]

Founder of TSMC criticizes Intel for not developing a sophisticated strategy for AI and instead focusing on becoming a foundry

TSMC founder says Intel should focus on AI, not advanced process technologies : Read more

 

[Kondamin]

Sure 200k per 2nm wafer

 

[Jame5]

That has to be the most self-serving criticism ever. "You focus on developing designs. Leave the foundry work to me so I can make even more money."

 

[hotaru251]

foundry has long term value.
"ai" has a low lifespan compared to being a foundry.

 

[-Fran-]

"You know that guy down the street that also has a bakery? Yeah, him. He's such a good carpenter, as he built his own shop, he should get out of the baking business and do wood work instead [that way I'm the only baker around town]. Totally. ".

Regards.

 

[logainofhades]

I will be glad when the AI bubble bursts. So sick of hearing AI this and AI that.

 

[abufrejoval]

Statements like this are like an iceberg.

There is so much more going on the public won't ever see within their attention span.

But I'm pretty sure he's trying to steer the course of silicon history in a way that benefits his legacy.

 

[why_wolf]

TSMC announces Intel should give them more money. I'm sure there would be no financial repercussions of their being only one company in the world who could make said chips.

 

[Dustyboy1492]

This guy is so arrogant and condescending. TSMC will have issues someday and I will be here for it.

 

[KraakBal]

*smug anime boss laugh*

 

[bit_user]

I will be glad when the AI bubble bursts. So sick of hearing AI this and AI that.

What if it's more like the Internet? Sure, we had a dot-com bubble, and it did burst, but the Internet is still with us.

 

[bit_user]

BTW, it's not as if Intel has been oblivious to AI. Their Xeon Phi product line was aimed at the same sorts of HPC workloads as Nvidia had targeted. Intel started pivoting more towards AI in 2017, with Knights Mill (the last Xeon Phi models they produced), not long after Nvidia's P100 had started making waves.

Going as far back as Broadwell or Skylake, their iGPUs also had packed fp16 dot product, which was a sort of weird thing to put in there, unless you were anticipating AI and convolutional neural networks. They did this even before Nvidia put it in the P100.

Then, in 2016, Intel bought Movidius. They also bought Nervana Systems, that year, which was their server-based AI strategy (with Movidius being decided more client/edge-focused). Then, in 2020, they upgraded their server AI stack with the acquisition of Habana Labs. They hedged this bet by building AI horsepower into Ponte Vecchio (marketed as the Xe Datacenter Max).

So, I think you can fault Intel for execution, but it's not like they were blind-sided or haven't been working on AI for long enough. Perhaps what they lacked was more focus and a better appreciation of how much harder it is to compete with Nvidia than practically anyone else they had faced.

 

[antoshkaplus]

He is trying to say that Intel did right going for 18A. And while ex-CEO is a great guy, Intel may need an outsider to have less arrogance for competition.

 

[King_V]

What if it's more like the Internet? Sure, we had a dot-com bubble, and it did burst, but the Internet is still with us.

I read it more like "AI" currently being a buzzword as opposed to actual useful artificial intelligence. Sort of like "dot com" was the buzzword rather than the Internet. Reminds me a little of the time not long ago when you couldn't go 24 hours without hearing "Big Data" multiple times.

Hell, we've all seen it. "AI" being the label slapped on simple things like keyword parsing, which has been around forever.

 

[bit_user]

Reminds me a little of the time not long ago when you couldn't go 24 hours without hearing "Big Data" multiple times.

Hell, we've all seen it. "AI" being the label slapped on simple things like keyword parsing, which has been around forever.

Well, modern AI is partly an evolution of Big Data. Without massive troves of data, you can't train very sophisticated models, which was a big driver behind the acquisition of so much data. And I'm not just talking about LLMs, which seem to garner most of the attention and focus, but a lot of the other things people are doing with AI.

 

[usertests]

Sure 200k per 2nm wafer

Actual cost is supposedly over $30k, less than $40k, or 1/5th that.

Even if it was $200k, that's maybe $300 for a Ryzen/Epyc CCD, or $2500 for an AD102-sized die. So if we eventually see that price for some super advanced monolithic 3D node, it will be worth it for some customers. Wafer Scale Engines already sell for $millions using the whole wafer.

 

[Pierce2623]

Well I mean obviously he doesn’t want the only company that can really compete with their process nodes to become a publicly available foundry….

 

[Pierce2623]

BTW, it's not as if Intel has been oblivious to AI. Their Xeon Phi product line was aimed at the same sorts of HPC workloads as Nvidia had targeted. Intel started pivoting more towards AI in 2017, with Knights Mill (the last Xeon Phi models they produced), not long after Nvidia's P100 had started making waves.

Going as far back as Broadwell or Skylake, their iGPUs also had packed fp16 dot product, which was a sort of weird thing to put in there, unless you were anticipating AI and convolutional neural networks. They did this even before Nvidia put it in the P100.

Then, in 2016, Intel bought Movidius. They also bought Nervana Systems, that year, which was their server-based AI strategy (with Movidius being decided more client/edge-focused). Then, in 2020, they upgraded their server AI stack with the acquisition of Habana Labs. They hedged this bet by building AI horsepower into Ponte Vecchio (marketed as the Xe Datacenter Max).

So, I think you can fault Intel for execution, but it's not like they were blind-sided or haven't been working on AI for long enough. Perhaps what they lacked was more focus and a better appreciation of how much harder it is to compete with Nvidia than practically anyone else they had faced.

I agree. Intel definitely wasn’t late to the AI party. Their design abilities are just significantly less advantageous than they were before.

 

[Pierce2623]

What if it's more like the Internet? Sure, we had a dot-com bubble, and it did burst, but the Internet is still with us.

That’s exactly what everyone expects. All these startups will go out of business eventually but nobody expects AI to just disappear back into the genie bottle….

 

[Pierce2623]

Sure 200k per 2nm wafer

What? 2nm wafers will be $20,000-$25,000 for big volume like Apple up to a max of $35,000 for small batch chips. 3nm small batch wafers are only about $28,000 now and high volume customers pay right at $20,000 a wafer.

 

[abufrejoval]

This guy is so arrogant and condescending. TSMC will have issues someday and I will be here for it.

If statements like this were a habit of his, I'd agree.

Here it sound a bit more like part of a farewell address, but with more to show than Mr. Gelsinger and fewer of such public statements in the past.

At that point at worst I'd call it more pride than arrogance, and while any advice from a mountain of success doesn't guarantee equal succes for anyone that follows it, anyone else quite literally hasn't ascended to that same level and has too look up to him looking down, even with a guy that isn't very tall.

 

[abufrejoval]

BTW, it's not as if Intel has been oblivious to AI. Their Xeon Phi product line was aimed at the same sorts of HPC workloads as Nvidia had targeted. Intel started pivoting more towards AI in 2017, with Knights Mill (the last Xeon Phi models they produced), not long after Nvidia's P100 had started making waves.

Going as far back as Broadwell or Skylake, their iGPUs also had packed fp16 dot product, which was a sort of weird thing to put in there, unless you were anticipating AI and convolutional neural networks. They did this even before Nvidia put it in the P100.

Then, in 2016, Intel bought Movidius. They also bought Nervana Systems, that year, which was their server-based AI strategy (with Movidius being decided more client/edge-focused). Then, in 2020, they upgraded their server AI stack with the acquisition of Habana Labs. They hedged this bet by building AI horsepower into Ponte Vecchio (marketed as the Xe Datacenter Max).

So, I think you can fault Intel for execution, but it's not like they were blind-sided or haven't been working on AI for long enough. Perhaps what they lacked was more focus and a better appreciation of how much harder it is to compete with Nvidia than practically anyone else they had faced.

And then I'd even add the neuromorphic hardware, the Loihi chips, where Intel tried to jump far beyond NN simulations on GPGPU.

Intel often has been a mix, of jumping too far ahead and not far enough.

Just think iAPX 80432, iAPX 850/860, Itanium and I don't know how many others.

 

[bit_user]

I agree. Intel definitely wasn’t late to the AI party. Their design abilities are just significantly less advantageous than they were before.

Their first mistake was going with x86, for Larrabee. They very nearly made the right decision of using their GMA IP for Larrabee, but the x86 team won that battle and lost them the war, ultimately dooming both Larrabee and Xeon Phi.

The Movidius team is designing the NPUs found in Meteor Lake, Lunar Lake, and their predecessors (GNA and GNA2, I think). So, that arguably worked out reasonably well.

I honestly don't know enough about the Nervana IP to say why Intel decided Habana was a better bet, but I assume they probably had good reasons.

Ponte Vecchio seems like it had the potential, but somebody went absolutely bonkers with tiles of all different process nodes and stacking them every which way! If they'd kept it far simpler, it could've shipped on time (or much closer, at least) and still would've offered compelling performance.

Habana is another case where I just don't know enough to say why they fell short. I wonder if Intel maybe erred in meddling too much with them, post-acquisition, or maybe erred in giving them too little support? Though, I'm also don't want to presume their shortcomings weren't due to other factors, entirely.