News Intel will spend $25 billion on manufacturing its new chips at TSMC: Report

[Admin]

In 2024, Intel will place $14 billion worth of orders with TSMC to fab 3nm Lunar Lake CPU tiles, per a report from semiconductor analyst Andrew Lu.

Intel will spend $14 billion on manufacturing its new chips at TSMC: Report : Read more

 

[-Fran-]

So... The baker is saying they'll be using the oven of the bakery across the street to make their stuff and tell other people to still use their ovens to bake their stuff?

Bad analogy, but you get the point XD

Welp. Good luck with your foundry, Intel.

Regards.

 

[Eximo]

I don't think anyone is going to complain about having more fab space. Either it drives prices down or it keeps them roughly where they are while meeting ever increasing demands.

AMD was switching to Samsung for GPUs, am I remembering that right?

 

[bit_user]

So... The baker is saying they'll be using the oven of the bakery across the street to make their stuff and tell other people to still use their ovens to bake their stuff?

Bad analogy, but you get the point XD

Welp. Good luck with your foundry, Intel.

When we first learned of their plans to fab GPU dies & tiles at TSMC, there was a claim that TSMC's node was more suitable for GPUs than Intel's. Something about the blend of clock speed, power, density, etc. that it's tuned for.

If true, it might still be the case that Intel's foundry is tuning their nodes for CPUs.

 

[dlheliski]

The EE times article says *4* billion in 2024, not 14 billion. And 10 billion in 2025. From the article:

Lu reportedly issued a note to clients saying that Intel will order US$4 billion of 3nm chips for its Lunar Lake personal computer processor from TSMC in 2024. This will be followed by orders worth US$10 billion in 2025.

 

[TerryLaze]

So... The baker is saying they'll be using the oven of the bakery across the street to make their stuff and tell other people to still use their ovens to bake their stuff?

Bad analogy, but you get the point XD

Welp. Good luck with your foundry, Intel.

Regards.

Why do you think that intel is making so many new fabs in the first place?!
And TSMC as well for that matter.
They are currently maxing out their production and need to use external fabs on top of their own.
With the margins for electronics having been destroyed, going for higher volume is the only way to keep making money.

 

[-Fran-]

Why do you think that intel is making so many new fabs in the first place?!
And TSMC as well for that matter.
They are currently maxing out their production and need to use external fabs on top of their own.
With the margins for electronics having been destroyed, going for higher volume is the only way to keep making money.

To waste taxpayers' money while everyone is afraid of "big bad red" in the USA and their allied friends?

Heh. Half joking, half serious, for sure.

Demand for electronics is definitely going up and it would be a fool's errand to argue otherwise. I just question the "bleeding edge" part of the story. Or I should say nitpick.

A simpler reading would be that Intel just doesn't want everything being fabbed internally and that's it.

Regards.

 

[Eximo]

A simpler reading would be that Intel just doesn't want everything being fabbed internally and that's it.

Probably more to do with margins and image. If they reserve their own fabs for large AI chips and ASICs for other companies, good for business. The average person won't know the Intel laptop/desktop chip they buy isn't manufactured by Intel so they get to stay in the public consciousness.

 

[-Fran-]

Probably more to do with margins and image. If they reserve their own fabs for large AI chips and ASICs for other companies, good for business. The average person won't know the Intel laptop/desktop chip they buy isn't manufactured by Intel so they get to stay in the public consciousness.

True. It's also a good source of income when you don't need the things you are fabbing outside.

In other words: everything being fabbed outside for Intel must imply they're not "business critical" to their plans. Or so I'd imagine.

Regards.

 

[JarredWaltonGPU]

The EE times article says *4* billion in 2024, not 14 billion. And 10 billion in 2025. From the article:

Lu reportedly issued a note to clients saying that Intel will order US$4 billion of 3nm chips for its Lunar Lake personal computer processor from TSMC in 2024. This will be followed by orders worth US$10 billion in 2025.

It's been updated, thanks.

 

[The Historical Fidelity]

When we first learned of their plans to fab GPU dies & tiles at TSMC, there was a claim that TSMC's node was more suitable for GPUs than Intel's. Something about the blend of clock speed, power, density, etc. that it's tuned for.

If true, it might still be the case that Intel's foundry is tuning their nodes for CPUs.

The tuning part I believe is correct. In fact, I’ve heard alleged testimony about Intel’s Fab division (before being spun off into IFS) as being beholden and second class to Intel’s CPU design teams. Specifically, forcing the Fab team to rework a fully taped out process node’s design rules to overcome lazy and inferior CPU design work, demanding the Fab team find ways to improve the CPU team design’s inherent power consumption problems instead of via architecture design, allegedly throwing the Fab team under the bus for things in released products that should have been caught by the CPU design validation team, etc.

Basically, it seems like Intel’s bleeding edge process nodes were tweaked for CPU designs and nothing else and over time re-worked as legacy nodes to be used for micro-controller, WiFi, etc. production. I will need to reach out to my friend and ask if things have changed since IFS. It would make sense that Intel 4, 3, 20A, 18A, etc. have generic design rules so they can fit the most design use cases.

 

[jp7189]

I for one will be very interested to see an AMD vs Intel comparison when they are both manufactured on the same TSMC node.

 

[svan71]

So basically Intel is admitting their inability to produce 5 or 3 nm and the excuses and reasoning behind owning their own foundries are all now irrelevant and paying TSMC is actually a good thing.

 

[thestryker]

When we first learned of their plans to fab GPU dies & tiles at TSMC, there was a claim that TSMC's node was more suitable for GPUs than Intel's. Something about the blend of clock speed, power, density, etc. that it's tuned for.

Historically Intel always optimized for CPUs first and then as the next node was being ramped up expanding the prior for other uses (if it's applicable). That's why 22nm is still alive and kicking and I believe is their most diverse node ever.

If true, it might still be the case that Intel's foundry is tuning their nodes for CPUs.

I'd bet this isn't as true for the refinement nodes (Intel 3/18A) and it wouldn't surprise me if there was a lot more node diversity going forward.

Probably more to do with margins and image.

More likely margins and capacity since Intel 3 isn't going anywhere and they've still got to bring 20A up to volume as well as 18A. I've not seen anything regarding their 20A/18A plans, but it seems likely Intel 3 will supplant Intel 4 for the most part.

 

[bit_user]

I for one will be very interested to see an AMD vs Intel comparison when they are both manufactured on the same TSMC node.

It already exists! Just compare the AMD RX 7600 to the Intel A770. Both made on TSMC N6!

Where they differ is that the AMD GPU is only 204 mm^2, while the Intel GPU is 406 mm^2. So, to make the comparison more fair, one might normalize the results by die area. That should also cover the fact that the AMD GPU has just a 128-bit memory interface, while the Intel GPU is 256-bit.

I'm having some trouble finding full benchmarks with the most recent drivers, however.

Intel's New GPU Drivers Boost Performance Up To 750% in DX11, 53% in DX12

Massive gains are incoming.

www.tomshardware.com

 

[8086]

So... The baker is saying they'll be using the oven of the bakery across the street to make their stuff and tell other people to still use their ovens to bake their stuff?

Bad analogy, but you get the point XD

Welp. Good luck with your foundry, Intel.

Regards.

Remember when AMD used to say "Real Men have their own FABs"?

 

[thestryker]

It already exists! Just compare the AMD RX 7600 to the Intel A770. Both made on TSMC N6!

While you're not wrong I think they were referring to modern mature architecture rather than a heavily flawed first shot.

I'm having some trouble finding full benchmarks with the most recent drivers, however.

​

Intel's New GPU Drivers Boost Performance Up To 750% in DX11, 53% in DX12

Massive gains are incoming.

www.tomshardware.com

​

GN put out a "1 year later" Arc video yesterday.

TPU's A770 in the charts here was retested on 2023-06-27 and I think that's likely to be the latest numbers in print: https://www.techpowerup.com/review/sparkle-arc-a750-titan-oc/

edit: unless Jarred retested for this https://www.tomshardware.com/reviews/intel-arc-a580-review-a-new-budget-contender

 

[bit_user]

While you're not wrong I think they were referring to modern mature architecture rather than a heavily flawed first shot.

Well, yes. Clearly, Alchemist is at a bit of a disadvantage, especially considering the RX 7600 is a 3rd generation RDNA design.

However, I don't really buy this excuse for Intel having quite so much trouble as they did, given how heavily-derivative their dGPUs are from their iGPUs and how long they've been in that business.

GN put out a "1 year later" Arc video yesterday.

Waste of time, if you're looking for benchmarks. He basically just recaps the status of all the issues they fixed, mitigated, and which still remain.

One thing he confirms is Jarred's claim that their idle power remains stubbornly high.

unless Jarred retested for this https://www.tomshardware.com/reviews/intel-arc-a580-review-a-new-budget-contender

Well, here are the results from that review:

​

So, if you adjust for die size, the RX 7600 is 115.1% faster per mm^2 on rasterization (i.e. more than twice as fast per die area) and 42.9% faster per mm^2 on ray tracing (nearly 1.5x as fast per die area).

 

[thestryker]

However, I don't really buy this excuse for Intel having quite so much trouble as they did, given how heavily-derivative their dGPUs are from their iGPUs and how long they've been in that business.

It is actually, and Raja admitted to as much. They made some fundamental errors due to assumptions they were working on from IGPs. He didn't show much (because who in their right mind is going to highlight flaws) but he did show some absolutely massive bottlenecking in the design. That revelation is part of why I've been so surprised they haven't released a revision (unless the problem was endemic and they had to fix it for Battlemage too).

 

[cyrusfox]

Pretty meaningless these numbers, especially without context to how it compares to what they spend fabbing their own (What is the internal cost/revenue equivalent for Raptor Refresh/Meteor/and Arrow for 2024 and 2025).

With those numbers we could piece together a story. Only conclusion I see is Intel secured a contract for a certain amount of leading edge for lunar lake (and GPU???/others), the big question is will their be a follow-up?

Is Intel going to do a dual source CPU strategy going forward long term or is this a one time deal due to their prior execution issues?

 

[bit_user]

Pretty meaningless these numbers, especially without context to how it compares to what they spend fabbing their own (What is the internal cost/revenue equivalent for Raptor Refresh/Meteor/and Arrow for 2024 and 2025).

That data is certainly non-public, but these numbers look very substantial if you compare them to Intel's revenues.

 

[thestryker]

That data is certainly non-public, but these numbers look very substantial if you compare them to Intel's revenues.

According to TSMC's financial data for 2022 Intel made up ~$3.87b so the numbers depend entirely on whether or not it's purely new spending (though the $10b is certainly an outlier). Also the report itself is highly suspect which is why my additions to this thread aren't really related to it.

 

[P1nky]

It's been updated, thanks.

You might wanna edit the following as well:

Those figures might seem strange considering the order values: 15,000 wafers per month in 2024 is worth $14 billion, while double the wafers would only be worth $10 billion in 2025.

 

[jp7189]

It already exists! Just compare the AMD RX 7600 to the Intel A770. Both made on TSMC N6!

Where they differ is that the AMD GPU is only 204 mm^2, while the Intel GPU is 406 mm^2. So, to make the comparison more fair, one might normalize the results by die area. That should also cover the fact that the AMD GPU has just a 128-bit memory interface, while the Intel GPU is 256-bit.

I'm having some trouble finding full benchmarks with the most recent drivers, however.

​

Intel's New GPU Drivers Boost Performance Up To 750% in DX11, 53% in DX12

Massive gains are incoming.

www.tomshardware.com

​

I guess I should have specified CPUs.

 

[Bluoper]

After reading all of the comments here I have a potential theory that's probably wrong but I thought I'd get it out there. Since metor lake will be a chiplet design wouldn't it be possible that only the igpu is being produced at tsmc? I see some comments saying tsmc is better optimized for gpus.