News Chips aren't getting cheaper — the cost per transistor stopped dropping a decade ago at 28nm

[Admin]

Google says transistor cost scaling stalled at 28nm and remains flat with every new node.

Chips aren't getting cheaper — the cost per transistor stopped dropping a decade ago at 28nm : Read more

 

[edzieba]

This is not 'new' news, it has been known for the last decade too. When the gate oxide thickness limit was reached ~22nm, it meant transistor scale basically remained fixed since then (barring geometry changes like GAA) and voltage scaling also ceased (why chips have been at ~1v for a decade). It's the main factor behind why for chips where performance scales closely with transistor count that those same designs have been reaching a price/performance asymptote.

 

[bit_user]

It's nice to see this issue get more publicity. I've picked up on it since the current pricing trends in GPUs became a topic of controversy. Here are a few more key details I've found.

Why is that happening? Let's start by looking at wafer price trends:

​

So, should we blame TSMC or ASML for being greedy?

​

Maybe not. Newer wafers are objectively more resource-intensive to produce.

Furthermore, design costs are also increasing at a similar pace:

​

It makes sense that dies with exponentially more components would take longer to design & test, especially to the extent that humans remain in the loop.

while new nodes may no longer make transistors cheaper, they still make a lot of sense for many designs that either cannot be disaggregated efficiently or are hard to disaggregate due to complications with their manufacturing.

And let's not forget performance (and its sister, efficiency)! Newer nodes are nearly always more efficient (e.g. offering more performance at the same power; or less power at the same performance). That's why cutting edge designs will always prefer newer nodes for logic, which still exhibits good scaling.

A big argument for chiplets is that SRAM and I/O don't scale as well as logic. So, if the benefits of a newer node are less (or even negligible), then why pay the higher price? Putting them on a separate die enables cost savings. If you're already doing a multi-die design for yield reasons, this becomes a very easy decision.

 

[bit_user]

This is not 'new' news, it has been known for the last decade too.

The article says as much, in the very first sentence.

When the gate oxide thickness limit was reached ~22nm, it meant transistor scale basically remained fixed since then (barring geometry changes like GAA)

Wasn't fin FET also intended to work around that?

It's the main factor behind why for chips where performance scales closely with transistor count that those same designs have been reaching a price/performance asymptote.

Your reasoning is too simplistic. Essentially, you're claiming that transistor scaling stopped, yet it hasn't.

I think the main story is that wafer costs have increased at a similar rate as density, because it's become more complex and required more expensive equipment to push density ever higher.

A side-story is that density improvements have plateaued for things like I/O and SRAM. However, it does appear still to be ongoing, for logic.

 

[Co BIY]

Also explains clearly why TSMC is still building capacity (And consolidating older customers) at 28nm and others like Global Foundries stopped investing in advanced nodes at around the same size. It's a sweet spot in a mature industry at that point.

 

[bit_user]

Also explains clearly why TSMC is still building capacity (And consolidating older customers) at 28nm and others like Global Foundries stopped investing in advanced nodes at around the same size. It's a sweet spot in a mature industry at that point.

Back in 2019, Eben Upton said, in an interview on this site, that the Raspberry Pi was likely to stay at 28 nm, since that's where the lowest cost per transistor was (and didn't appear likely to change, in the near future).

"28nm brings quite substantial improvements in energy. It's also the current "value node" (lowest cost per transistor), and I think will remain so for the foreseeable future."

https://forums.tomshardware.com/thr...-ask-your-questions-now.3492239/post-21119030
(Eben Upton, June 27, 2019)
​

I was therefore a bit surprised that the Pi v5 seems to have moved to a smaller node, but I guess they decided the market demands for more performance were sufficient to justify a higher price.

 

[PaulAlcorn]

Here's a fun throwback; Lisa Su pointed this out back in 2013 with a chart:

https://www.design-reuse.com/articles/36150/moore-s-law-is-dead-long-live-soc-designers.html

and again in 2020 at Hot Chips.

 

[bit_user]

Here's a fun throwback; Lisa Su pointed this out back in 2013 with a chart:

https://www.design-reuse.com/articles/36150/moore-s-law-is-dead-long-live-soc-designers.html

and again in 2020 at Hot Chips.

Well spotted!

Perhaps this is one of those things the semiconductor industry almost considered common knowledge, these days. Us outsiders are only now beginning to take notice, as we start to see those semiconductors price trends having a meaningful impact on product pricing.

It does put Intel's E-core & AMD's C-core strategy in an interesting light. If transistor costs aren't going down, then you'd expect to see greater emphasis on "area-efficiency", in new designs. As long as they can increase performance per mm^2, they can increase core count and thereby perf/$ can continue to scale faster than the efficiency gains of new process nodes.

P.S. I think this article deserves to be featured in one of the panels on the front page, no?

P.P.S. Go Chiefs!!!

​

: )

 

[refillable]

This analysis is too overrated, too simplistic. GM200 had almost ten times less transistors than AD102, yet the 980 Ti didn't launch anywhere near 300 USD. People think "5nm chips" are just a bunch of 5nm class transistors packed around a die. No they're not.

 

[bit_user]

This analysis is too overrated, too simplistic. GM200 had almost ten times less transistors than AD102, yet the 980 Ti didn't launch anywhere near 300 USD.

They didn't say anything about end product prices, did they? It was just an analysis of manufacturing costs. A lot else goes into a graphics card and its price, including software (i.e. drivers).

I think that explains why we didn't feel the impact of higher silicon pricing until more recently. In those days, the actual GPU chip cost was probably a much smaller proportion of the total cost of a graphics card.

 

[refillable]

You have to have some degree of specificity, you can't analyze "manufacturing costs" and proceed to conclude that "transistor costs stopped going down at 28nm". It's too blanket. What are they specifically talking about? The node? What is a "node", precisely? The whole IC? Those are made up of an assortment of different transistors!

It's also a moving target due to node maturity. 28 nm in its introduction (2011) had huge capacity and yield problems, making it not any cheaper than 5 nm now per die size (GK104, 2012, 294 mm^2 was $668, adjusted for inflation). Sound's familiar? Yes, that's the 4070 Ti and the 4070 super (AD104).

They didn't say anything about end product prices, did they? It was just an analysis of manufacturing costs. A lot else goes into a graphics card and its price, including software (i.e. drivers).

I think that explains why we didn't feel the impact of higher silicon pricing until more recently. In those days, the actual GPU chip cost was probably a much smaller proportion of the total cost of a graphics card.

There's a reason why I didn't say exactly 90% less (that'd be 159.99), or 124 USD, adjusted for inflation.

If you want a more representative example, take a look at 2700K vs 14700K. Former had a 450 USD (adjusted) MSRP, 995M transistors, fabbed on 32 nm while latter has a 420 USD MSRP and 14B transistors (estimated). 28 nm was considered the "half shrink" of 32. "Transistors became 15 times cheaper", no?

Go back 4 nodes from 32, you have Northwood at 130 nm. It was released at ~500 USD (adjusted, prices hard to track) and had 55M transistors. "Transistors had became 20 times cheaper", this time yes?

Wait what?

 

[bit_user]

You have to have some degree of specificity, you can't analyze "manufacturing costs" and proceed to conclude that "transistor costs stopped going down at 28nm". It's too blanket. What are they specifically talking about? The node? What is a "node", precisely? The whole IC? Those are made up of an assortment of different transistors!

The statement surely presumes some mixture of components characteristic of CPUs and GPUs, I'd imagine. We're not talking power management ICs, here.

It's also a moving target due to node maturity.

Yes, it's a fair question whether they're looking at launch prices or legacy prices. I'd wager they're probably looking within the window of when a node is yet to be superseded, because that's what's relevant to the cutting-edge products. Even so, that's still a couple years, during which wafer pricing & yield probably changes a decent amount.

Regardless of what point they picked in nodes' lifecycles, I'm reasonably confident they're comparing equivalent datapoints from each node.

If you want a more representative example, take a look at 2700K vs 14700K.

Why those two? One is a full die, the other isn't. The equivalent comparison would be i7-2700K vs. i9-14900K.

"Transistors became 15 times cheaper", no?

They're talking about fabrication prices, not final product pricing. More goes into the final sales price of a CPU than that.

Go back 4 nodes from 32, you have Northwood at 130 nm.

That was still on the curve.

 

[refillable]

The statement surely presumes some mixture of components characteristic of CPUs and GPUs, I'd imagine. We're not talking power management ICs, here.

Of course not power ICs, MOSFETs on your motherboard usually has ~3 μm transistors. Larger ones are probably used in PSUs.

If we widen the definition a bit to include "transistors that makes up computing clusters", sure, that's also possible, but how is that supposed to be calculated? How about the SRAMs, are they considered too? IO/interface channels? I'm almost certain they don't consider any of these lest they calculate them one by one.

But for the sake of simplicity, let's go ahead with the presumption and have it at that definition.

Yes, it's a fair question whether they're looking at launch prices or legacy prices. I'd wager they're probably looking within the window of when a node is yet to be superseded, because that's what's relevant to the cutting-edge products. Even so, that's still a couple years, during which wafer pricing & yield probably changes a decent amount.

Regardless of what point they picked in nodes' lifecycles, I'm reasonably confident they're comparing equivalent datapoints from each node.

Fair enough, 14700K was unrepresentative. If you change the 14700K to 12900K, you'll get 9.5 instead of 15. Lower, but narration remains. Where is the "stagnation" in 9.5-fold? The 12900K was the first Intel 7 mainstream desktop processor, and the first 32 nm was Clarkdale, an MCM CPU+GPU. In this case, the 13900K is more appropriately compared to the 2700K. Now we have 10.7, not bad, isn't it? But you get me, it's not easy.

You might have told me "they got it from insider information!" But consider that most Intel CPUs sold in 2011 were 32nm while in 2022, most were Intel 7. Gross margins dropped from 60 to to 39. But can you explain how the drop remains at just 21pp with a comparable revenue when they're making 15 times more transistors? (Which according to this "analysis" costs about the same to produce each)?

Now, if you consider that this cost refers not to the cost imposed on the fabs, with many companies going fabless, then what does it refer to? Surely they normalized the effect of the economics of scale, right?

That was still on the curve.

Eh, what? That was quite a plot twist.

 

[bit_user]

Where is the "stagnation" in 9.5-fold? The 12900K was the first Intel 7 mainstream desktop processor, and the first 32 nm was Clarkdale, an MCM CPU+GPU. In this case, the 13900K is more appropriately compared to the 2700K. Now we have 10.7, not bad, isn't it? But you get me, it's not easy.

As I said before: they're talking about fabrication prices, not final product pricing. More goes into the final sales price of a CPU than that.

You're not accounting for the costs of: R&D, support, marketing, or packaging (either chip packaging or product packaging, for that matter). Don't forget that R&D includes chip design, testing, and lots of software Intel has to write, from Windows device drivers to Linux kernel patches. These costs shouldn't have increased nearly as fast as the wafer fabrication costs, which means they're diluting any increases in actual silicon production costs.

Next, consider that Intel 7 is still a DUV node, and therefore should be behind the cost curve of TSMC's 7 nm. The data is on TSMC wafer pricing, because they're a lot more transparent than Intel traditionally has been. I think only Intel truly knows what their fabrication costs have been.

Eh, what? That was quite a plot twist.

I mean that you're looking back before transistor cost flattened out.

 

[refillable]

Next, consider that Intel 7 is still a DUV node, and therefore should be behind the cost curve of TSMC's 7 nm. The data is on TSMC wafer pricing, because they're a lot more transparent than Intel traditionally has been. I think only Intel truly knows what their fabrication costs have been.

Neither is 20nm, 16nm, 10nm, and 7nm DUV from TSMC. If Intel used EUV on 7, what difference does that make? Yet this report showed that "transistor costs has been plateauing since 28nm".

I mean that you're looking back before transistor cost flattened out.

I don't know how this is a problem in my narrative.

As I said before: they're talking about fabrication prices, not final product pricing. More goes into the final sales price of a CPU than that.

You're not accounting for the costs of: R&D, support, marketing, or packaging (either chip packaging or product packaging, for that matter). Don't forget that R&D includes chip design, testing, and lots of software Intel has to write, from Windows device drivers to Linux kernel patches. These costs shouldn't have increased nearly as fast as the wafer fabrication costs, which means they're diluting any increases in actual silicon production costs.

Let's create a simple model. Assuming that "other costs" are constant.

Y is total production cost
X1 is transistor production cost (including anything that is related to transistor fabrication: initial R&D for the node, new equipment, initial testing, node deployment, yield increasing, fine tuning, you name it)
Y-X1 is everything else, let's name it X2. X2 is, as you said, constant.
so, Y = X1 + X2

Y+ZY is revenue
Z is gross margin

In 2011 when sandy bridge was released, Intel's quarterly revenue was 13.9 billion, with a margin of 60%. When raptor lake was released, Intel's quarterly revenue was 14 billion, with a margin of 39%.

In 2011:
Y + 0.6Y = 13.9B
Y = 8.69B
Therefore,
8.69B = X1 + X2

In 2022:
Y + 0.39Y = 14B
Y = 10.1B
10.1B = 14X1 + X2

Here, with the magic of linear algebra, we get that:
X1 is roughly 108.5M
X2 is roughly 8.58B

Therefore, in order for your assumption to be true, the fabrication cost was only 0.8% the total cost of Intel's operating cost in 2011.

This also implies that intel only spend 108.5M dollars for EVERY ONE OF THEIR CPUs manufactured in Q4 2011. We know that AMD paid global foundries $300 million every quarter between 2010 and 2012, when they still made their GPUs in TSMC!

This definitely did not happen. TSMC would've been bankrupt a long time ago with those rates!

 

[bit_user]

...

All I'm going to say is that I'm not an expert on semiconductor manufacturing and, by the looks of it, neither are you. If you were, I'm sure you wouldn't be trying such a backwards and simplistic approach of estimating Intel's fab costs.

So, go ahead and disregard what they're saying. All it means is that you'll mis-attribute the blame, when you find yourself having to fork over more money for your next CPU. I guess that's not really my problem, though.

 

[refillable]

All I'm going to say is that I'm not an expert on semiconductor manufacturing and, by the looks of it, neither are you. If you were, I'm sure you wouldn't be trying such a backwards and simplistic approach of estimating Intel's fab costs.

Well, an oversimplified approach, maybe. With a 1000% error rate I'll have my numbers shoot up to 8.8%, which still won't satisfy anyone.

But, Backwards? How?

So, go ahead and disregard what they're saying. All it means is that you'll mis-attribute the blame, when you find yourself having to fork over more money for your next CPU. I guess that's not really my problem, though.

Sure, you can say that, this is not a forum with huge restrictions. I'll keep disregarding what they're saying, because the fact always contradicts them. 28 nm does not bring the cheapest per-transistor prices for these applications. Do I have to blame someone to really tell to myself that I won't have to fork more money for your next CPU? Maybe not!

If I'm not an expert, then Lisa surely is, right?

Pitcairn --> 28 nm, ~212 mm2, 2.8B Transistors --> Lisa's estimated cost: 1.7x
Polaris --> 14 nm, ~232 mm2, 5.7B Transistors --> Lisa's estimated cost: 2.2x
Navi 10 --> 7 nm DUV ~251 mm2, 10.3B Transistors --> Lisa's estimated cost: 3.8x
Navi 23 --> 7 nm EUV ~237 mm2, 11.1B Transistors --> Lisa's estimated cost: 3.8x
Navi 32 GCD --> 5 nm ~200 mm2, 19.9B Transistors --> Lisa's estimated cost: 4.9x

You tell me!