Forget the White House Sideshow. Intel Must Decide What It Wants to Be.

[OneEng]

Let me just start with one man's saying "Real Men Have Fabs" .

AMD thought that the reason Intel was able to beat them consistently was because Intel had their own fabs. I think that ironically, this is both true AND at the same time false.

If it were completely true, then AMD would have been able to equal Intel's product line .... but they didn't (mostly).

The REASON AMD couldn't make the vertical integration model work for them like it worked for Intel was because (IMO) Intel was able to amortize their considerable foundry investments across TONS more units. Intel was cash rich and was able to command a foundry lead over the entire industry... resulting in products that had great margins .... thus letting them continue to fund the foundry (which honestly, I consider to be the key to Intel's past success).

This was a business model created at a time when the majority of chips made in the world were made by Intel. I speculate that this fact is why the vertical integration worked so well for Intel for so long.

Fast forward to more recent history ......

New process nodes are exponentially more expensive to implement and are much more complex than those of the past. Ironically, Intel is making the same mistake (IMO) that AMD made all those years ago. They are trying to compete with a fab (TSMC) that makes nearly 2 times more wafers per year than Intel. This is a tough position for Intel IMO.

Another point to consider is Intel's risk taking in the foundry. Just like with 10nm, Intel's 18A is a very risky node IMO. Intel can't afford another 14++++ moment.

I believe that it would be in Intel's best interest to sell the fab to someone who could run it like a business instead of a cost center. The foundry can't be run "the Intel way". It needs to be run "the industry way" using common tools and processes. The foundry needs to become a world class service provider .... and that may well be a much bigger change than many imagine it to be.

Intel's design group is also in for an overhaul IMO. I think that they have long relied on the foundries industry leading processes to bolster their designs. For the first time with ARL we are able to analyze a leading edge Intel design against a leading edge AMD design (ARL vs. Zen 5) on process technologies that are well know relative to one another.... and honestly, it isn't very impressive for ARL IMO. ARL using N3B (the best of the best ... and most expensive TSMC has to offer) vs Zen 5 on N4P and Zen 5 nearly sweeps the field.

Finally, AMD has correctly adopted a "Server First" design approach that is greatly boosting their profits. Intel had dominated this high margin market for years and now, when that market is growing faster than ever, Intel has lost the leadership in this area to AMD (by quite a lot).

I am eager to hear your thoughts on my analysis.

 

[siliconbruh999]

AMD thought that the reason Intel was able to beat them consistently was because Intel had their own fabs. I think that ironically, this is both true AND at the same time false.

i think it can be said out who has the best fabs

I believe that it would be in Intel's best interest to sell the fab to someone who could run it like a business instead of a cost center. The foundry can't be run "the Intel way". It needs to be run "the industry way" using common tools and processes. The foundry needs to become a world class service provider .... and that may well be a much bigger change than many imagine it to be.

like i have said it would make more sense for Intel to give up leading edge logic keep fabs to themselves 18A/AP is going to be a long node like 10nm/14nm no point in selling off which is going to return itself that is poor decision making and short sided. They can simply get their investment back from the fabs by not doing any leading edge after 18A/AP. Just say Uncle Sam go to TSMC if you want leading edge we can only do 18A if you don't fund it like they have done LBT played 4D Chess.

Intel's design group is also in for an overhaul IMO. I think that they have long relied on the foundries industry leading processes to bolster their designs. For the first time with ARL we are able to analyze a leading edge Intel design against a leading edge AMD design (ARL vs. Zen 5) on process technologies that are well know relative to one another.... and honestly, it isn't very impressive for ARL IMO. ARL using N3B (the best of the best ... and most expensive TSMC has to offer) vs Zen 5 on N4P and Zen 5 nearly sweeps the field.

On this part Intel is pretty competitive in Mobile it's just ARL sucked for Desktop

 

[hist78]

They are trying to compete with a fab (TSMC) that makes nearly 2 times more wafers per year than Intel.

I’m curious how you arrived at this “2×” figure. I’m thinking it’s much larger.

 

[Xebec]

Intel's design group is also in for an overhaul IMO. I think that they have long relied on the foundries industry leading processes to bolster their designs. For the first time with ARL we are able to analyze a leading edge Intel design against a leading edge AMD design (ARL vs. Zen 5) on process technologies that are well know relative to one another.... and honestly, it isn't very impressive for ARL IMO. ARL using N3B (the best of the best ... and most expensive TSMC has to offer) vs Zen 5 on N4P and Zen 5 nearly sweeps the field.

i agree; though N3B is "the first N3" node, while N4P is a heavily refined N5 node. I think the performance of the nodes at the transistor level may be basically equal.

An older chart shows N4P as +11% performance vs N5, and N3 (B) as having "+10-15%" performance vs N5. Power is also listed as -22% vs -25-30%, so I'd say the playing field is roughly level as far as transistor performance and power go.

 

[MKWVentures]

I’m curious how you arrived at this “2×” figure. I’m thinking it’s much larger.

@OneEng @Xebec

Good question. So test numbers

1) how many N543 wafers per month does TSMC run
2) How many N543 wafers does TSMC run for Intel
3) How many Intel 4/3 wafers per month does Intel run for itself

My Over-simplistic model is that 8% of TSMCs sub N7 wafers are for Intel

I would love to hear estimates

 

[hist78]

@OneEng @Xebec

Good question. So test numbers

1) how many N543 wafers per month does TSMC run
2) How many N543 wafers does TSMC run for Intel
3) How many Intel 4/3 wafers per month does Intel run for itself

My Over-simplistic model is that 8% of TSMCs sub N7 wafers are for Intel

I would love to hear estimates

My estimate for Intel’s total wafer production in 2024 is between 0.65 million and 1.3 million 12-inch wafers, under the best assumptions for Intel.

In comparison, TSMC produced 12.9 million 12-inch–equivalent wafers across its 6, 8, and 12-inch fabs in 2024, with the majority being 12-inch. This means Intel’s wafer output volume is far smaller than TSMC’s. TSMC’s 12-inch wafer output is probably 6 to 8 times greater, or even more, than Intel’s.

To avoid dragging this thread too far off topic, I’ll post my calculation in a separate thread. Link is listed below:

How Many Wafers Did Intel Produce in 2024

My assumptions and method: 1. Server outsourcing volume – Assume Intel has very limited outsourcing of server products. 2. Wafer outsourcing ratio – Intel CFO David Zinsner stated that Intel’s goal is to keep no more than 30% of wafer needs outsourced. Since Intel did not publish exact...

semiwiki.com

 

[MKWVentures]

My estimate for Intel’s total wafer production in 2024 is between 0.65 million and 1.3 million 12-inch wafers, under the best assumptions for Intel.

In comparison, TSMC produced 12.9 million 12-inch–equivalent wafers across its 6, 8, and 12-inch fabs in 2024, with the majority being 12-inch. This means Intel’s wafer output volume is far smaller than TSMC’s. TSMC’s 12-inch wafer output is probably 6 to 8 times greater, or even more, than Intel’s.

To avoid dragging this thread too far off topic, I’ll post my calculation in a separate thread. Link is listed below:

How Many Wafers Did Intel Produce in 2024

My assumptions and method: 1. Server outsourcing volume – Assume Intel has very limited outsourcing of server products. 2. Wafer outsourcing ratio – Intel CFO David Zinsner stated that Intel’s goal is to keep no more than 30% of wafer needs outsourced. Since Intel did not publish exact...

semiwiki.com

Good data, but lets focus on advanced nodes. How many Intel 4/3 wafers are made by Intel and how many N543 wafers are bought by Intel.

BTW: Intel 7 Is still dominating Intel 4,3,18A combined at Intel. Its actually quite shocking how high Sapphire/Emerald Rapids and Raptor lake (or even alder lake) volumes are.

 

[siliconbruh999]

Good data, but lets focus on advanced nodes. How many Intel 4/3 wafers are made by Intel and how many N543 wafers are bought by Intel.

Intel 4/3 wafers at this point of time would be less than N543 but this will start reverse next year with GNR still ramping.
Clearwater Forest and Diamond Rapids Nova Lake Panther Lake and more SKU using IFS 18A/3 and N6.

 

[count]

Good data, but lets focus on advanced nodes. How many Intel 4/3 wafers are made by Intel and how many N543 wafers are bought by Intel.

BTW: Intel 7 Is still dominating Intel 4,3,18A combined at Intel. Its actually quite shocking how high Sapphire/Emerald Rapids and Raptor lake (or even alder lake) volumes are.

Intel 7 is good enough for CPUs now. It's only mobile and GPU that really needs more performance.

Even mobile, I feel like 3nm is probably good enough...

 

[hist78]

Good data, but lets focus on advanced nodes. How many Intel 4/3 wafers are made by Intel and how many N543 wafers are bought by Intel.

BTW: Intel 7 Is still dominating Intel 4,3,18A combined at Intel. Its actually quite shocking how high Sapphire/Emerald Rapids and Raptor lake (or even alder lake) volumes are.

It’s difficult to estimate Intel’s wafer output for a particular node, especially since Intel is very tight-lipped about each fab’s capacity.

 

[MKWVentures]

It’s difficult to estimate Intel’s wafer output for a particular node, especially since Intel is very tight-lipped about each fab’s capacity.

Add Meteor Lake+GR+SF product sales. I have a "model" graph by fab but that is NOT based on confidential Information

 

[Daniel Nenni]

Should the US Government Invest in Intel? - Semiwiki

“Most companies don’t die because they are wrong; most die because they don’t commit themselves. They fritter away their valuable resources while attempting to make a decision. The greatest danger is standing still.” Andy Grove's Only the Paranoid Survive, first published in 1996. Looking back...

semiwiki.com

 

[hist78]

BTW: Intel 7 Is still dominating Intel 4,3,18A combined at Intel. Its actually quite shocking how high Sapphire/Emerald Rapids and Raptor lake (or even alder lake) volumes are.

Pat Gelsinger used to say that Intel 7 is neither technology nor cost wise competitive. Now if Intel 7 gets so much production volume compare to Intel 3, 4, and 18, it might be a bad sign for Intel overall. The possible reasons could be:

1. Intel 7 is neither technology nor cost wise competitive, but Intel Intel 4, 3, and 18A might be even worst.

2. Intel started using EUV with the Intel 4 while Intel 7 does not use EUV tools. Intel might not have enough EUV machines and/or enough Capex to acquire enough EUV machines.

3. Intel OEM customers, such as DELL, Lenovo, HP, Asus, and Acer, are shy away Intel more advanced products due to market demand and profit/cost consideration.

4. Intel 3 and Intel 4 yield is not ideal and capacity and output is limited too. Intel must use Intel 7 related products to fill the gap.

 

[hist78]

Add Meteor Lake+GR+SF product sales. I have a "model" graph by fab but that is NOT based on confidential Information

Can you share some numbers?

 

[count]

Pat Gelsinger used to say that Intel 7 is neither technology nor cost wise competitive. Now if Intel 7 gets so much production volume compare to Intel 3, 4, and 18, it might be a bad sign for Intel overall. The possible reasons could be:

1. Intel 7 is neither technology nor cost wise competitive, but Intel Intel 4, 3, and 18A might be even worst.

2. Intel started using EUV with the Intel 4 while Intel 7 does not use EUV tools. Intel might not have enough EUV machines and/or enough Capex to acquire enough EUV machines.

3. Intel OEM customers, such as DELL, Lenovo, HP, Asus, and Acer, are shy away Intel more advanced products due to market demand and profit/cost consideration.

4. Intel 3 and Intel 4 yield is not ideal and capacity and output is limited too. Intel must use Intel 7 related products to fill the gap.

Cost competitiveness also depends on yield and volume.... it's possible that Intel 7 is cost competitive from unit cost standpoint if it's yielding well and the volumes are high... and that might also explain why they are keeping productions on Intel 7... because it makes that node more profitable (at the expense of other nodes, but that is tomorrow's problem).

 

[hist78]

but that is tomorrow's problem

With shrinking revenue and negative free cash flow, Intel doesn’t have many options.

 

[Xebec]

Good data, but lets focus on advanced nodes. How many Intel 4/3 wafers are made by Intel and how many N543 wafers are bought by Intel.

BTW: Intel 7 Is still dominating Intel 4,3,18A combined at Intel. Its actually quite shocking how high Sapphire/Emerald Rapids and Raptor lake (or even alder lake) volumes are.

Fwiw TSMC N3 still hasn't overtaken N5 - 36% of TSMCs revenue was N5 last quarter vs 24% for N3, and that's despite N3 wafers costing more.

It does show TSMC is the only company running "massive volume" on "3nm" but it also argues for cost / capacity challenges even at TSMC for "3nm" class processes.

(If a N3 wafer costs 2X N7 then TSMC may also still be running as many N7 wafers as N3).

 

[MKWVentures]

Cost competitiveness also depends on yield and volume.... it's possible that Intel 7 is cost competitive from unit cost standpoint if it's yielding well and the volumes are high... and that might also explain why they are keeping productions on Intel 7... because it makes that node more profitable (at the expense of other nodes, but that is tomorrow's problem).

@hist78

Intel has stated Margins are higher on Intel 7 than Intel 4.... and it is fully capacity limited. Intel 4, 3, 18A, etc are demand limited, not capacity limited. And Intel stated that customers want Raptor lake.... Intel is not pushing them to it. I just bought another computer and the deals on Raptor lake Notebooks were great

This is another culture change.... back in my day, we would encourage people move to the next node with price increases on the old node. Can't do that anymore in a competitive environment.

 

[OneEng]

like i have said it would make more sense for Intel to give up leading edge logic keep fabs to themselves 18A/AP is going to be a long node like 10nm/14nm no point in selling off which is going to return itself that is poor decision making and short sided. They can simply get their investment back from the fabs by not doing any leading edge after 18A/AP. Just say Uncle Sam go to TSMC if you want leading edge we can only do 18A if you don't fund it like they have done LBT played 4D Chess.

This would only work for the short run IMO. Just like Global Foundries, Intel would eventually end up producing low margin parts. Now, if they make enough of those parts, it adds up and makes money.

A side issue is that I am unsure if all designs will work well with BSPDN. Unlike TSMC that has created a common library for both N2 and A16 (which I think they are still planning BSPDN for) allowing chip designs to try out one process or the other depending on which one works best for the design (If I understand it correctly). 18A appears to be an "all in" jump for a designer.

Of course, one would think that Intel will work this out in the long run.

On this part Intel is pretty competitive in Mobile it's just ARL sucked for Desktop

Yes, but they did so with a much more expensive process node.

I’m curious how you arrived at this “2×” figure. I’m thinking it’s much larger.

I threw a question at ChatGPT and did a few searches and went conservatively thinking that someone would call me out for being to aggressive with my estimate . I am going to go with your numbers though. Still, my point is even more valid with your numbers. The much higher volume makes it so much easier to pay for the equipment and future investment.

With shrinking revenue and negative free cash flow, Intel doesn’t have many options.

.. sigh .. and this is really the crux of the problem. Intel went "all-in" with 18A and with High NA equipment. The numbers I have heard are in excess of 20bn USD in investment in the development of 18A. I have stated on several occasions that this is the cost of a US Ford class aircraft carrier.

My opinion is that this "Foundry Race" is simply not a race that Intel can win. The realities of today's lithography equipment cost, and the competitive landscape are conspiring to destroy the vertical integration model Intel has mastered over the decades.

@hist78

Intel has stated Margins are higher on Intel 7 than Intel 4.... and it is fully capacity limited. Intel 4, 3, 18A, etc are demand limited, not capacity limited. And Intel stated that customers want Raptor lake.... Intel is not pushing them to it. I just bought another computer and the deals on Raptor lake Notebooks were great

This is another culture change.... back in my day, we would encourage people move to the next node with price increases on the old node. Can't do that anymore in a competitive environment.

Wow. I hadn't thought of that. I now distinctly remember a design I made with an Intel Q7 board that was "guaranteed" for like 10 years of availability. Such CRAP. About 4 years into that 10 years the parts got so expensive it was cost effective for me to spend millions of dollars to do a redesign with the newer version than to keep paying for the old part.

Yea, I am still butt hurt about the whole debacle .

With this in mind, if Intel's foundry DOES decide that they will simply stick with 18A for a long time vs spending another 20bn to move to the next advanced node, TSMC could simply price them out of profitability by making N2 more cost effective when they move on to A14 and beyond.

I believe that the only way Intel Foundry can continue (and I think they should) is with help. I think that it is in the nations best interest to ensure that the foundry continues; however, I don't know how much appetite the US government has for spending 20bn a year to keep our edge in semiconductor tech.

Also, there is the fact that each advanced node is giving less and less advantage over the previous node. If things keep going this way (and I can't see why they wouldn't), we can all expect the technology ramp to slow down across the board.

In fact, this may have some really profound economic impacts (says the guy typing on a Zen 3) beyond the discussion we are having here, and for far more than just Intel.

 

[hist78]

This is another culture change.... back in my day, we would encourage people move to the next node with price increases on the old node. Can't do that anymore in a competitive environment.

When Pat Gelsinger and Intel were pushing the ‘5 Nodes in 4 Years’ plan, I questioned how Intel could convince OEM customers (Dell, Lenovo, HP, Asus, Acer, etc.), as well as consumers and businesses, to keep upgrading to the latest and greatest in a short 4-year time frame. If each node truly succeeded and went into high volume production, then where would the customers come from?