Arthur Hanson
Well-known member
Very dangerous strategy to achieve success, it's more like a plan to fall further and further behind.
Why Lunar Lake changes (almost) everything
- Thread starter TICin852
- Start date
tooLongInEDA
Moderator
Historically true, but given Intel's recent financial performance for its data center products (recall that this used to be the golden goose only a few short years ago), I don't think they have as much margin to play with in servers as you suggest. Surely, cost matters for pretty much all Intel's chips now - remember they're taking on PE financing for fabs now to avoid a credit rating downgrade.
That's the reason I'm a little sceptical about Intel's stated plans to bring all the chip manufacturing back in house. They need the TSMC option to keep the Intel fabs honest and help them drive down costs. Remove the potential competition and they'll lose that. I don't think that Intel foundry cometing with TSMC will have anything like the same leverage on getting Intel fab costs down.
Given Intel Foundry size and costs (and yield?) compared to TSMC, I think there's zero doubt that a comparable chip -- server or laptop -- will cost more.
However gross margins on server products -- especially high-end ones -- are traditionally very high, at least 60% IIRC, so if that's reduced (e.g. to 50%) by Intel fab higher prices, they're still making a shedload of money on each sale while keeping the Intel foundry busy.
Of course that doesn't mean there won't be pressure on Intel to get foundry wafer costs down, but just the facts of where it is and how big it is means there's no way it can match TSMC on pricing.
tooLongInEDA
Moderator
The funny thing is that last time I checked the Intel financials it was CCG (PC side) that was making the bulk of the profits and DCG (servers) that was under-performing [I may not have kept up with the latest BU names here]. So while what you're saying makes perfect sense to me, it's not the story Intel's public financials are telling us today.
As others have suggested, Intel fabs must (in the short term at least) increase their pricing to the product groups to make their margin goals under the new, transparent internal accounting regime we're promised. You seem to confirm that there's no way for Intel to get level on pricing in the short term. How does that not directly hurt the product groups sales if the option to use TSMC instead is taken off the table in 6 months or a year ?
It doesn't hurt sales volume, assuming Intel have to sell at an ASP that is competitive with AMD to make customers buy them. What hurts sales is not having a competitive product, and that's why Intel server sales are dropping and AMD sales are rising.
It hurts margins/profits, since they make less money on each CPU sold. Still plenty, but less than they would have made using TSMC -- and less than AMD are making, presumably... ;-)
Last edited:
The problem is that when Intel uses TSMC not only does Intel wafer volumes go down, TSMC wafer volumes go up. Economies of scale, yield learning, supply chain advantages, etc... This is the downside of the frenemy strategy in the semiconductor industry, TSMC gets stronger and Intel gets weaker.
Intel wafer costs must come down. I highly doubt Intel will be using TSMC beyond the N3 node. Intel 18A is competitive to TSMC N2 and it is made by Intel. Intel 18A with BSPD is even more competitive. That will be the goto node for Intel design moving forward, my opinion.
I agree that if Intel pull off their process they will have a technology -- and presumably products -- that are technically competitive (performance/power consumption) with AMD, for the first time in several years.
The problem is that if Intel foundry can't compete on price with TSMC -- which seems impossible, not just "challenging" -- they'll be making less money on each HPC/server CPU sold than AMD. So either the product group accepts this or they squeeze the foundry down on price, which means the foundry makes less money, which makes it harder for them to invest in process technology to keep up with TSMC...
It's easy to say "Intel wafer costs must come down" but how can they possibly compete with TSMC who have lower cost and much bigger volume fabs?
What is lucky about using N3? Intel didn't just accidentally stumble into it. This was a conscious choice made when prior management gave up instead of spending money on a neglected manufacturing arm. Calling that luck makes even less sense than Pat saying Jensen was lucky, because at least in that case there are external factors at play (as opposed to intel outsourcing which is only the result of internal pressures).
Because intel wants to make money? Even on cost ineffective intel 7 IF makes a small profit per wafer. This margin stacking allows the wider intel corp to have a lower COGS for each chip than any fabless company. Any external customers which is an extra revenue stream/risk mitigation/diversification strategy is just gravy on top. Of course due to the amount of cash intel is investing/idle fabs, IF has an operating loss rather than a small profit so it is currently an anchor rather than a tailwind. But as I have said MANY times before, the second IF hits breakeven+1 cent intel as an IDM is more profitable than what they could ever do as a fabless company. Even if they can only do small time numbers like a tower (10% op margin) instead of more avg number like UMC (30%) intel corp would be wildly more profitable than AMD without intel products even having to lift a finger. That profitability cannot happen without volume, and intel products is the single largest fabless customer by wafer volume.
As for flexibility why would intel keep huge volumes with TSMC if there intention is to surpass them technologically, doing so would be gimping their products. So unless intel knew that they would never be competive process tech wise there is no reason to commit huge volumes at TSMC beyond derisking a product with the design spread between the two, if TSMC offers a specialty node with no intel equivalent, or to gain pricing pressure over intel foundry when negotiating wafer pricing. Even the intel foundry slides for their 2030 end state still show some outsourcing (albeit at lower levels than historical).
intel BUs should pay less for an intel 3 wafer than an N3 wafer due to the worse PPA. I also wouldn’t assume intel 3 is that far from TSMC structural cost given intel 3 has 30% fewer mask layers and 5% fewer process steps overall than intel 7 (assuming the intel 4 numbers roughly hold true). To further support this, intel rates intel 3 as (-) for areal cost rather than 4x (-) for intel 7, and said they expected their EUV nodes to have 40ish percent margins. So maybe 30-35% on intel 3 wafers?
There is also the fact if you didn’t do these big die Xeons on i3 Oregon and Ireland would just be sitting idle due to how tiny mtl compute die is (like 1/5th the size of a normal a last gen client die), lower volume (only like 30-35M units this year (if we assume not all of the 40M aipc are mtl), and intel 4 yield being described as “VERY good”. When intel was doing capacity planning in the BS days Ireland was in flight and they would have known that TSMC can never be cheaper than an empty fab.
Last edited:
Why in the world would Intel choose to use Intel 7 for the IO die in its server chips? It's common knowledge among everyone that TSMC N7 is much cheaper - and Intel 7 doesn't need any more products to improve yield or increase volume. They already have enough volume on that node.
And why would they even keep it around? I mean, it is a good idea not to leave the fabs idle - but absolutely nobody (apart from Intel) will buy Intel 7 wafers because of how much it costs to manufacture. Isn't it better for them to cut their losses and retool the fabs because the node is simply not profitable?
And why would they even keep it around? I mean, it is a good idea not to leave the fabs idle - but absolutely nobody (apart from Intel) will buy Intel 7 wafers because of how much it costs to manufacture. Isn't it better for them to cut their losses and retool the fabs because the node is simply not profitable?
Last edited:
Doesn’t your question about foundry pricing get to the heart of the matter of the wisdom of Intel pursuing a foundry model? If cost is intrinsically an issue, then why not purely go fabless and rely on TSMC and Samsung for silicon?
It's not realistic to say that IF can operate with low wafer margins to help the product group make more money on selling CPUs, because the foundry is now a separate business and has to make enough profit to be able in invest in new technologies and fabs to keep up with TSMC.
If Intel use IF to make CPUs to compete with AMD using TSMC, there's simply less profit to share out inside Intel -- no matter which division gets it. foundry or product -- to develop new products and technologies.
Because that would be admitting that Intel Foundry was a failure, which wouldn't go down well either inside Intel or in the USA in general.
We went through exactly this same process when I worked for Fujitsu -- going from a traditional IDM to "IDMv2" (using both internal and external foundries) to separating off the foundry business -- which eventually died because it couldn't compete with TSMC. All the same drivers are there in Intel -- the reluctance to admit that you can't compete any more, but being unwilling to admit it and do what is actually necessary... :-(
The only way an IDM-style foundry -- which is what Intel still is today, regardless of the separation of the businesses -- is if it offers some fundamental advantage that customers are willing to pay for, for example security of supply (US-based) or guaranteed early access to the latest technology (no risk of being squeezed out by the likes of Apple).
If IF can find enough such (big enough) customers -- not the guvmint or military, they're *way* too small -- then they will succeed, though on a smaller scale than TSMC. If they can't then the lower costs and bigger volumes (and higher yield, which always comes with volume...) of TSMC will drive them into the ground, regardless of whether their raw technology is up to scratch or not.
Of course there's also a downside to ending up with a TSMC monopoly, which is why many customers (and governments) don't want this to happen -- the problem is that to stop it happening will take a truly enormous amount of money, and when push comes to shove few want to cough up the cash...
I think what you said is only true if IF cost is higher than TSMC_cost + TSMC_profit.
IF cost is likely higher than TSMC cost, but as long it is lower than TSMC cost+profit, IF can still sell to Intel products at the same wafer price, and allow Intel Corp as a whole to be better off.
They could, but as a smaller volume foundry with lower profits than TSMC they'd then be unable to invest enough in developing the next-generation process.
I suspect they also wouldn't be able to sell much more cheaply to Intel than other IF customers now the two have been legally and financially separated (they could do this when they were an IDM) -- and if they have to sell at lower margins to all customers (including non-Intel) to compete with TSMC on pricing, they have even more of a funding problem.
With SRF that already seems to have happened. Of course N3 Turin-Dense will win back perf/watt when it eventually comes out, but the shot has been fired so to speak.
How is IF going to get better if they can just charge BUs cost+some amount to have TSMC margins? If they continue to do pricing based on PPA vs TSMC with a small undercut that forces IF to focus on cost to improve their P&L and gives insight into what they need to improve. Also are we going to pretend TSMC is the only foundry on planet earth? Not even intel says they plan to have better margins or revenue than TSMC. Their 2030 end state is how ever much internal business and more external business than Samsung internal+external plus 40% GM 30% op. Those sort of margin values are in the GF/UMC regime. I don't see why doing that is completely out of the question?
As for how you go about achieving that profitability intel told us:
1) Have technology leadership so you can charge more than TSMC wafers.
2) IDM1.0 was not focused on cost due to the final product being high ASP chips rather than the wafers. The focus was on technology, yield, volume, and minimizing TTM. Now that they are standalone they must be cost conscious and be efficient.
3) Have process flows with competitive structural cost for their density.
1 is what 5N4Y is for. 2 is a mindset change, but folks like GF figured it out so it can't be impossible. 3 is just using the best equipment for the job rather than trying to reuse ancient tools for your new process and have alternatives when something doesn't work (eg 14A being able to easily switch off of high-NA if it doesn't work).
That is only a recent trend. While TSMC total capacity vs intel total is in TSMC's favor. The nodes intel runs are at higher capacity than TSMC. For example intel has/had more capacity for intel 7 than TSMC N5 family. The older nodes were even more in favor of TSMC. Given the use of disag intel's demand for the latest node will always be less than it was, so intel will need external to take their economics of scale lead back from TSMC. If intel products can come back that maybe sounds do-able (but more likely they will only close the gap). The only wrench is if mobile AP folks ever go to chiplets. If they do TSMC's most advanced node at any time will see lower demand (as the demand shifts to older nodes). In that situation intel taking scale back at leading nodes seems much more do-able.
By this logic TSMC would have never gotten anywhere near Samsung and intel due to both firms being wildly larger/more profitable until only recently (in the grand scheme of things anyways). As long as IF op margin is positive intel investing in process and building fabs generates profit for the wider intel corp wile also paying for itself (ie it is self sustaining). So I don't see why intel would need to throw in the towel if TSMC has even 1% better margins. All that matters is if intel has the scale to amortize TD/ramp costs. If they can do that there is no reason not to push forward. The root of the reason folks like GF and TI stopped is because the TD/ramp costs grew large enough that at their volumes moving forward would be unprofitable not because they were less profitable than TSMC.
I think it is a given that IF wafer prices will have to beat TSMC prices for it to establish itself. The question is whether IF will have positive margin at those price points. And I assume Pat and Dave would have changed course if positive margin isn't clearly possible.
Also, intel 3/4 and even A20 are now intermediate steps, they can't skip steps like before (10nm), but it also makes no sense to heavily ramp them, so it makes sense to leverage TSMC.
I think Arrow Lake is on A20, and if it passes muster, then Intel will be fine as that will offer very good visibility on A18 (which will ramp heavily)
In reply to your last point, Intel as an IDM succeeded by having an effective stranglehold on the wildly successful x86 CU market by having the best products after AMD cocked it up -- for quite a few years they had both leading technology and products and were absolutely coining money. This is all fine so long as you stay in the lead and don't drop the ball. But it's worth noting that everything Intel did apart from x86 CPUs failed, because they didn't have either a technology or product lead, or were too expensive -- or just plain screwed up. The list is very long -- WiMax, cellular modems, NVRAM to name but a few.
The problem came with the 10nm process disaster and the decision of AMD to go chiplet-based while Intel stuck to monolithic solutions -- the two combined meant that Intel bled market share (and profits) and fell behind on process.
Getting back from where they are now is going to be really hard from both points of view -- and separating out the foundry and process divisions makes the kind of IDM cross-subsidy referred to above impossible, which makes things even harder for them.
Small but positive margin is not enough -- to invest in and develop the next technology you need big gross margins (typically around 50%) but also big enough volume to fill and pay for the fabs.
I don't understand how anybody nowadays thinks that a (standalone) bleeding-edge fab can succeed on thin margins, it's simply not possible -- Samsung are only being kept alive by internal business from the other BUs in the conglomerate, which is also willing to spend a lot to keep the foundry going. Intel can't play that cross-subsidy card any more now IF has been separated out, even if it's not illegal it would totally piss off all their other (who?) customers who found they were paying more per wafer than the Intel product group.
IF has separate accounting but isn't much different from Samsung's situation. IF does not have to sell to Intel products any lower than other customers, as long as they are selling at above cost. Every dime of margin IF makes is accretive to Intel Corporation. Until IF goes IPO, the corporation margin is what matters.
But future fab and process development has to be paid for by IF, not Intel Corp, because that's how the businesses are structured.
You can't keep on pretending that Intel is still an IDM -- when what you said might be true, as it is for Samsung -- when the separation specifically means that they're not any more...