News Intel IFS Plans Are a Strategic Mistake, Hints AMD Exec

[versed.perception]

Someone bookmark this article so in three years when Intel is fabbing AMD chips we can all rub it in his face.

Exactly. Intel spinning up a FAB business is not bad! the issue will be how they license access to the fab and if Intel follows their own rules, else the FTC and SEC will have something to say about it.

However, I do not trust Gelsinger at all and historically Intel's leadership has been rot with corruption (https://techcrunch.com/2018/01/04/a...arise-about-timing-of-intel-ceos-stock-sales/). Hell when Gelsinger was brought in as CEO and endorsed by Dan Loeb, a few of us predicted that before 2025 Intel would be either selling off Fabs or creating a Fab sub/business as a service for short term profits, while diging into that cookie jar. So we just have to now keep an eye on that cookie jar.

 

[cyrusfox]

The 2 things Intel has that TSMC does not currently is Backside power delivery/ power all around, and packaging(3d stack foveros, glass substrate, etc...).

View: https://www.youtube.com/watch?v=HGCaGIxIa2k&t=2h23m30s


If IFS is providing this to others, can't afford not to look at Intel

 

[thestryker]

Given Intel's manufacturing operations which AMD lacks, it's a false comparison. If you add TSMC's R&D spending to AMD's, I think the comparison would look entirely different.

Comparing R&D spending to total revenue isn't a false comparison at all it shows the dramatic difference in company size.

Intel R&D spending is more than both combined by a lot. TSMC adjusted to USD (in billions) 2018-22: 2.67, 2.85, 3.41, 3.88, and 5.08. AMD same period: 1.434, 1.547, 1.983, 2.845, and 5.005.

2018: 13.54 vs 4.104
2019: 13.36 vs 4.397
2020: 13.56 vs 5.393
2021: 15.90 vs 6.725
2022: 17.53 vs 10.085

It's a question of priorities. Intel went over 2 decades of not touching their dividend scheme, and then have been doing considerable share buy-backs on top of that. All the while, doing countless layoffs and divesting from numerous businesses that often seemed like they could still go either way. If your priority is generating returns for investors, then it makes total sense that you protect profitability even if it means underinvesting in the business.

We saw the clear results of this practice with a decade of middling IPC improvements and consumer core-count stagnation, countless corners being cut on security, and repeated failures to deliver on their process roadmap. In my opinion, what's been Intel's biggest challenge isn't its fabs, but rather its investors. A series of CEOs has been unable to meet this challenge, but Pat may turn out to be successful enough. He already cut the dividend and didn't lose his job over it.

Intel had all engineers at the helm until the mid 00s and that's when the prioritization on revenue started taking over everything. Investors never care about company health so long as the money is rolling in so it's up to having good management. I'm hopeful Gelsinger continues to be successful and is replaced by another with engineering background.

 

[JayNor]

intel spent over $10B in R&D each year since 2012. It is currently at about $17B for ttm.

they continued to work on their euv process in parallel with fixing their duv multi-patterning issues.

their intel-7 duv process is producing 6GHz cpus. where are AMD's 6GHz cpus?

 

[JayNor]

I recall Intel cranking out something like 70 million chips for Apple in the last year they had that 4g business. It goes to show that a contract can guarantee a customer priority over Intel's internal fab needs.

 

[bit_user]

their intel-7 duv process is producing 6GHz cpus. where are AMD's 6GHz cpus?

Oh, I thought you knew better than that.

The race to peak clock speed is a bit of a false pursuit, in that it only really helps you in the gaming desktop market, which is rather niche. It doesn't help you in laptops or servers, which are the main bread-and-butter for these companies.

And it's not like it doesn't come at a cost, as we probably all remember from Intel's disastrous NetBurst architecture of the Pentium 4 - chasing higher clock speeds forces you to increase pipeline lengths, which both sacrifices efficiency and increases your branch-mispredict penalty. So, it's not a harmless endeavor. Instead, AMD smartly focused on efficiency*, rather than getting sidetracked on the race for peak clock speeds.

* Before someone replies with data of Ryzen 7000 operating at the edge of its envelope, I want to clarify that I'm talking about efficiency within the frequency range that servers and laptops actually use. For reference, the flagship Genoa model (9654) has a base/turbo clock of just 2.4/3.7 GHz. The fastest mainstream Genoa base clock is 3.1 GHz (9554; 64c) and the fastest mainstream Genoa turbo is just 4.15 GHz (9254; 24c). They have some HPC variants with base & turbo clocks as high as 4.1/4.4 GHz (9174F; 16c).

Source: https://www.tomshardware.com/review...6-cores-zen-4-and-5nm-disrupt-the-data-center
​

If you want to see how this worked out for them, the last page of Phoronix' Genoa-X review also has performance & power data for EPYC 9654 and Xeon 9480.

AMD EPYC 9684X Genoa-X Provides Incredible HPC Performance Review - Phoronix

www.phoronix.com

There, we see the EPYC 9654 (x2) had an average power consumption of 446.57 W and a Geomean score of 299.36, yielding a perf/W of 0.670 points/W.

By contrast, the Xeon 9480 (x2) had an average power consumption of 537.85 W and a Geomean score of 143.43, yielding a perf/W of 0.267.

So, in this benchmark suite, that AMD processor was 2.51x as efficient as its Intel counterpart. That's what I mean, by a focus on efficiency. That's why it's what really matters - not the desktop performance horse race we tend to focus on.

 

[bit_user]

I recall Intel cranking out something like 70 million chips for Apple in the last year they had that 4g business. It goes to show that a contract can guarantee a customer priority over Intel's internal fab needs.

Supply contracts all have penalty clauses, because (as we've seen), there's always the possibility of unforeseen events which cause production delays or shortfalls.

What @versed.perception was talking about is Intel making a conscious decision to pay the penalties associated with under-supplying customers, in order to prioritize their own internal production needs. As such a move could be catastrophic for IFS' future business, I don't expect that will happen. Also, their separation within Intel should reduce their incentives to do so.

 

[bit_user]

Intel R&D spending is more than both combined by a lot. TSMC adjusted to USD (in billions) 2018-22: 2.67, 2.85, 3.41, 3.88, and 5.08. AMD same period: 1.434, 1.547, 1.983, 2.845, and 5.005.

2018: 13.54 vs 4.104
2019: 13.36 vs 4.397
2020: 13.56 vs 5.393
2021: 15.90 vs 6.725
2022: 17.53 vs 10.085

Thanks for digging up that data! It's interesting.

However, there are still some Apples vs. Oranges, underneath it. For instance Intel owns Altera & other businesses. In fact, I'll bet the 2022 numbers still mostly include their storage business unit, since that's probably from FY22, not CY22 and the sale of that unit happened sometime probably towards the end of their FY.

Meanwhile, AMD's acquisition of Xylinx didn't close until 2022 and probably explains the sudden jump in their R&D spending. Perhaps FY23 will be the first full FY after the acquisition, in which case we'll see it continue to jump.

Anyway, my point is that just looking at R&D spending over disparate product lines doesn't tell you how their spending compares within the same class of products. Factoring in TSMC gets us closer to a like-for-like comparison, but still doesn't get us 100% of the way there.

Intel had all engineers at the helm until the mid 00s and that's when the prioritization on revenue started taking over everything.

Yes, I have some dim recollection of Otellini's appointment as CEO. His timing probably couldn't have been better, since he took the helm shortly before Core 2 launched. He had a chance both to do some cost-cutting (i.e. layoffs) and then ride high on the success of that new generation and the couple which followed.

 

[thestryker]

AMD smartly focused on efficiency, rather than getting sidetracked on the race for peak clock speeds.

I do think a fair argument can be made that with Zen 4 they did fall into this trap on desktop with the X SKUs. The architecture itself is absolutely fantastic efficiency wise, but they're not being run that way.

Anyway, my point is that just looking at R&D spending over disparate product lines doesn't tell you how their spending compares within the same class of products. Factoring in TSMC gets us closer to a like-for-like comparison, but still doesn't get us 100% of the way there.

Yeah Intel is simply a much bigger company with way more business groups.

I hadn't thought about it before, but I bet Intel's R&D spending on graphics is probably higher on a 1:1 comparison the last 5 or so years as they were basically building from scratch.

 

[Ogotai]

where are AMD's 6GHz cpus?

intel is the one that needs the higher clocks fo get tne performance it needs to compete with amd. bit user, kind of showed that with his comparison between epyc and xeon ..

 

[bit_user]

I do think a fair argument can be made that with Zen 4 they did fall into this trap on desktop with the X SKUs. The architecture itself is absolutely fantastic efficiency wise, but they're not being run that way.

What happened is they took an architecture which wasn't designed to clock high, and then jacked up the clock speeds to chase the last few % of performance. It's well beyond the point of diminishing returns. It's misleading to take its efficiency data, at those extreme clocks, and conclude that the architecture is inefficient. Zen 4 is quite efficient at the lower clock speeds it was designed to run at.

Sadly, I think Chips & Cheese hasn't published any frequency-scaling data on Zen 4. So, we can only look at power-limited SKUs, or data from people limiting power consumption in BIOS, to draw any conclusions.

Yeah Intel is simply a much bigger company with way more business groups.

Heh, it's getting to be fewer and fewer. They're starting to resemble AMD, in terms of what markets they play in (i.e. besides IFS). Before long, they'll have spun out Altera, making them potentially even more focused than AMD! I guess Intel does have Habana, while AMD is trying to cover its AI bases with a combination of Xylinx' IP and their GPUs. So, AMD is getting more from Xylinx than Intel will lose by spinning out Altera.

I hadn't thought about it before, but I bet Intel's R&D spending on graphics is probably higher on a 1:1 comparison the last 5 or so years as they were basically building from scratch.

+1. Not only that, but they had like 4 architectures: Xe-LP, Xe-HP, Xe-HPG, and Xe-HPC. I think I've got that right. They killed off Xe-HP, so it's now down to 3.

 

[bit_user]

intel is the one that needs the higher clocks fo get tne performance it needs to compete with amd.

My take on the situation is that Intel basically said: "okay, we have the E-cores to address efficiency (both power & area), so that unleashes us to optimize the P-cores for single-thread performance".

Except, when their P-cores are still the mainstay of their Xeon line, you really can't stop caring about either power-efficiency or area-efficiency, because then AMD and ARM will run circles around you. We'll see how well Sierra Forest fixes this situation.

 

[JayNor]

I believe Sierra Forest will support MCR DIMMs and the improved DSA and CXL 2.0. Looks like ARM and AMD will have troubles matching it... PG also held a 288 core version.

 

[bit_user]

I believe Sierra Forest will support MCR DIMMs and the improved DSA and CXL 2.0. Looks like ARM and AMD will have troubles matching it... PG also held a 288 core version.

Built on from a new E-core, on Intel 3, it could indeed take the efficiency crown from Siena. Although, probably just 144 cores won't be enough to beat the 128 core Siena on performance. The dual-die (up to 288 core) version will probably be required for that.

Even so, it'd be interesting to know how many Siena customers are using AVX-512, because that could be an area where Sierra Forest still provides no answer. And not long after, AMD will be launching Zen 5c on TSMC N3E.

So, Intel is not just going to walk away with the efficiency crown. It will likely have a fight on its hands.

 

[jkflipflop98]

Exactly. Intel spinning up a FAB business is not bad! the issue will be how they license access to the fab and if Intel follows their own rules, else the FTC and SEC will have something to say about it.

However, I do not trust Gelsinger at all and historically Intel's leadership has been rot with corruption (https://techcrunch.com/2018/01/04/a...arise-about-timing-of-intel-ceos-stock-sales/). Hell when Gelsinger was brought in as CEO and endorsed by Dan Loeb, a few of us predicted that before 2025 Intel would be either selling off Fabs or creating a Fab sub/business as a service for short term profits, while diging into that cookie jar. So we just have to now keep an eye on that cookie jar.

I'm not in the business part of the business. . . but it seems to me the new Ohio superfactory is intended for only fabbing chips for external customers. I don't think there's going to be much cross-over once that facility is fully online. I could possibly be wrong on this, but from what I see - no, IFS isn't going to be spun off, and no, there won't be competition between external customers and internal customers. No one can pump out devices like Intel can. And with all this new capacity coming online I don't see that being an issue in the future.

 

[rootami]

The only way for intel to survive is to be in the foundry business, almost every big CSP is designing their own chips. Many can design chip but only a few can manufacture it. Intel certainly in the right track.

 

[bit_user]

The only way for intel to survive is to be in the foundry business, almost every big CSP is designing their own chips. Many can design chip but only a few can manufacture it. Intel certainly in the right track.

Advanced semiconductor manufacturing is absolutely a key business, for Intel. Pat clearly understands this. The only real question is whether it belongs under the same corporate entity & ticker symbol as the IP business.

As for CSPs creating their own chips, they also can't do that without licensing someone else's IP.

 

[Co BIY]

I'll say right now I think that could happen. Even so, I've been saying for a while that Intel will probably spin off IFS.

Who would buy it ? Who could afford it ?

 

[bit_user]

Who would buy it ? Who could afford it ?

IPO it - spin it out as its own company.

That sort of thing happens all the time. Existing shareholders get equity in the new entity.

 

[Co BIY]

IPO it - spin it out as its own company.

That sort of thing happens all the time. Existing shareholders get equity in the new entity.

That would be a sign of great success in this case and not comparable to discarding a dead end. I could see a sale to a consortium of large customers who all get a large stake at some point. Similar to how ASML sold a large stake to Intel at one point.

I think IFS would need to be very successful to the point that the design/sell side of Intel was holding it back.

 

[bit_user]

I think IFS would need to be very successful to the point that the design/sell side of Intel was holding it back.

I think it's the other way around. IFS is holding back the rest of Intel, by being capital-intensive and having comparatively lower margins (as manufacturing usually does).

I think Intel wants to at least be in a position to spinoff IFS, to "unlock value" in each part (but mostly the design side).

 

[TerryLaze]

I think it's the other way around. IFS is holding back the rest of Intel, by being capital-intensive and having comparatively lower margins (as manufacturing usually does).

Comparative to what?!
Your own manufacturing doesn't have margins, it has/is manufacturing costs.
The only comparable thing would be intel having to buy wafers from tsmc or someone reducing their manufacturing down to 20% maybe, the same level (in marketshare) amd is at now, and intel would probably get even less than that, at least for several years.
Intel couldn't get the same amount of wafers from a different source for all the money in the world, so in that regard "the margins" for their own IFS are super.

 

[bit_user]

Comparative to what?!

Look at fabless semiconductor companies. They have economics like that of a software company. Big margins and relatively few assets to take a depreciation charge on.

Intel couldn't get the same amount of wafers from a different source for all the money in the world,

It's not like IFS goes away - Intel could continue using them for a portion of their chip manufacturing, just as they are today.

The real value of being fabless probably comes when the economy is down. You don't post huge losses, due to having to maintain costly fabs that are underutilized. Sure, if you have to cancel some wafer orders, you take a small hit (maybe 10%), but the burden of having idle capacity falls on whoever you're using for contract manufacturing.

 

[TerryLaze]

Look at fabless semiconductor companies. They have economics like that of a software company. Big margins and relatively few assets to take a depreciation charge on.

You look at them...both arm and amd are doing terrible net margins, apple is doing around 25% which is about the historical average that intel did before they started building out fabs. Nvidia had around 25-30% mostly since 2017, with some weaker quarters, only shooting up the last couple of months with the whole AI craze.

 

[thestryker]

Look at fabless semiconductor companies. They have economics like that of a software company. Big margins and relatively few assets to take a depreciation charge on.

You hit the nail on the head with regards to depreciation and the other half is capital expenditure. Fabless companies don't ever have huge capital expenditures crop up every time a fab needs to be built or repurposed. This is why a certain group of investors wants IFS spun off.

Of course I think spinning off IFS would likely be a wreck for the industry unless there are serious fabrication guarantees. One just has to look at Global Foundries to see what happens when investment groups get a hold of a capital intensive business. The best way to knock down the capital costs is not pushing new nodes.

You look at them...both arm and amd are doing terrible net margins, apple is doing around 25% which is about the historical average that intel did before they started building out fabs. Nvidia had around 25-30% mostly since 2017, with some weaker quarters, only shooting up the last couple of months with the whole AI craze.

nvidia has had market leading gross margins for the last 6 or so years. They're the company that every investor wants every other to be (until they get knocked back down of course). AMD has gotten significantly better, but didn't pass Intel until more recently as Gelsinger has been driving hard on fabrication. It's not about rationality with regards to health to the business for investors but rather quarterly profits.