Speculation: Intel will become fabless

[Gideon]

Ian says that 16FF is theoretically 28.2 whereas Intel 14 is 37.5. And as I said 10FF was just a stopgap which is probally not being produced anymore.

Yeah sorry read that incorrectly in a haste, but nevertheless disqualifying TSMC's 16/12nm but keeping Intel's 14nm seems kinda arbitrary.

If one considers Intel's 14nm competitive (ignoring power draw in that case) then IMO TSMCs 12nm is so as well (which is essentially 16nm with higher rectangle limit). Let's not forget that it was considered a better node than Samsungs 14nm at least for Nvidia GPUs which scaled better on it.

We can take TU102 as an example. RTX 2080 Ti has RTX 3070 level performance with a 250 W TDP and has some units disabled, RTX Titan is a 280W TDP (with ~8% more performance) while there is decent headroom compared to RTX 3090 and RTX 3080 which use 350W and 320W respectively.

Re-release TU102 it with all units enabled, 12GB of 15.5 Gbps GDDR6 (2080 SUPER has it), the current gen coolers and 350W TDP and it will at least equal if not beat the upcoming 3070 Ti while being cheaper to produce than the 3080. Highly competitive in my book.

 

[Vattila]

Who would Intel make it a joint venture with?

GlobalFoundries still seems to be a good match to me (as I wrote in the OP back in 2018), with IBM aboard as a research partner. The three of them could make a joint foundry funded by USA and EU government initiatives. GlobalFoundries has established a good business in speciality nodes, but needs a plan for the future and the cutting edge, EUV, new transistor structures (nanosheets), chiplets, 2.5D/3D stacking, packaging etc. They have presence and history in USA (through the AMD heritage), and a big campus established in upstate New York, where they also are close to IBM research and former manufacturing facilities (which they acquired). They also have presence in Europe (Dresden, Germany), which Intel says it wants to serve, playing into EU semiconductor funding initiatives.

Seems to me they could make a strong western manufacturing giant together, to counter-balance the dominant foundries in Asia.

 

[moinmoin]

GlobalFoundries still seems to be a good match to me

TAM wise maybe. But the only case for spinning off IFS for Intel is to reduce cost for overall Intel, and there GloFo is in no position at all to help financially. Mubadala wants to IPO GloFo until next year at the latest. So an independent IFS or IFS+GloFo would still be completely dependent on Intel's money, at which point Intel may as well keep all the control as well.

 

[LightningZ71]

I keep Intel's (out of necessity) well refined 14nm node in the discussion for the simple reason that it is still producing processors that perform as well as competing leading edge nodes, such as TSMC N7. The three current poster children are the various leaked 5750g benchmarks, the 5800x (which has to use an older node for it's I/O to keep the use of costly N7 wafer capacity low), and Rocket Lake. Those products all benchmark quite close to each other. Yes, when pushed to the limits, the Intel parts do consume considerably more power, but, as we have heard from the Intel camp, for desktop chips, that doesn't matter. I won't argue that it is at the end of it's leading edge life, but it's hardly not in the discussion TODAY.

I exclude 10ff, as well as 16/12ff, simply because 10ff was not heavily used for high performance parts, and 16/12 don't reach the same performance/density ratio as Intel's 14nm nodes. You can get them rather dense, but you don't see them clock that high and the chew through power to do so. When you run them with more relaxed rules, then density isn't
competitive.

When people point out the mass numbers of SoCs produced for mobile and tablet, they conveniently forget that those die sizes are usually substantially smaller than the ones found in desktop, x86 mobile, and x86 server products. When your die is a quarter to a half the total surface area, you can get many more per wafer and yield numbers are easier to get higher. Case in point is that the high volume, lower and lower middle market phones that were sold over the last few years that offered only clusters of A53/55 cores in 4 and 6 total count were tiny compared to the higher end chips, yet they are captured in those same metrics.

 

[Vattila]

Having digested this news, I am actually surprisingly supportive of the moves Intel CEO Pat Gelsinger is making with regards to Intel Foundry Services. Although they are not quite there yet, they are moving towards a business model with much greater independence between product design and manufacture. As with Samsung and their manufacturing division, that independence is crucial for competitors to embrace IFS as a manufacturing option.

The demise of Intel manufacture would not be a good outcome — not for Intel, nor for the overall state of the semiconductor industry. Even as an AMD shareholder, it is not something I would want to see. I strongly support government appreciation of science and technology, and in particular, the initiatives to strengthen the semiconductor industry in the home markets. Cost-saving outsourcing to cheaper parts of the world has done good, but has gone too far. There needs to be rebalancing. As such, Intel's manufacturing division is a big part of the current capability in the west.

So, just as Gelsinger uttered the statement that Intel may make use of GlobalFoundries (AMD's former facilities), I want to see Intel get to a position where AMD CEO Lisa Su could reciprocate that statement by considering the use of Intel Foundry Services — or better yet, the use of my proposed joint venture between Intel, GlobalFoundries and IBM, spun off from IFS at some point in the future.

We are not at that point yet. But I can see us moving there.

 

[Vattila]

[longwinded argument about AMD's growth limitations due to wafer supply constraints]

How much wafer supply and growth do we need to see at AMD before you find it decent enough?

 

[moinmoin]

Yes, when pushed to the limits, the Intel parts do consume considerably more power, but, as we have heard from the Intel camp, for desktop chips, that doesn't matter. I won't argue that it is at the end of it's leading edge life, but it's hardly not in the discussion TODAY.

As we speak Intel is scrambling to move more of its mobile and server business to 10nm, and Intel's resulting bleeding edge products are still only competitive with AMD at best. So yes, it very much does matter today already.

 

[Stuka87]

I keep Intel's (out of necessity) well refined 14nm node in the discussion for the simple reason that it is still producing processors that perform as well as competing leading edge nodes, such as TSMC N7.

N7 isn't bleeding edge. 5nm is. And while Intel's 14nm may match on the performance side in some use cases, its not even remotely close on the performance per watt side. All of us here know that it takes a lot more to cool a 125W Intel than it does to cool to 65-95W AMD. Especially since Intel's routinely draw over their rating by a larger amount.

In the chart below, the Ryzen 5900X has a TDP of 105W. The Intel 10900K has a TDP of 125W. The Ryzen is using 31W over that 105W, but the Intel (with fewer cores) is using 82W over its 125W rating (non-OC) for a total of 207W, just for the CPU (this isn't system power).



Gaming performance, Intel's high end 10900K is losing to mid range AMDs:


And in non-gaming workloads, things get worse off for Intel:


In short, 14nm is not competitive today. Much less a year down the road when they will still be selling 14nm chips mixed in with tweaked 10nm chips. While AMD will be transitioned to 5nm. Intel *MUST* do something soon. It will be years before these new fabs come on line. And even when they do, Intel needs to figure out their issues with smaller nodes, or they will remain in the same situation that they are in now.

 

[IntelUser2000]

In short, 14nm is not competitive today. Much less a year down the road when they will still be selling 14nm chips mixed in with tweaked 10nm chips.

Yup. 14nm only made sense because of being optimized to the edge for 5 years they were able to take the clocks beyond any reasonable sanity.

10nm is way better, even the one in Icelake.

Some of the faults may have to do with their inefficient core. Tigerlake uses about the same power but you can fit twice the number of Zen 2 cores.

 

[LikeLinus]

"In short, 14nm is not competitive today. " That is true as far as performance or TDP, but the issue is everyone here is using benchmarks and not looking at real world usage. Intel 14nm has lasted far longer than anyone expected. The fact that Intel still has around 80% of the market and even gained some market share in 2020, should tell you something. You are looking strictly at those metrics. Completely ignoring that 95% of the population has no freaking clue what you are talking about. Intel is a significant brand name.

Newsroom Home

Intel Newsroom home page

newsroom.intel.com

I've said this 2 years ago when everyone was saying that Intel was doomed and the end was near (and I got a lot of crap for it). Intel is a brand name that people trust. AMD? If I asked 100 people who AMD was....

The point is, the Intel 14nm chips will perform pretty much exactly the same as the AMD chips in everyday task. Even if they didn't, people don't see the small advantage of saving a document. The reality is the marketing. AMD just doesn't have that.

Seriously, how many people do you know go out and say "what is the TDP or 3Dmark bench of this CPU?" LOL

AMD needs high performance CPUs, no doubt. But AMD also needs something that they just don't currently have. Plus, they simply can't supply a vendor like Dell with the amount of CPUs they need. This is why there is a huge shortage of PS5, Xbox and AMD 5000. Great CPUs, but their manufacture doesn't have the capacity that Intel does, at this point.

FYI: I have an AMD 5950x and 5800x for my editing systems. My servers are still Xeon. This post isn't bias, it's just trying to point out that Marketing is a major factor in success.

 

[DrMrLordX]

The fact that Intel still has around 80% of the market and even gained some market share in 2020, should tell you something.

The market expanded, and 14nm CPUs aren't completely useless.

 

[LikeLinus]

The market expanded, and 14nm CPUs aren't completely useless.

I completely agree. I use the AMD 5000 series but I'm an enthusiast. I'm not sure some people here understand that 14nm (no matter the power consumption) is still widely used. We really want both companies to win and stay competitive.

 

[Thunder 57]

I've said this 2 years ago when everyone was saying that Intel was doomed and the end was near...

I'm so tired of that strawman. I've seen exactly zero people say that. Plenty have said that Intel is in trouble or not doing all that great, but I've heard no one saying they were flat out going under. One would have to be a dumbtard to think that. Or a troll looking to fan the flames.

 

[LikeLinus]

I'm so tired of that strawman. I've seen exactly zero people say that. Plenty have said that Intel is in trouble or not doing all that great, but I've heard no one saying they were flat out going under. One would have to be a dumbtard to think that. Or a troll looking to fan the flames.

I'm not going through multiple years of post. Since you've never seen anyone talk about the death of Intel.

But I guess I'm making it up, right? I'm not the only person who has seen these comments.

"I would hope rumors of Intel's death are greatly exaggerated." IEC must be a dumbtard!

 

[IntelUser2000]

The market expanded, and 14nm CPUs aren't completely useless.

Only reason Comet Lake has been relevant was because of "muh gaming", and Zen 3 took that away from them. Not even Rocketlake will save them here.

Yes their brand clout will last for a while, but it will expire. If you are buying something you still want the better one, whether you need it or not.

Also the shortages are preventing AMD from taking more share. Especially in the lower end Intel chips are cheaper and widely available. Try to build a cheap system using LGA 1151/1200 or AM4. With AM4 you have to go for pre-Ryzen chips which suck.

 

[DrMrLordX]

"I would hope rumors of Intel's death are greatly exaggerated." IEC must be a dumbtard!

Oh please. Even you ought to know that @IEC wasn't predicting Intel's impending doom with that comment. He was responding to a thread about one of Charlie's rants. Though it is amusing to see predictions of Q4 2020 for Ice Lake-Sp in that thread. Intel wishes they could have hit that target.

Intel will continue to survive so long as someone, somewhere has a reason to buy their existing 14nm silicon. With Cooper Lake and Cascade Lake-AP dead and Ice Lake-SP shipping in tiny quantities, Intel is stuck selling old product to the server sector, and AMD can't replace all those Intel 14nm installations in one go (yet). Over time, though, I expect most Cascade Lake-SP installations to be retired for something that is not Intel-based. Sapphire Rapids still isn't here, and AMD can eventually saturate the market with silicon that is objectively better than Cascade Lake-SP in every conceivable way.

Also the shortages are preventing AMD from taking more share.

More or less.

 

[IntelUser2000]

Over time, though, I expect most Cascade Lake-SP installations to be retired for something that is not Intel-based. Sapphire Rapids still isn't here, and AMD can eventually saturate the market with silicon that is objectively better than Cascade Lake-SP in every conceivable way.

I can't even say Granite Rapids will bring back competitive server chips. AMD is simply on fire. The Nvidia driver overhead issue is also being known at a time when we're at a shortage and AMD's resurgence is happening in graphics too.

Zen 4 with AVX-512, 96 cores, 5nm, and 30% higher perf/clock will make Sapphire Rapids look like Icelake-SP.

Intel wants to coast with Xe for 2-3 years while next year AMD's iGPU moves from Vega to RDNA2, with potentially more CUs. Getting curbstomped in every category.

 

[Stuka87]

The fact that Intel still has around 80% of the market and even gained some market share in 2020, should tell you something.

Yes, when you strong man OEMs into only using your equipment, its easier to hold markets share. OEMs are specifically not allowing higher end AMD systems, and this all traces back to Intel's power over OEMs.

We know there is a market for AMD systems, and we know they perform better. But good luck trying to buy an AMD workstation, because they either don't exist, or they are being put into non-competitive chassis. AMD has been able to make some inroads on the mobile side, but even that is going slower than it should.

 

[DrMrLordX]

I can't even say Granite Rapids will bring back competitive server chips.

Even if it did, how many will Intel be able to sell per quarter? How many installations of Cascade Lake-SP are there out there now? Right now, anyone looking to upgrade is either:

1). Going to find out that Intel doesn't have a better product on the open market (and only has a handful of Ice Lake-SP and Cooper Lake shipped to select clients)
2). Going to find better products from AMD that are probably sold out already, or are on backorder

It's pretty staggering to think that Intel hasn't been able to volume launch a significant upgrades since April 2019, and that's only if you consider Cascade Lake-SP to have been a significant upgrade from Skylake-SP.

In any case, Intel will have to ship an enormous amount of Sapphire Rapids and/or Granite Rapids to permit upgrades of that installed base. AMD has already demonstrated that they're in no hurry (either by choice or by limitation) to saturate the market with their own hardware. If/when Intel proves unable to do so, it'll be clear that many shops will be stuck on Cascade Lake-SP for awhile.

 

[moinmoin]

Ooof, seems like Intel hid bad news within the whole IDM 2.0 announcement package. Via:

"We have customers testing ‘Sapphire Rapids’ now, and we’ll look to reach production around the end of the year, ramping in the first half of 2022."

https://newsroom.intel.com/wp-content/uploads/sites/11/2021/03/CEO-Gelsinger-quotes-57285720.pdf

To put a positive spin on it: Looks like Ice Lake SP gets its fair chance to find at least some consumers now before Sapphire Rapids arrive. Ouch...

 

[Stuka87]

We already knew that Sapphire Rapids was not expected until 2022 at some point. If anything them actually giving a tentative time frame of the end of 2021 is positive news. Rather than remaining unknown.

 

[biostud]

More competition is better. I don't want TSMC to be the only facility to produce bleeding edge CPUs.

 

[moinmoin]

We already knew that Sapphire Rapids was not expected until 2022 at some point. If anything them actually giving a tentative time frame of the end of 2021 is positive news. Rather than remaining unknown.

I was told Ice Lake SP would only have a short shelf life due to Sapphire Rapids coming in 2021 as well and that Intel's 10nm is a solved problem now. Is Alder Lake S still for 2021?

 

[Stuka87]

I was told Ice Lake SP would only have a short shelf life due to Sapphire Rapids coming in 2021 as well and that Intel's 10nm is a solved problem now. Is Alder Lake S still for 2021?

Alder Lake S is also expected very late this year into early next year.

 

[jpiniero]

I was told Ice Lake SP would only have a short shelf life due to Sapphire Rapids coming in 2021

Rumors are Sapphire Rapids will only be for the high end and that Ice Lake Server is going to stay on the market for awhile.