Howdy, Stranger!
It looks like you're new here. If you want to get involved, click one of these buttons!
Quick Links
Intel radically changes its foundry plans
Intel made a number of announcements yesterday that together amount to a dramatic change of direction for the company. The changes basically amount to betting the company on their new direction being better. Which may sound dramatic, but without changes, Intel was on the path to irrelevance and ultimately bankruptcy. So what are the changes?
1) Intel will no longer exclusively manufacture their own products. While they had previously been inching away from that stance, they made it official in a huge way yesterday. The next Intel CPU you buy might have been manufactured by TSMC--or Samsung, Global Foundries, or even UMC. Well, probably not GloFo or UMC for CPUs, but Intel did say that some products will be manufactured there. Maybe they'll be chipsets, I/O dies like what AMD builds at GloFo, or random other things like ethernet controllers.
2) Intel will no longer exclusively manufacture their own products. This might look like the same as (1), but it's in the other sense. Intel foundries will be open to anyone who wants to manufacture anything, just like TSMC and Samsung are. Well, not quite absolutely anyone, as they're not going to violate laws by helping Iran or North Korea manufacture things. But while in the past, they've been extremely restrictive about who else can manufacture anything at Intel fabs, now they're going to open it up and accept most companies who want to manufacture products there. A future GPU that you buy might be fabricated by Intel, even if it says GeForce or Radeon on the box.
3) Intel will open up their designs and IP for use by others, much like ARM does now. It's not clear just how far they'll go in this direction, as they may be unwilling to, for example, let Nvidia build consumer SoCs with the latest and greatest Intel CPU cores to compete with Intel's Core lineup (and AMD's Ryzen Mobile). But they do very much intend to let some other companies build chips using Intel x86 CPU cores at Intel fabs.
4) Intel is promising a return to tick-tock, with new process nodes coming a lot more frequently. How do you do that? By making smaller changes from one process node to the next. Their 10 nm node tried to make far too large of a jump all at once, and it hurt the company badly.
5) Intel CPUs are going chiplets, like what AMD has done, starting (at least for consumer use) with Meteor Lake on 7 nm in 2023. Intel calls the approach "tiles" rather than chiplets, though I expect them to have some difficulty if they really want to dislodge the chiplet nomenclature. There are some technical differences between what Intel is going to do and what AMD CPUs are doing, but they will fundamentally be piecing together several smaller chips to make products instead of sticking with monolithic dies.
Overall, I'd say that Intel is really just doing what they had to do to survive. This is far short of closing their struggling fabs entirely or even spinning it off as a separate company. But with their traditional approach, either CPU design or fabs faltering could kill the whole company. Had Intel decided to make some CPUs at TSMC years ago like AMD did, they'd probably still be competitive with AMD today.
This is not just a case of Intel's new CEO changing things for the sake of justifying his existence. Intel needed dramatic changes in order to save the company, and this might do it. Or it might not. We'll see how it works out.
1) Intel will no longer exclusively manufacture their own products. While they had previously been inching away from that stance, they made it official in a huge way yesterday. The next Intel CPU you buy might have been manufactured by TSMC--or Samsung, Global Foundries, or even UMC. Well, probably not GloFo or UMC for CPUs, but Intel did say that some products will be manufactured there. Maybe they'll be chipsets, I/O dies like what AMD builds at GloFo, or random other things like ethernet controllers.
2) Intel will no longer exclusively manufacture their own products. This might look like the same as (1), but it's in the other sense. Intel foundries will be open to anyone who wants to manufacture anything, just like TSMC and Samsung are. Well, not quite absolutely anyone, as they're not going to violate laws by helping Iran or North Korea manufacture things. But while in the past, they've been extremely restrictive about who else can manufacture anything at Intel fabs, now they're going to open it up and accept most companies who want to manufacture products there. A future GPU that you buy might be fabricated by Intel, even if it says GeForce or Radeon on the box.
3) Intel will open up their designs and IP for use by others, much like ARM does now. It's not clear just how far they'll go in this direction, as they may be unwilling to, for example, let Nvidia build consumer SoCs with the latest and greatest Intel CPU cores to compete with Intel's Core lineup (and AMD's Ryzen Mobile). But they do very much intend to let some other companies build chips using Intel x86 CPU cores at Intel fabs.
4) Intel is promising a return to tick-tock, with new process nodes coming a lot more frequently. How do you do that? By making smaller changes from one process node to the next. Their 10 nm node tried to make far too large of a jump all at once, and it hurt the company badly.
5) Intel CPUs are going chiplets, like what AMD has done, starting (at least for consumer use) with Meteor Lake on 7 nm in 2023. Intel calls the approach "tiles" rather than chiplets, though I expect them to have some difficulty if they really want to dislodge the chiplet nomenclature. There are some technical differences between what Intel is going to do and what AMD CPUs are doing, but they will fundamentally be piecing together several smaller chips to make products instead of sticking with monolithic dies.
Overall, I'd say that Intel is really just doing what they had to do to survive. This is far short of closing their struggling fabs entirely or even spinning it off as a separate company. But with their traditional approach, either CPU design or fabs faltering could kill the whole company. Had Intel decided to make some CPUs at TSMC years ago like AMD did, they'd probably still be competitive with AMD today.
This is not just a case of Intel's new CEO changing things for the sake of justifying his existence. Intel needed dramatic changes in order to save the company, and this might do it. Or it might not. We'll see how it works out.
Comments
------------
2024: 47 years on the Net.
Both of those are driven by US government subsidies, not market conditions. With IBM and AMD no longer having their own foundries anywhere remotely near the cutting edge and uncertainty over how much longer Intel will do so, the government is paranoid about having sensitive (e.g., military) components fabricated on foreign soil so near major rivals, especially China. The government would much rather have military equipment use computer chips fabricated on a TSMC 5 nm process node in Arizona (or anywhere else in the US, really) than on exactly the same TSMC 5 nm process node in Taiwan.
Another announcement yesterday is that Intel plans to finally move the majority of their production to 10 nm later this year. For now, most of what they have is Tiger Lake chips in fairly low volumes, most likely because the yields are terrible.
------------
2024: 47 years on the Net.
The 2023 date is when Intel expects Meteor Lake CPUs to launch for consumer ("client") use. Or at least it's when they claim to expect it. After the 10 nm debacle, it's possible that internal estimates might not match public claims. Depending on yields and fab capacity, those might be laptop-only, as with Ice Lake and Tiger Lake.
Intel's fabs are certainly well behind TSMC right now. Intel's CPU design teams still seem to be competent, but there's only so much you can do when you designed your chips according to the rules that your foundry listed, but they just can't build them. It's basically the reverse of the situation where AMD's fabs were behind Intel's for decades. AMD had some good CPU designs during that time, but the only time that they were able to make CPUs better than Intel's were when Intel had an awful design, mainly Netburst (Pentium 4 and D). It wasn't enough for AMD to merely have a better CPU design than Intel. To produce a better CPU overall, the CPU design had to be better by a large enough margin to overcome Intel's fab advantage.
It's also possible that Intel will build some cutting edge CPUs at TSMC, as in point (1). If they do that, then AMD's current fab advantage disappears. They could do what Apple did for at least one generation and design a chip at multiple foundries (in Apple's case, both Samsung and TSMC), and then actually build it at whichever foundry can deliver something that works (or both). That adds a lot of development cost, certainly, but that added cost is a lot less bad than what Intel is going through right now where they just can't build any competitive products.
But there's no guarantee that Intel will remain behind forever, just as they didn't remain ahead forever. Intel can buy the same foundry equipment from ASML and other companies that TSMC can buy, and it's just a question of how well you can make everything work.
It's also very possible that TSMC will fall on their face for a generation or two. Their 10 nm and 20 nm process nodes were rather bad, for example. That's why AMD and Nvidia mostly ignored them, and other customers moved off of those nodes quickly as soon as the next one was ready.
But for just purchasing manufacturing like AMD and NVidia do you can move your manufacturing to a new node relatively quickly. It's likely to take more than 6 months but less than a year, if you decide you want to use your development resources that way instead of using the same resources to develop a completely new product.
Things can be simpler if your design isn't that demanding and you're using an older, very mature process node. If you're starting on a brand new chip today and TSMC's 28 nm node is plenty good enough, for example, then you could probably reach a commercial launch significantly faster than if you tried to build exactly the same thing on TSMC's 5 nm node. Well, at least if you assume that the foundries have plenty of capacity to build whatever you want on all nodes, which definitely isn't the case today. But it's still going to take quite a while.
No, a Core i7-10700K is not just as fast as a Ryzen 7 5800X. Not even close. Or at least not in typical CPU-heavy workloads for which the CPU is the bottleneck. For example, see here:
https://www.anandtech.com/show/16214/amd-zen-3-ryzen-deep-dive-review-5950x-5900x-5800x-and-5700x-tested/12
If you want to compare the 5800X to the 10700K in the 39 CPU tests, the 5800X wins 37 out of 39, and often by large margins. The 5800X often even beats the 10900K in multi-theaded tests, even though the latter has two more cores and uses nearly double the power.
In the gaming tests, at minimum settings (which are most CPU-intensive), all four of the Ryzen 5000 series CPUs beat all tested Intel CPUs (including the top of the line Core i9-10900K) in a clean sweep of all 13 games tested. In only three of the games is it even close, and the 5800X often beats the 10700K by 20% or so. At higher settings when there is a GPU bottleneck, the Intel CPUs can catch the AMD CPUs, but that's not really a CPU test anymore. For comparison, in exactly the same review, the Core i7-10700K beats even the Ryzen 9 3950X in 12 of those 13 games.
And no, it's not a difference of 20 W. The Core i7-10700K has a PL2 power of 250 W, and you'll typically use somewhere around there in heavy CPU workloads that can push all of the cores. The Ryzen 5000 series CPUs have a max turbo power of 142 W. That's a gap of over 100 W. You can decide to care about that gap or not, but it is not a small one.
You seem to be focusing on TDP, but TDP isn't meaningful anymore in desktop CPUs, as basically all motherboards will ignore it. Turbo power is what matters, as that's what the CPU is going to use when you push it hard. Intel and AMD have some nominal limits on how long the CPUs will use the turbo power, and those limits may be enforced in laptops, but they allow motherboard vendors to just ignore the TDP and run max turbo forever. Basically all motherboards do that, as if you don't, then people will think your motherboard is slow and not buy it.
Even Intel's upcoming Core i9-11900K usually isn't as fast as a Ryzen 7 5800X, though it does narrow the gap quite a bit. But the power consumption and temperatures are pretty awful, and the prices probably won't be pretty, either.
The higher end Comet Lake and Rocket Lake CPUs are the most power-hungry mainstream consumer CPUs ever, and the only other CPUs to come close some old super high clocked Piledriver CPUs such as the FX-9590. (I exclude HEDT from that comparison.) Maybe you don't care about the difference between 105 W and 125 W, but 250 W is really a lot of power for a consumer CPU.
If your big complaint about the Ryzen 7 5800X is the price tag, then why not compare it to a Ryzen 5 5600X instead? That's still faster than a Core i7-10700K at gaming, and much closer to the same price, at least if you can find one at MSRP. And if you want to argue that MSRP doesn't matter when parts are out of stock, have you had a look at GPUs lately? The Ryzen 5 5600X is going to be widely available at MSRP before any GeForce 3000 or Radeon 6000 series GPUs, and gaming is the goal, you usually want a new GPU, too.
7700k @ 5ghz with GTX 1080ti = avg 28fps limited by CPU
7700k @5ghz with RTX 3090 = avg 28fps limited by CPU
Ryzen 3950 @4.5ghz with GTX 1080ti = avg 39fps jumping between CPU and GPU limited.
Ryzen 3950 @4.5ghz with RTX 3090 = avg 42fps CPU limited.
Ryzen 5900 @4.7ghz with RTX 3090 = avg 57fps GPU limited
Conclusion: the new Ryzen 5000 series are a beast when running MS Flight Simulator
"Be water my friend" - Bruce Lee
Logic, my dear, merely enables one to be wrong with great authority.
While Intel does have a significant presence in Israel, their struggles with 10 nm will ultimately cost the company tens of billions of dollars. And that's even if they get everything fixed for 7 nm, which is far from certain. If Intel's only goal were to reincorporate in Israel, they surely could find a much cheaper way to do that.
Cant give you a source I'm afraid. Do your own research, maybe youll learn something.
Its always rubbish...until it isnt.