Intel's Process Roadmap to 2025: with 4nm, 3nm, 20A and 18A?

You hit the nail.

Potential fab clients look now at Intel, and keep guessing what's been going on with both their 10nm process, and their fab offering for the last 5 years.

Now people keep seeing Intel setting ambitious new targets, and keeping missing them.

Would they bet their tapeout on a fab like this now? No.

The least worst thing Intel can do for itself now is to disclose what in the world was actually happening which led to the current situation.

They long missed the window when their leading edge offer can capture first few clients ready to pay any premium for it.

Sony and Microsoft consoles are running x86 hardware so demand is certainly not just driven by COVID-19, people would want those machines regardless. There is no current consumer option to pair an ARM processor with a high end graphic card. I don't know what the future holds but it's pretty obvious why people want x86 at the moment.

These are the 10nm and 7nm processes. They're just retconning the names. 10nm+ is Intel 7, 7nm is now Intel 4, 7nm+ is Intel 3, 5nm is Intel 20A.

On the one hand, x86 is (eventually) going the way of the dodo for some of the biggest datacenter customers, which is a large part of the reason that Intel has once again opened itself up to the prospect of foundry.

On the other hand, x86 is so ubiquitous for PC that it would take a sea change for Apple Silicon to take over. Users and businesses are entrenched and this isn't the first time that Apple has run a different architecture from the rest of the PC crowd. I, for one, prefer my x86 MBP to the newfangled M1 because of RAM, compatibility, external display support, and number of ports. These limitations are all a function of implementation, not inherent to Apple Silicon, but they influence my preferences nonetheless.

COVID-19 has changed the way that work is balanced against everything else in our lives, and I view its impact on the chip market as demand creation rather than a demand spike. We will never go back 100% to the way that things were prior to the pandemic; big shocks all have that in common. The hybrid model of work from home as a complement to office work means that people/companies will need to buy and maintain more hardware.

I share your hope that this doesn't turn out the way the delayed nodes did. Intel's 10nm and 7nm setbacks started in the Krzanich days during the doldrums of the PC glut. People were mobile, phone innovations were booming, and no substantial PC advancements came to the fore during that time. I think we can give Intel some benefit of the doubt and assume that they aren't dumb enough to repeat the error of announcing a road map for which there is no road map...especially because the tick-tock model is already a thing of the past and exerts no expectations nowadays.

If architecture doesn't matter, why on earth would you think x86 is a negative. The most ARM can ever hope for is that people don't care or notice it's ARM.

> I'm still left wondering why most people would want / need x86 when Apple's M1 chips clearly demonstrate that ARM-based solutions are not just sufficient, but potentially superior.

Simple, I use Linux and MacOS cannot run the software that I need to use on a daily basis. The most powerful Apple Silicon you can buy right now still can't quite beat my 2014 beater-PC in raw performance, which also doesn't exactly instill confidence in it.

Retail PC sales are only a small portion of the x86 market. The vast majority of corporate PCs will continue to use x86 in the foreseeable future for compatibility reasons. On the datacenter / cloud side providers are trying to make inroads with ARM, such as Amazon's Graviton, but are meeting similar resistance.

Additionally we have yet to see x86 on the TSMC 5nm process. Making that leap once Apple's reservations expire may remove much of the performance gap.

>t's still unclear to me what the market for x86-based machines is going to look like 3-5 years down the road.

~750M of PC are used in Cooperate and Business. And if IBM mainframe has taught us anything business couldn't give a f what ISA they are using. They want 100% ( when possible ) backward compatibility. Considering most Business PC are already using low end x86 CPU, and they have not been CPU bottlenecked for long, I mean most of them are still using PC with HDD for all sort of reasons. So I dont see them switching any time soon.

Excluding that the vast majority of PC are Laptop and Gaming PCs. Gaming wont switch. Laptop is a complicated issue. And I think it sort of depends on Microsoft and how aggressive will Apple price their MacBook Air. I see no reason why they cant have the lowest cost MacBook starting at $799 while maintaining the same margin.

On Server I think x86 is pretty much done, at least in the HyperScaler market. The whole reason why Amazon is an Intel foundry customer is because Amazon are moving everything they could to Graviton. So you either loss Amazon as a customer leaving them to buy some x86 for some smaller part of their AWS, or you try and Fab them a chip they want.

And I believe, for political reason, Intel will gain lots of foundry customers.

So while a potential lost of 30-40% of their DC market in the next 5 years. x86 should still be a healthy state in 5 years time.

TSMC Announced a 2nm soon after Intel announced it can catch up https://9to5mac.com/2021/07/28/tsmc-2nm-production-set-for-2...

Yeah, it’s dishonest, but the consumers of their processors most often look at benchmarks, not nm width. The only people this might actually fool will be casual stock traders.

Dishonesty aside, this hits like a ton of bricks: FOUR YEARS to leadership. Even with Moore's law slowing, that is a loooong time. I remember when Intel had seemingly a two year lead in process, around 65nm I think.

And now, project four years to leadership, and assuming the economics and competition won't respond.

If Moore's Law hits a major wall (it has to ... sometime, right?), then that may be the catch-up.

Intel 3 is what used to be called Intel 7nm+, which is more or less equivalent to newer batches of TSMC N5.

I mean, this is nothing new. If you want newer processes, AMD does a pretty good job of keeping their Zen architecture on the cutting edge without under-engineering the end product. Intel's bag has always been about selling very specific things to specific audiences: bare-metal virtualization for people on Windows/MacOS (rip), super-wide SIMD busses, special overclockable models, IMEI, the list goes on. Intel, much like Apple, is selling a product that supposedly offers a better "experience" than their competitors, which moves like gangbusters in the corporate sector.

Those numbers refer to what a planar transistor would have to look like to achieve the same density. Of course nobody is actually using planar transistors anymore.

x Nanometer or y Ångström isn't based on physical size anymore and is just a marketing term.

From the article:

>A lot of the industry, for whatever reason, hasn’t learned that these numbers aren’t actually a physical measurement. They used to be, but when we moved from 2D planar transistors to 3D FinFET transistors, the numbers became nothing more than a marketing tool.

The Intel 3 i7 13N65GK is pretty good, but the Intel 3 Xeon Platinum 9451 v6 is better for virtualization and stuff.

1 angstrom = 0.1 nm

10 A = 1 nm

You joke, but CPUs can draw hundreds of amps.

It’s actually quite little. Most modern CPUs run at ~1V, which means 20A ~ 20W CPU. Ultrabooks are able to pull off 45W, so double of that.

With TSMC.

One of the announcements here is that Intel will be getting literally the first of the new batch of ASML's machines.

They do buy from them (and invest in them). But as far as understand, ASML is one piece of the puzzle.

There are several other parts of the process that are not touched by ASML

Intel owns just under 5% of ASML. TSMC owns a bit of ASML as well. So does Samsung.

Ha, I first thought they meant “Ampere” (vaporizing the conductor at those numbers). Using “Å” would make it clear they mean “Ångström”. Someone should tell Intel that Unicode exists.