Discussion Intel current and future Lakes & Rapids thread

[moinmoin]

We don't yet know the reason why N3 broke TSMC's cadence. It might have been due to issues getting the process to work, or it might have been issues related to sufficient delivery of EUV scanners. After all, TSMC has been expanding N7+/N6 wafer starts which use a limited amount of EUV, and is in the process of expanding N5/N4 wafer starts which use even more EUV. N3 will reportedly use EUV for over 20 layers and they have some pretty aggressive targets even for risk production (30k wpm) let alone the first phase of mass production (105k wpm) If it is the latter ASML does look to be making progress with increasing yearly shipments so it may not be too much of a gating factor. I guess the wildcard is how heavily DRAM fabs get into EUV.

Before worrying that Intel may be "stuck" at Intel 4, how about worrying that they can actually get there in the promised timeline? They haven't exactly been good at keeping their promises for the past half dozen years, and all the while were saying they were on track and claiming just temporary delays.

Good call. Reminder we are still in a pandemic situation that effects shipments even at a very basic level. And if TSMC is affected by shipment delays of EUV equipment by ASML due to the current circumstances chances are Intel is as well.

 

[mikk]

Intel 3 is simply a revision of Intel 4. It is like 14+ vs 14 nm.

Intel 3 is more than a simple 14->14+ revision: https://images.anandtech.com/doci/16823/Intel Accelerated Briefings FINAL-page-006.jpg

Performance/watt gain is very similar to 7->4

 

[Exist50]

Intel didn't release much of anything exciting from ~2017 to early 2021. You are correct that the delays were far longer than Intel wanted or publicly stated. This was due to bad management, poor investment in necessary equipment, firing of great engineers, overconfidence, etc. That said, at what point do we say that they are back? What is the metric you are using to judge this?

Intel had 2 generation releases in the last 8 months. Raptor Lake has at least engineering samples in the wild. Meteor Lake is taped in and taped out. So, Intel has a reasonable shot of 4 CPU generations in just over 2 years. Is that enough, or does Intel need more to be back on track in your mind?

I think the criteria is simple. Can Intel release new products on a proper yearly cadence, as promised, and are those products leadership? I don't think they should get any credit for MTL slipping a year and inserting a refresh gen (Raptor Lake), nor for a flop like Rocket Lake.

 

[dullard]

Intel 3 is more than a simple 14->14+ revision: https://images.anandtech.com/doci/16823/Intel Accelerated Briefings FINAL-page-006.jpg

Performance/watt gain is very similar to 7->4

Intel 4 used to be called 7 nm and Intel 3 used to be called 7+ signaling that it was just a revision. Intel 3 is supposed to have a lot of the same features as Intel 4. Yes, the power levels should drop giving a performance/watt gain. But, the changes are not nearly as technically challenging as other steps between nodes. https://www.anandtech.com/show/16823/intel-accelerated-offensive-process-roadmap-updates-to-10nm-7nm-4nm-3nm-20a-18a-packaging-foundry-emib-foveros

2023 H2, Intel 3: Previously known as Intel 7+. Increased use of EUV and new high density libraries. This is where Intel’s strategy becomes more modular – Intel 3 will share some features of Intel 4, but enough will be new enough to describe this a new full node, in particular new high performance libraries. Nonetheless, a fast follow on is expected. Another step up in EUV use, Intel expects a manufacturing ramp in the second half of 2023 with an 18% performance per watt gain over Intel 4.

 

[dullard]

I think the criteria is simple. Can Intel release new products on a proper yearly cadence, as promised, and are those products leadership? I don't think they should get any credit for MTL slipping a year and inserting a refresh gen (Raptor Lake), nor for a flop like Rocket Lake.

"Leadership" and "flop" are vague or undefined (what if it leads in some software but not others?). That isn't a simple criteria.

I agree with yearly cadence as a criteria. It is measurable and pretty easy to agree on. Not allowing refreshes seems a bit harsh and arbitrary. Intel's heyday years included ticks and tocks, 2 generations per node size.

 

[mikk]

Intel 4 used to be called 7 nm and Intel 3 used to be called Intel 7+. Intel 3 is supposed to have a lot of the same features as Intel 4. Yes, the power levels should drop giving a performance/watt gain. But, the changes are not nearly as technically challenging as other steps between nodes. https://www.anandtech.com/show/16823/intel-accelerated-offensive-process-roadmap-updates-to-10nm-7nm-4nm-3nm-20a-18a-packaging-foundry-emib-foveros

It doesn't matter how it was called before, it's not a simple 14-->14+ refresh. Your 14+ comparison is completely off and that's why it was a good idea from Intel to change the naming. A better comparison would have been 10+ to 10SF.

 

[Exist50]

"Leadership" and "flop" are vague or undefined (what if it leads in some software but not others?). That isn't a simple criteria.

I admit that's vague, but at Intel's height, there was zero question who the leader was. Thinking back to the Core 2 launch, I feel like that kind of reception is the bar to reach. Gelsinger is the one always talking about "unquestioned leadership", right? Then it should be self evident.

I agree with yearly cadence as a criteria. It is measurable and pretty easy to agree on. Not allowing refreshes seems a bit harsh and arbitrary. Intel's heyday years included ticks and tocks, 2 generations per node size.

Refreshes are just delays or deviations from a yearly cadence by another name. A node shrink is fine, but something like Coffee Lake or Comet Lake? Nah. Raptor Lake seems far more like those than like Ivy Bridge.

 

[IntelUser2000]

I'm thinking the opposite, actually. Intel 4 will get Grand RIdge, Meteor Lake, and Granite Rapids, but what's on Intel 3? Conflicting rumors about Lunar Lake, and that's it. Maybe a future server gen? I think it'll probably end up used more-so for IO dies and the like, especially as Intel's last node without the PowerVia complexity.

It makes no sense that a subpar node similar to Icelake's will be long lived while a one that could be considered mature won't be.

We know that Intel 4 is a delayed process. Delayed products almost always suck because it's due to not meeting expectations. They have said "12 month yield delay cut down to 6 months" using fancy tricks.

So you are saying they'll practically abandon 3 and go to 20A somehow? I don't believe it. They'll need their 7nm process to be mature before progressing.

Intel 3 is more than a simple 14->14+ revision: https://images.anandtech.com/doci/16823/Intel Accelerated Briefings FINAL-page-006.jpg

Performance/watt gain is very similar to 7->4

Yes and I bet you part of the reason for the big gain is that 7 is a disappointment just like we got so much gains on 10SF due to 10 nm being sucky.

 

[Markfw]

I admit that's vague, but at Intel's height, there was zero question who the leader was. Thinking back to the Core 2 launch, I feel like that kind of reception is the bar to reach. Gelsinger is the one always talking about "unquestioned leadership", right? Then it should be self evident.



Refreshes are just delays or deviations from a yearly cadence by another name. A node shrink is fine, but something like Coffee Lake or Comet Lake? Nah. Raptor Lake seems far more like those than like Ivy Bridge.

I think core2 (or core??) 2006ish was their first unquestioned leader. (in the last 20 years) Then AMD was at Zen3 release.

Alder lake is not quite "unquestioned leader" IMO. But a step in the right direction. Also in x86 server AMD is currently the "unquestioned leader". Before Intel also had the server locked up.

 

[DrMrLordX]

I think the criteria is simple. Can Intel release new products on a proper yearly cadence, as promised, and are those products leadership? I don't think they should get any credit for MTL slipping a year and inserting a refresh gen (Raptor Lake), nor for a flop like Rocket Lake.

The real question is whether Intel can regain performance and perf/watt leadership in the server/workstation sector where they have minted so much of their money. Right now most of their server sales are still Cascade Lake. IceLake-SP has not been broadly successful, and Sapphire Rapids is still MIA.

 

[maddie]

Intel didn't release much of anything exciting from ~2017 to early 2021. You are correct that the delays were far longer than Intel wanted or publicly stated. This was due to bad management, poor investment in necessary equipment, firing of great engineers, overconfidence, etc. That said, at what point do we say that they are back? What is the metric you are using to judge this?

Intel had 2 generation releases in the last 8 months. Raptor Lake has at least engineering samples in the wild. Meteor Lake is taped in and taped out. So, Intel has a reasonable shot of 4 CPU generations in just over 2 years. Is that enough, or does Intel need more to be back on track in your mind?

My concern is why 4 generations in 2-3 yrs? As soon as you start marketing one, it will be replaced. Isn't this counterproductive?

 

[Exist50]

I think core2 (or core??) 2006ish was their first unquestioned leader. (in the last 20 years) Then AMD was at Zen3 release.

Alder lake is not quite "unquestioned leader" IMO. But a step in the right direction. Also in x86 server AMD is currently the "unquestioned leader". Before Intel also had the server locked up.

Agreed on both counts. Alder Lake is decent enough, but the problem is they are now going to coast on that for like 2 more years, and that is a problem. They have a similar problem in servers. Even if Sapphire Rapids matches Milan, Genoa will follow not long after, and then Intel has nothing till god knows when in hopefully '23 they get out Granite Rapids. They don't need a good product gen. They need many.

 

[dullard]

My concern is why 4 generations in 2-3 yrs? As soon as you start marketing one, it will be replaced. Isn't this counterproductive?

Marketing and reputation are well past my knowledge set.

From the best that I can tell, most consumers don't know the generation or what it means. For the enthusiasts, we want the latest and greatest, so I'd personally take many iterations if possible. I think it is mostly OEMs and motherboard manufacturers that are really impacted. Socket compatibility is important there. But if Meteor Lake is mobile only, even that doesn't matter too much.

 

[Doug S]

Intel had 2 generation releases in the last 8 months. Raptor Lake has at least engineering samples in the wild. Meteor Lake is taped in and taped out. So, Intel has a reasonable shot of 4 CPU generations in just over 2 years. Is that enough, or does Intel need more to be back on track in your mind?

If Intel is able to deliver "Intel 4" (f.k.a. 7nm) chips in large quantities across their whole product portfolio then I might believe they are back in stride. If we see a launch like 10nm had where it is only being used for certain product segments for a long time due to terrible yields then I'm going to figure it is same old same old.

 

[Doug S]

Good call. Reminder we are still in a pandemic situation that effects shipments even at a very basic level. And if TSMC is affected by shipment delays of EUV equipment by ASML due to the current circumstances chances are Intel is as well.

I wasn't suggesting EUV shipments were delayed due to covid (though that is of course quite possible) but more that the growth in TSMC's business for N7+ and newer processes may have been larger than originally forecast. In fact that's almost assurdly true, given the pandemic boosted sales in electronics. They are taking delivery of only so many EUV scanners per year even if ASML is delivering as many as promised (all we know is they are making them as fast as they can but will still have orders backlogged for years)

If the number of EUV scanners TSMC owns at any given time is a gating factor on decisions whether to expand existing processes or make the schedule for newer ones, it is quite possible they might decide to prioritize expanding N7+/N6/N5/N4 capacity in exchange for a delay in N3. Any customers on those current nodes they turn away might go to Samsung, and it could be difficult to get them back (due to e.g. sunk costs for design, mask sets, product qualification, etc.) but they don't have to worry about N3 customers defecting over a moderate delay since no foundry offers anything comparable during that time period.

The only customer that would potentially care is Apple, and if feathers are ruffled TSMC can probably find some way to make them happy using the giant piles of cash they are making from the N7+/N5/etc. wafers they raised prices on earlier this year!

 

[Abwx]

It doesn't matter how it was called before, it's not a simple 14-->14+ refresh. Your 14+ comparison is completely off and that's why it was a good idea from Intel to change the naming. A better comparison would have been 10+ to 10SF.

Actually 14nm was greatly improved during its life cycle, about a full node shrink power efficency wise.

A third improved process, "14nm++", is set to begin in late 2017 and will further allow for +23-24% higher drive current for 52% less power vs the original 14nm process.

14 nm lithography process - WikiChip

The 14 nanometer (14 nm) lithography process is a semiconductor manufacturing process node serving as shrink from the 22 nm process. The term '14 nm' is simply a commercial name for a generation of a certain size and its technology, as opposed to gate length or half pitch. The 14 nm node was...

en.wikichip.org

 

[mikk]

Actually 14nm was greatly improved during its life cycle, about a full node shrink power efficency wise.

14 nm lithography process - WikiChip

The 14 nanometer (14 nm) lithography process is a semiconductor manufacturing process node serving as shrink from the 22 nm process. The term '14 nm' is simply a commercial name for a generation of a certain size and its technology, as opposed to gate length or half pitch. The 14 nm node was...

en.wikichip.org

Yes from 14 to 14+++ was quite big over 5 years but this is not what he said. He said it's like 14+ vs 14 nm. According to Intel there is a 5.5% performance improvent from 14 to 14+. A better comparison is 10+ to 10SF with a 17-18% better transistor performance. Do we know if the compute tile for consumer CPUs uses the HP library on Intel 3/4? In this case there are some area improvements as well.

 

[Hans de Vries]

Was it BapCo that started operating from Intel offices?

You must post your own commentary to the link/image/story you have posted. This is your one-time zero point warning. Its in the CPU rules


esquared
Anandtech Forum Director

 

[Hans de Vries]

Was it BapCo that started operating from Intel offices?

Shervin Kheradpir - VP, Sales & Marketing Group & GM, Platform Evaluation & Competitive Assessment @ Intel - Crunchbase Person Profile

Shervin Kheradpir is vice president of the Sales and Marketing Group and general manager of Platform Evaluation and Competitive Assessment a...

www.crunchbase.com

You must post your own commentary to the link/image/story you have posted. This is your one-time zero point warning. Its in the CPU rules



esquared
Anandtech Forum Director

 

[moinmoin]

Shervin Kheradpir - VP, Sales & Marketing Group & GM, Platform Evaluation & Competitive Assessment @ Intel - Crunchbase Person Profile

Shervin Kheradpir is vice president of the Sales and Marketing Group and general manager of Platform Evaluation and Competitive Assessment a...

www.crunchbase.com

View attachment 54095

Good find. He joined Intel in 1989, BAPCo was founded in 1991...

 

[Exist50]

So you are saying they'll practically abandon 3 and go to 20A somehow? I don't believe it. They'll need their 7nm process to be mature before progressing.

What I'm assuming is that by the time Meteor Lake and/or Granite Rapids roll out in '23, Intel 4 will be in decent enough shape, i.e. more like 10++ (Tiger Lake) than 10+ (Ice Lake). Then, looking at '24, I think 20A will be in good enough shape for at least some small chiplets by year's end, but maybe not a big die like the server chips use (assume Granite Rapids topology holds for another gen).

But even if 20A is in good shape in '24, Intel 3 still has value because PowerVia in particular adds a ton of issues for stuff like post-Si validation.

 

[mikk]

If it turns out that Arrow Lake is based on Intel 3 there is not so much with Intel 4 for consumer apparently.

 

[Joe NYC]

View attachment 54094

Thank you Hans! I had this vague recollection, but could not find the specifics - that BapCo most likely started just as an Intel Front, there the mission really was deception on behalf of Intel.

 

[igor_kavinski]

ASML's High-NA EUV lithography machine is set to transform chipmaking (cnbc.com)

Multi-Patterning EUV Vs. High-NA EUV (semiengineering.com)

Really enjoyed these articles.

For context, each of the current EUV machines reportedly has over 100,000 components, and they take 40 freight containers or four jumbo jets to ship. They cost around $140 million each, according to reports.

WOAHHHHH, dude!

 

[moinmoin]

ASML's High-NA EUV lithography machine is set to transform chipmaking (cnbc.com)

Multi-Patterning EUV Vs. High-NA EUV (semiengineering.com)

Really enjoyed these articles.



WOAHHHHH, dude!

Yeah, before your edit you asked what's the limiting factor. Aside the quote you now added there's also this:

EUV machines take years to build and ASML can only ship so many of them in any given year. Last year, it sold just 31, according to its financials, and it’s made around only 100 in total.

It's simply the very definition of high tech. Also constant improvements are being made, so these are the opposite of mass production.