Discussion Intel current and future Lakes & Rapids thread

[Markfw]

Like I said Client.

-S: Desktop
-H: High Performance Laptop
-P High Performance Ultrathin
-U Low Power Ultrathin
-M Ultra Low Power

So desktop Raptor lake desktop in October of this year(after Zen 4 I think), and nothing but laptop/mobile variants in the foreseeable future, like server or HEDT. That's pitiful.

So is Intel going to move to being a laptop/mobile only vendor ? Until 2025 or so ?

 

[cortexa99]

Not true, base speeds are different (<<50% alder lake), boost is same only on P-cores, but again will be different in practice because of power limit differences.

“””
Do note that the P-Cores for this engineering sample ran at a base clock of 1.4 GHz and a max boost clock of 3.8 GHz (P-Cores) while the E-Cores were split with 2 clusters running at 2.8 GHz and the other 2 clusters running at 1.0 GHz so we can already see some inconsistency in the clock speeds … In terms of performance, the single-core performance of Intel's Raptor Lake Core i9-13900 CPU is slower than the Alder Lake Core i9-12900K chip. This is because the Alder Lake CPU also has the full backing of an optimized BIOS and a higher TDP limit versus the Non-K Raptor Lake chip. And also, as said before, the clock speeds for the Raptor Lake CPU aren't stable.
“””
- https://wccftech.com/intel-raptor-lake-core-i9-13900-es-cpu-benchmarks-leak-out-20-faster-than-core-i9-12900k-in-multi-threading/

The original article and video stated the Ecore max boost is 2.8Ghz, there's no "2 clusters running at 1Ghz" statement.

E-Core最大睿频2.8GHz

Deepl translation: E-Core up to 2.8 GHz

Maybe WCCF saw 1Ghz clocks that readed by CPUZ which is the base clock of E cores of this Raptor sample. But no evidence or statements claimed E cores stall at 1Ghz or boost to 2.8Ghz when being benched.

 

[Exist50]

My current expectation of client chips based on reliable leakers(read: NOT MLID, RGT, Adored)

Intel 7
Raptorlake-S: October 2022
Raptorlake-P/H: Jan 2023

Intel 4

Meteorlake-S Low/Mid: Q1 2024
Meteorlake-M 5W: Mid 2023
Meteorlake-P/H: Jan 2024

Intel 20A
Arrowlake-S: Mar 2024

I think MTL M/P/H will probably all start shipping to OEMs around mid '23 (maybe 1-2Q difference, with the possibility of the first one slipping into Q3). No reason for a huge gap for two almost identical compute dies. Though I'm doubtful they'll try to push MTL down to 5W. Probably wouldn't be a good product.

Also, I think it's very unlikely they'd have a 20A ARL product ready for March '24. That would be beating their official process roadmap by a quarter, and even if I consider myself begrudgingly optimistic about the state of 20A/18A, Intel certainly has given no reason to believe they're capable of that. I certainly don't.

There's also the question of what happened to N3-based ARL-P/H. Still on the roadmap? For when?

 

[IntelUser2000]

I think MTL M/P/H will probably all start shipping to OEMs around mid '23 (maybe 1-2Q difference, with the possibility of the first one slipping into Q3). No reason for a huge gap for two almost identical compute dies.

Yes there is, to allow products 1 year time on the market. Unless they are skipping Raptorlake P and H which I doubt. Perhaps it's also possible we'll see U considering the low volume on Alderlake so they can skip it for Raptorlake.

 

[Thunder 57]

My current expectation of client chips based on reliable leakers(read: NOT MLID, RGT, Adored)

Intel 7
Raptorlake-S: October 2022
Raptorlake-P/H: Jan 2023

Intel 4

Meteorlake-S Low/Mid: Q1 2024
Meteorlake-M 5W: Mid 2023
Meteorlake-P/H: Jan 2024

Intel 20A
Arrowlake-S: Mar 2024

Where does Intel 3 fit in? Isn't that supposed to be a more refined Intel 4 with both HP and HD libraries?

 

[JasonLD]

Where does Intel 3 fit in? Isn't that supposed to be a more refined Intel 4 with both HP and HD libraries?

Products announced so far for Intel 3 are Granite Rapids and Sierra Forest in 2024.

 

[Exist50]

Yes there is, to allow products 1 year time on the market. Unless they are skipping Raptorlake P and H which I doubt. Perhaps it's also possible we'll see U considering the low volume on Alderlake so they can skip it for Raptorlake.

I really don't think that is good enough reason to let an uncompetitive product line remain.

 

[jpiniero]

I really don't think that is good enough reason to let an uncompetitive product line remain.

But unless they are going to dual source the CPU chiplet at TSMC they aren't going to have anywhere near the volume.

 

[Exist50]

But unless they are going to dual source the CPU chiplet at TSMC they aren't going to have anywhere near the volume.

They should have enough Intel 4 for Meteor Lake. They were planning a server chip the same year, after all.

 

[IntelUser2000]

I really don't think that is good enough reason to let an uncompetitive product line remain.

Yes there is. They do not build and sell the device, it's only sold then companies build systems around it. One will be skipped in favor of the other if the timeframes are too short. Intel has a lot of power but they are ultimately at the mercy of the companies as well.

And if one product generation is skipped in favor of the other then Intel has to think of the volume it'll be shipped and whether it makes sense at all for one generation to be in the market. It may make sense for Intel to lower their revenue a bit and sell the next gen notebook chip 6 month after the previous but they need to take into system builders and they will *not* see it the same way. If they do it it'll be "too bad so sad" and they'll lose.

I've read articles and references previously that Intel was pressured to release it a bit later because the vendors thought the duration was too short and that was for 12 months! In fact the 12 month period was said by Intel to be an optimal ratio between technological advances and revenue. <12 results in lower revenue because you didn't fully use the previous generation. >12 results in lower revenue because loss in competitiveness coupled with loss in interest. In fact nearly every tech industry from cars to phones to computers settled for the 12 month cadence.

Trust me on this. They can shorten it a bit for desktop chips with sockets and two generations(if you count one generation as an annual product) but you will not see that for laptops with two thirds of the revenue and volume share.

*I also am not sure if you are referring to Meteor or Raptor when saying "not competitive device". I am assuming Raptorlake. Considering that Alderlake's battery life regression is likely unrelated to core uarch design but rather the platform/SoC, it's not implausible we'll see improvements in the Raptorlake generation. Highly doubt they'll greatly improve it but will be a welcome change if they go back to the Cometlake days.

 

[IntelUser2000]

Where does Intel 3 fit in? Isn't that supposed to be a more refined Intel 4 with both HP and HD libraries?

Intel 3 based on the recent reveal of Intel 4 is much more than just having HD libraries. It'll be the first official bleeding edge Intel foundry process with IO, capacitors, and standardized design methodologies needed for customers. If you wonder when companies like Nvidia can test the waters using some GPUs or something Intel 3 would be the start.

I could see maybe two generation successor to Jasper Lake using Crestmont cores made on it? HD libraries are quite dense so it would be a good low cost/low power chip process.

But yea no major client volume will be on it. Obviously to me their goal is gaining process leadership back and skipping lineups is the only way to do so. Even so, quite a bold move akin to what a smaller company would do. If they were much stable such "practices" will be essentially forbidden.

 

[moinmoin]

I've read articles and references previously that Intel was pressured to release it a bit later because the vendors thought the duration was too short and that was for 12 months! In fact the 12 month period was said by Intel to be an optimal ratio between technological advances and revenue. <12 results in lower revenue because you didn't fully use the previous generation. >12 results in lower revenue because loss in competitiveness coupled with loss in interest. In fact nearly every tech industry from cars to phones to computers settled for the 12 month cadence.

I guess this could be summarized as window of opportunity. All participants in the market both need lead time for new products but also need to refresh their product range regularly to not get outmoded.

 

[DrMrLordX]

Question: We have seen a lot of dies and a couple of ES chips, but when will we see any real retail silicon from Intel, server, desktop or HEDT ?

There's no clear answer. At the latest, it'll be after Intel officially launches Sapphire Rapids into retail channels. If there were more QS/early ODM silicon available then leakers much have something to work with, which kind of makes you wonder about the state of Sapphire Rapids. At the same time, leakers were mostly mum about Milan performance for a very long time despite it sampling all the way back in 2019. So maybe Intel is just doing a good job of controlling leaks except on a few stray engineering samples. Maybe.

 

[dullard]

I've read articles and references previously that Intel was pressured to release it a bit later because the vendors thought the duration was too short and that was for 12 months! In fact the 12 month period was said by Intel to be an optimal ratio between technological advances and revenue. <12 results in lower revenue because you didn't fully use the previous generation. >12 results in lower revenue because loss in competitiveness coupled with loss in interest. In fact nearly every tech industry from cars to phones to computers settled for the 12 month cadence.

It is true that Intel tries for about a 12 month cadence. But, there are very notable exceptions. There are timeline slips and a few odd chips that break up that timeline.

Desktop Chip​	Generation​	Months As Top Generation​
Nahalem​	​	14.2​
Westmere​	​	12.1​
Sandy Bridge​	2​	15.6​
Ivy Bridge​	3​	13.1​
Haswell​	4​	12.0​
Haswell Refresh​	4​	12.0​
Broadwell​	5​	2.1​
Sky Lake​	6​	12.8​
Kaby Lake​	7​	13.2​
Coffee Lake​	8​	12.5​
Coffee Lake Refresh​	9​	10.1​
Comet Lake​	10​	19.3​
Rocket Lake​	11​	7.2​
Alder Lake​	12​	10.9? Estimated​

​

Haswell Refresh wasn't even called a new generation, so you could argue there was a 24 month gap. Broadwell was barely released for desktop, so its 2.1 month reign as the top chip is just a one time oddity. Rocket Lake was just a filler chip since Alder Lake wasn't ready. Basically, 12 months is a goal rather than a hard set rule.

 

[dullard]

Where does Intel 3 fit in? Isn't that supposed to be a more refined Intel 4 with both HP and HD libraries?

As of July 2021, Intel 3 was supposed to be the top process for an entire year from Q3 2023 to Q2 2024. Note: this is when the process is ready to begin production, not when chips are selling as there is time necessary for testing and volume production ramp up. This was from Intel's Accelerated event. https://www.anandtech.com/show/16823/intel-accelerated-offensive-process-roadmap-updates-to-10nm-7nm-4nm-3nm-20a-18a-packaging-foundry-emib-foveros

But, both Intel 20A and 18A are ahead of that schedule. Intel has moved up Intel 20A to Q1 2024. That gives Intel 3 top reign for just a brief period in Q3 and Q4 2023. https://www.anandtech.com/show/17344/intel-opens-d1x-mod3-fab-expansion-moves-up-intel-18a-manufacturing-to-h22024
Thus Intel is not rumored to be producing any client (desktop or mobile) chips on Intel 3. This is for timing purposes (Intel does want to try to avoid multiple 6 month chip cadences), and also it takes the production pressure off of the limited Intel 3 production equipment. It won't be shared between client and server chips, but instead can be dedicated.

Obviously to me their goal is gaining process leadership back and skipping lineups is the only way to do so. Even so, quite a bold move akin to what a smaller company would do. If they were much stable such "practices" will be essentially forbidden.

To try and actually hit those timelines, Intel has inserted a test node that will not reach consumers.

https://images.anandtech.com/doci/17344/Intel_Test_Node.jpg

 

[Exist50]

*I also am not sure if you are referring to Meteor or Raptor when saying "not competitive device". I am assuming Raptorlake. Considering that Alderlake's battery life regression is likely unrelated to core uarch design but rather the platform/SoC, it's not implausible we'll see improvements in the Raptorlake generation.

Meant Raptor Lake. It's an Alder Lake refresh. It might be convincing enough vs Rembrandt, at least aside from graphics, but no way it can compete with Phoenix Point. Push comes to shove, I think skipping it for Meteor Lake would be very reasonable, or even doing a 10th-gen-like split.

But, both Intel 20A and 18A are ahead of that schedule. Intel has moved up Intel 20A to Q1 2024. That gives Intel 3 top reign for just a brief period in Q3 and Q4 2023.

20A is scheduled for H1'23, i.e. Q2.

 

[jpiniero]

Push comes to shove, I think skipping it for Meteor Lake would be very reasonable, or even doing a 10th-gen-like split.

I suspect what ends up happening is that Raptor Lake gets a full burn and you only see a tiny amount of Meteor Lake M in the middle of 23. Then we'll see about what they can do at CES 24.

20A is scheduled for H1'23, i.e. Q2.

No they claim 1H24... and that's "manufacturing ready", not production shipments. And that's before the inevitable delays.

 

[pakotlar]

I suspect what ends up happening is that Raptor Lake gets a full burn and you only see a tiny amount of Meteor Lake M in the middle of 23. Then we'll see about what they can do at CES 24.



No they claim 1H24... and that's "manufacturing ready", not production shipments. And that's before the inevitable delays.

Given that they pulled up 18A by a quarter, and 18A is a refined 20A process, there is little chance that 20A gets pushed back. They wouldn’t pull up 18A unless they were near certain. If anything, 20A may just get cancelled in favor of 18A-only.

 

[shady28]

So is anything other than the Apple A17 (and M2 Pro / M2 Max) going to be on TSMC N3 in 2022-2023?

SemiWiki has an in depth analysis of Intel 4 and concluded that it is more dense than TSMC N5 but not as dense as N3. It is also expected to perform better than TSMC N3.

"These values are consistent with a 4 designation since it slots between N5 and N3 for the leading foundry company TSMC, although it is closer to TSMC N3 than TSMC N5. We also believe Intel 4 will have performance slightly better than TSMC N3. I didn’t include Samsung in Figure 4 but based on my current estimates Intel 4 is denser than Samsung GAE3. "

Source:

Intel 4 Deep Dive - Semiwiki

As I previously wrote about here, Intel is presenting their Intel 4 process at the VLSI Technology conference. Last Wednesday Bernhard Sell (Ben) from Intel gave the press a briefing on the process and provided us with early access to the paper (embargoed until Sunday 6/12). “Intel 4 CMOS...

semiwiki.com

 

[jpiniero]

So is anything other than the Apple A17 (and M2 Pro / M2 Max) going to be on TSMC N3 in 2022-2023?

Arrow Lake?

 

[Timmah!]

Do you think we will see multiple compute tile (well 2) products on the Meteor Lake, similarly to Ryzen 9?

 

[IntelUser2000]

@shady28 The density comparisons are irrelevant for Intel 4 because they aren't using it to full advantage and no one outside of Intel will use it. The greater than 2x was claimed for both 14nm and 10nm. The Core chips ignored that as business as usual and had density of roughly 2x. At least for Atoms and graphics the density gains took hold but that's about it.

Intel might be one of the first to use TSMC N3 with graphics tiles starting with Meteorlake.

@Timmah! Definitely not for Meteorlake. No high end products and they've already shown a 6+8 compute tile which will be enough.

Haswell Refresh wasn't even called a new generation, so you could argue there was a 24 month gap. Broadwell was barely released for desktop, so its 2.1 month reign as the top chip is just a one time oddity. Rocket Lake was just a filler chip since Alder Lake wasn't ready. Basically, 12 months is a goal rather than a hard set rule.

Good chart. I am addressing specifically for mobile and the rule is that it has to be 12 months minimum.

You can see the process delays from the chart. 3 months for 22nm, about 6 months for 14nm and everyone knows how long for 10nm.

 

[shady28]

Arrow Lake?

That's supposedly Intel 20A.

What I'm getting at is, if we assume Meteor Lake launches on Intel 4 mid 2023, what will be on TSMC N3? As far as a I know the only mass market competing products will be the higher end M2's and later in 2023 the iPhone 15's A17 SoC. Unless some N3 GPUs are planned, which is possible.

It really sounds like the node advantage that TSMC has over Intel will, outside of Apple, dissipate entirely in 2023.

 

[JasonLD]

Definitely not for Meteorlake. No high end products and they've already shown a 6+8 compute tile which will be enough.

Intel's range of Meteor Lake is 5-125W, I assume there are SKUs higher than 6+8, just that it is not shown yet.

 

[Timmah!]

@shady28 The density comparisons are irrelevant for Intel 4 because they aren't using it to full advantage and no one outside of Intel will use it. The greater than 2x was claimed for both 14nm and 10nm. The Core chips ignored that as business as usual and had density of roughly 2x. At least for Atoms and graphics the density gains took hold but that's about it.

Intel might be one of the first to use TSMC N3 with graphics tiles starting with Meteorlake.

@Timmah! Definitely not for Meteorlake. No high end products and they've already shown a 6+8 compute tile which will be enough.

Wait, 6 + 8 as in 6P + 8E?