Discussion Intel current and future Lakes & Rapids thread

[dullard]

It would be a PR disaster for Intel if the median improvement is 5%. I'm sure Intel knows that and they will do their best to at least target 10%. There is a slight chance that Intel may not want to be too ambitious with Intel 4 and hence keep architectural enhancements to a minimum but it may turn out to be their funeral in the mindshare department among gamers/enthusiasts.

Is IPC what gamers/enthusiasts most care about? Or do they care more about pure performance? Intel 4 will give Meteor lake some performance/watt gains that could be used for pure performance (higher clock speeds) without much IPC improvement needed.

Unless there is still some surprise still lurking (like a not yet revealed CPU tile with more cores than we are expecting), I'm not too hopeful for Meteor Lake desktop. Desktop Meteor Lake will exist, but I think Intel is going more for new features and power savings over pure IPC gains. But those new features won't be applicable to everyone and software that uses them won't be instantly ready at launch. Take the rumored VPU for example. If you need it, and once software is available, it will be a great feature with huge IPC gains (on that software only). But, honestly, it might not do much at all for gamers who might not ever use it.

I still think Raptor Lake desktop will be better than people think and Meteor Lake desktop will be worse than they hope. Arrow Lake is where it really comes together. We'll know a lot more in 3 weeks at Hot Chips when Meteor Lake and Arrow Lake are discussed. https://hotchips.org/advance-program/

 

[igor_kavinski]

Intel 4 will give Meteor lake some performance/watt gains that could be used for pure performance (higher clock speeds) without much IPC improvement needed.

RPL is at 5.7 - 5.8 GHz. MTL will do what? 6 GHz on air (roughly 5% gain or even less)? That can't be their only gambit. They need something more in the architecture.

 

[mikk]

RPL is at 5.7 - 5.8 GHz. MTL will do what? 6 GHz on air (roughly 5% gain or even less)? That can't be their only gambit. They need something more in the architecture.

Does it even matter on desktop? Some leaker claimed MTL on desktop is a low-midrange CPU and ARL-S will be the higher end replacement for desktop.

https://twitter.com/x/status/1533759998186393600

 

[DrMrLordX]

Zen 3 is old now so possibly AMD can let Intel license it

LOL I hope you're joking. But Intel is never going to let that happen. x86-64 was already like eating crow for them.

Does it even matter on desktop? Some leaker claimed MTL on desktop is a low-midrange CPU and ARL-S will be the higher end replacement for desktop.

That seems highly probable. We still haven't seen anything larger than 6P+8e in mockups of dice or anything. Well I haven't. I might be behind the curve here.

 

[igor_kavinski]

But Intel is never going to let that happen.

Let them have another abysmal quarter...

 

[Exist50]

It would be a PR disaster for Intel if the median improvement is 5%. I'm sure Intel knows that and they will do their best to at least target 10%.

Since when have Intel's desires and their ability aligned? They haven't put in the engineering to get 10%, so they won't get it.

Does it even matter on desktop? Some leaker claimed MTL on desktop is a low-midrange CPU and ARL-S will be the higher end replacement for desktop.

That still leaves them a ~2 year gap in which all they'll have is Raptor Lake. Might hang on in raw performance against Raphael, but that's a significant gap in efficiency and platform features, and Raphael-X will be the clear performance champion for most of that time period.

But then what happens when Arrow Lake comes out? Let's say Lion Cove does get 10-20% IPC over RWC. Is that enough? It'll be up against Zen 5, with probably a similar, if not greater increase. And that's ignoring the V-Cache models that should be becoming more high volume by then. Seems to me like Arrow Lake won't be enough to fundamentally change Intel's competitive situation. Might take Panther Lake, or even Nova Lake for that. Feels like a repeat of the Pentium 4 days, in some sense.

Ironically, Intel's biggest advantage with Arrow Lake vs Granite Ridge might be a process advantage, regardless of whether that's from N3 or 20A.

 

[tamz_msc]

Not so good news for Meteor Lake:

Intel Orders Delayed, TSMC Slows Three-Nanometer Expansion, Says TrendForce

According to TrendForce research, Intel plans to outsource the tGPU chipset in Meteor Lake to TSMC for manufacture. Mass production of this product was initially planned for 2H22 but was later postponed to 1H23 due to product design and process verification issues. Recently, the product’s mass...

www.trendforce.com

 

[DrMrLordX]

Not so good news for Meteor Lake:

Wow. That plus the Sapphire Rapids delay is a real kick in the pants. Plus TSMC has got to be charging Intel fees for all these delays.

 

[Zucker2k]

Plus TSMC has got to be charging Intel fees for all these delays.

What if the delay is a result of TSMC in this particular instance?

 

[jpiniero]

What if the delay is a result of TSMC in this particular instance?

Well, there's still Apple...

So it's looking like Raptor will get a full burn. And maybe we will get a full refresh too for desktop. If Intel thinks they can't even get acceptable yields of the tiny 40 mm2 chiplet on Intel 4 until the end of next year, who knows when they would be able to do anything bigger. This Arrow Lake on TSMC better be true.

 

[mikk]

According to TrendForce research, Intel plans to outsource the tGPU chipset in Meteor Lake to TSMC for manufacture. Mass production of this product was initially planned for 2H22 but was later postponed to 1H23 due to product design and process verification issues. Recently, the product’s mass production schedule has been postponed again to the end of 2023 for some reason

Intel Orders Delayed, TSMC Slows Three-Nanometer Expansion, Says TrendForce

According to TrendForce research, Intel plans to outsource the tGPU chipset in Meteor Lake to TSMC for manufacture. Mass production of this product was initially planned for 2H22 but was later postponed to 1H23 due to product design and process verification issues. Recently, the product’s mass...

www.trendforce.com

If this report is true it means no MTL in the market until late Q1-Q2 2024??? I mean for desktop it wasn't expected before 2024 and it's only low-midrange apparently but for the laptop market this is bad, really bad. Not even a tiny 2+8+64EU? It makes more and more sense that MTL and ARL are both 14th gen desktop and coming at roughly the same time. Unless ARL is not going to plan which can happen given Intels delay history recently.

It's not even a process node issue if the report are true. What are they doing with Intel 4 until then? Could they shift other products to Intel 4?

 

[Exist50]

Not so good news for Meteor Lake:

Intel Orders Delayed, TSMC Slows Three-Nanometer Expansion, Says TrendForce

According to TrendForce research, Intel plans to outsource the tGPU chipset in Meteor Lake to TSMC for manufacture. Mass production of this product was initially planned for 2H22 but was later postponed to 1H23 due to product design and process verification issues. Recently, the product’s mass...

www.trendforce.com

This is a bit of a confusing rumor. I certainly don't doubt that Meteor Lake could be delayed well into H2, but Meteor Lake doesn't use N3, so it can't directly affect TSMC's buildout. So the rumor is either referring to Arrow Lake delays (possibly indirectly from MTL slips) or it's trying to make a connection where there isn't one.

 

[tamz_msc]

This is a bit of a confusing rumor. I certainly don't doubt that Meteor Lake could be delayed well into H2, but Meteor Lake doesn't use N3, so it can't directly affect TSMC's buildout. So the rumor is either referring to Arrow Lake delays (possibly indirectly from MTL slips) or it's trying to make a connection where there isn't one.

Wasn't Meteor Lake supposed to use N3 for the iGPU tile, which was later rumored to use N5 according to Raichu?

 

[Karnak]

but Meteor Lake doesn't use N3,

It does. Slide is an official one from Intel 6 months ago. "External N3".

 

[Exist50]

Wasn't Meteor Lake supposed to use N3 for the iGPU tile, which was later rumored to use N5 according to Raichu?

No. There are rumors to that affect, but they're wrong. Raichu has it right.

It does. Slide is an official one from Intel 6 months ago. "External N3".

That N3 is for Arrow Lake. The only MTL tile Intel has officially confirmed is compute on Intel 4.

 

[Karnak]

N3 is for both of them.

 

[nicalandia]

N3 is for both of them.

Redwood Cove found on Meteor Lake is set in Stone for N4 no two ways about it. The core that Granite Rapids will use is on N3 perhaps a Redwood Cove+ or enhanced Redwood cove of sorts.

 

[Karnak]

It's not about the CPU part, the GPU tile is on N3 for both.

 

[Exist50]

N3 is for both of them.

Then do you also think MTL is using 20A? No, they just bundled the two together for that column.

 

[dullard]

It's not about the CPU part, the GPU tile is on N3 for both.

Read the Key Takeaways from this Intel slide for Arrow Lake: "Introduces LNC+SKT on N3, GT3 N3..." How do you introduce something on N3 that was already on N3 on the previous generation?

 

[João Bortolace]

N3 is for both of them.

On the old intel slide the MTL had 192EU, and it looks like it now has 128EU, intel may have ported tGPU to the N5.

 

[Exist50]

On the old intel slide the MTL had 192EU, and it looks like it now has 128EU, intel may have ported tGPU to the N5.

The leak above predates the slide with 196EUs, which was wrong from day 1. Intel marketing has been terrible about basic product details.

Also, there's the simultaneous rumor that Intel is shifting products onto N3 because of unfulfilled wafer obligations. Obviously it's unlikely for both to be true.

 

[shady28]

The leak above predates the slide with 196EUs, which was wrong from day 1. Intel marketing has been terrible about basic product details.

Also, there's the simultaneous rumor that Intel is shifting products onto N3 because of unfulfilled wafer obligations. Obviously it's unlikely for both to be true.

"Isaiah outlined that TSMC had initially planned to churn out 15,000 to 20,000 3nm wafers per month by the end of 2023 but this has now been reduced to 5,000 to 10,000 wafers per month."

So 10K wafers per month is roughly 6M 100mm2 chips per month. That's actually not very many considering how big the phone and x86 markets are. Just the x86 CPU market itself is about 5X that yearly output.

The end of 2023 number is not anywhere near enough volume to satisfy the major customers we know about - Apple M2 Pro and M3, Apple A17 for the 2023 iPhones which use it (bifurcation on Apple's lineup with 'new' phones using old processors starting this year), and Intel's MTL GPU tile.

It looks to me like N3 is not going to be a thing outside a few high end iPhones until 2024.

 

[mikk]

Apparently Intel denied the story.

https://twitter.com/x/status/1555297656183762944

 

[shady28]

Something I noticed today, maybe because of the unofficial/official release date announcements for Zen 4 and RPL, but I think DDR5 is starting to sell out.

I've been going back on forth on getting ADL on sale or waiting for RPL, doing builds and so on online for fun and reading memory reviews and so on, and just noticed a sudden reduction in the available selection of DDR5 kits.

Prices are still good on what's there, for now, but not quite as good as last week... Just anecdotal.