Intel Meteor Lake Technical Deep Dive

[Chrispy_]

With 14th-gen Raptor Lake refresh imminent, the timing on this reveal is a bit weird. Are we getting Meteor Lake laptops before 15th-Gen?

The cynic in me is wondering of Intel is just releasing all of this now to trick some investors into thinking that the next gen CPUs about to go on sale are Meteor Lake.

 

[DavidC1]

With 14th-gen Raptor Lake refresh imminent, the timing on this reveal is a bit weird. Are we getting Meteor Lake laptops before 15th-Gen?

The cynic in me is wondering of Intel is just releasing all of this now to trick some investors into thinking that the next gen CPUs about to go on sale are Meteor Lake.

Huh?

You know both Raptorlake Refresh and Meteorlake is under the new Core without the i branding right? The Ultra is only for Meteorlake.

 

[watzupken]

I bet we'll see this in a lot of NUC like devices. That new iGPU might finally give AMD some competition on that front.

I think it is good to have competition, but I feel Intel's GPUs are still not there yet. It is a great first attempt at the dGPU space, but driver issues will drag it down big time.

 

[theouto]

The three tiered core setup sounds like latency hell honestly, and a great way to confuse programs or windows on what to even use.

Very disappointed that this is pretty much "guys, look, AI, now give us investor money"
Oh well, so long as the efficiency improvements are substantial then this should still be alright, on a thin and light laptop efficiency is what most care about, and last I checked that wasn't doing so well

 

[las]

I hope they will be able to use 20A process for Arrow Lake, hopefully 16th gen in about a year from now. 20A should be comparable to TSMC N3/3nm.

Samsung might also take a leap with its 3nm and 4nm nodes, not sure about yields tho

 

[Chrispy_]

Huh?

You know both Raptorlake Refresh and Meteorlake is under the new Core without the i branding right? The Ultra is only for Meteorlake.

What on earth does the branding have to do with my question?

 

[Zareek]

I think it is good to have competition, but I feel Intel's GPUs are still not there yet. It is a great first attempt at the dGPU space, but driver issues will drag it down big time.

Supposedly, Intel graphics drivers have improved leaps and bounds. I don't currently own any newer Intel products, so I can't really confirm that.

Tiger Lake's iGPU was actually a bit faster than AMD's Vega iGPU at the time. But Alder Lake and Raptor Lake use the exact same iGPU as Tiger Lake so yeah this will be the first time in a while it's improved and with a 2x improvement it should rival the current RDNA3 iGPU. Intel's lower trims are also less cut down than AMD's, so certain i5 and i3 models I think still compete very favorably against AMD.

Maybe so but what about the drivers? If you listen to Nvidia fanboys, AMD drivers are bad. I've personally owned as many AMD cards as Nvidia cards over the past 25+ years. They both have had times when their drivers lacked polish. In my experience, Intel's drivers have never had polish, they just plain stink. That being said, I haven't had many experiences with them over the past 3 to 5 years, so hopefully things have changed.

 

[RandallFlagg]

I hope they will be able to use 20A process for Arrow Lake, hopefully 16th gen in about a year from now. 20A should be comparable to TSMC N3/3nm.

Samsung might also take a leap with its 3nm and 4nm nodes, not sure about yields tho

Intel 4 is comparable to TSMC N3E. Intel 4 is basically Intel's high-power node, whose equivalent is TSMC's N3E, while N3 is the low power (and more dense) version from TSMC. Hence, you'll see N3E in high power desktop/laptop, but not N3.

Intel also has Intel 3 node, which is supposed to be just for IDM aka their foundry business. That is a direct competitor to TSMC N3 as it is lower power and more dense than Intel 4, but AFAIK Intel does not plan to use it internally as they are going to 20A.

20A, if it releases on schedule (end of 2024) with Arrow Lake, will be ahead of TSMCs N2 node by about 3 quarters. It will be in a different class from TSMC N3 / N3E.

So that would be the first time since N7 released in 2019 that Intel's fab node was ahead of TSMC.

All that said it's pretty clear that Meteor Lake laptops will be impressive, especially for those who rely on the iGPU and/or don't care too much about dGPU - which is the majority of that market. Let's just hope the price tag isn't exorbitant, but I suspect it will be.

 

[PlatinumProclaim]

Thanks for always including the full slides! But I guess these are from a PDF, you couldn't just upload that too?

 

[Squared]

Maybe so but what about the drivers? If you listen to Nvidia fanboys, AMD drivers are bad. I've personally owned as many AMD cards as Nvidia cards over the past 25+ years. They both have had times when their drivers lacked polish. In my experience, Intel's drivers have never had polish, they just plain stink. That being said, I haven't had many experiences with them over the past 3 to 5 years, so hopefully things have changed.

I use an Nvidia GT 1030 DDR4 in my HTPC and it doesn't sees only a little gaming. I use an AMD RX 5600 XT for gaming (before that RX 480 and before that Nvidia 750 Ti). And I have a Tiger Lake i5 laptop (80 Xe EUs) which has also seen only a little gaming. If I were to blame all the problems I've had on drivers, then I would rank AMD as best and Nvidia as worst. But I've seen very few issues on any of them that I would blame on drivers. However the Intel laptop renders a lighting effect in Age is Empires 4 that follows the camera and isn't really a problem but it's not suppose to be there. (My HTPC frequently disconnects from my TV and so I have to reconnect the HDMI cable, which is why I would rank Nvidia worst.)

 

[Steevo]

One thing I have learned about Intel is the bigger the slideshow, the more mediocre the product. I will wager that they will sell and only half the functions will ever be used, and 10% will become disabled through security patches.

There is a countdown to the first virus to use hardware AI, I could see a use for inference learning to mutate and either avoid detection or learn to disable detection covertly.

 

[DavidC1]

What on earth does the branding have to do with my question?

Meteorlake is only for mobile and Raptorlake Refresh which comes at almost the same time is desktop. And those will be further differentiated by branding. No confusion between the two. If they did not change the branding it's 14th Gen Meteorlake for mobile and 14th Gen Raptorlake Refresh for desktop.

Also if RPL-R is coming in October I doubt Arrowlake of any kind will come in couple of months. Then 15th gen will consist of Arrowlake and Lunarlake, neither of which we will see soon.

 

[las]

Meteorlake is only for mobile and Raptorlake Refresh which comes at almost the same time is desktop. And those will be further differentiated by branding. No confusion between the two. If they did not change the branding it's 14th Gen Meteorlake for mobile and 14th Gen Raptorlake Refresh for desktop.

Also if RPL-R is coming in October I doubt Arrowlake of any kind will come in couple of months. Then 15th gen will consist of Arrowlake and Lunarlake, neither of which we will see soon.

Arrow Lake should be 20A in Q4 2024.

Raptor Lake refresh is a gap closer but will probably last a full generation.

If Arrow Lake is delayed, I think Intel will be in trouble vs Zen 5 @ 3nm (Q2 2024) and later Zen 5 with 3D Cache (Q4 2024).

 

[R0H1T]

Zen5 isn't on 3nm, at least not on any of the "leaks" I remember.

 

[Chrispy_]

Meteorlake is only for mobile and Raptorlake Refresh which comes at almost the same time is desktop. And those will be further differentiated by branding. No confusion between the two. If they did not change the branding it's 14th Gen Meteorlake for mobile and 14th Gen Raptorlake Refresh for desktop.

Also if RPL-R is coming in October I doubt Arrowlake of any kind will come in couple of months. Then 15th gen will consist of Arrowlake and Lunarlake, neither of which we will see soon.

Do we have confirmation that Meteorlake is not coming to desktop?

Intel have historically shared silicon between mobile and desktop. They did this for:

2nd Gen Sandy Bridge
3rd Gen Ivy Bridge
4th Gen Haswell
5th Gen Broadwell
6th Gen Skylake
7th Gen Kaby Lake
8th Gen Coffee Lake
10th Gen Comet Lake
12th Gen Alder Lake

9th gen was a desktop only rebrand of 8th gen, owing to 10nm manufacturing failures
11th gen was a desktop only port of Rocket lake back to 14nm owing to 10nm manufacturing failures.
13th gen is just a rebrand of Alder Lake (source)

So, manufuacturing failures aside, Every generation of Intel going back 12+ years has been shared between desktop and mobile.

 

[Daven]

I looked over everything in the slides but I couldn’t find the following:

Any stated IPC and clock increase in P-cores. TPU says it but not the slides.

Any stated process node for all four tiles:

Compute - Intel 4
SoC - ? (commentors say TSMC)
GPU - TSMC ?
I/O - ?

Were some of the above mentioned verbally during Intel’s presentations?

Edit: Okay I found this article from a month ago.
https://videocardz.com/newz/intel-reveals-meteor-lake-p-6p8e-cpu-diagram-for-2023-laptops

Basically Meteor Lake is built by TSMC. And I hope this isn’t a further attempt by Intel to buy up TSMC capacity away from competitors.

 

[ToTTenTranz]

So where are the Adamantine Meteor Lake models? They're coming later?
Is Intel sandbagging until Strix Point and Strix Halo arrive?

https://twitter.com/i/web/status/1672626063095255045

 

[AnotherReader]

Do we have confirmation that Meteorlake is not coming to desktop?

Intel have historically shared silicon between mobile and desktop. They did this for:

2nd Gen Sandy Bridge
3rd Gen Ivy Bridge
4th Gen Haswell
5th Gen Broadwell
6th Gen Skylake
7th Gen Kaby Lake
8th Gen Coffee Lake
10th Gen Comet Lake
12th Gen Alder Lake

9th gen was a desktop only rebrand of 8th gen, owing to 10nm manufacturing failures
11th gen was a desktop only port of Rocket lake back to 14nm owing to 10nm manufacturing failures.
13th gen is just a rebrand of Alder Lake (source)

So, manufuacturing failures aside, Every generation of Intel going back 12+ years has been shared between desktop and mobile.

Since 14 nm, older Intel processes have allowed processors to clock much higher than the first iteration of their new process. Broadwell didn't clock as high as Haswell and Ice Lake failed to approach the clocks of the later Skylake iterations. It took until Tiger Lake to approach those frequencies. It's rumoured that Intel 4 can't deliver clock frequencies as high as the final variant of Intel 7. Since Meteor Lake doesn't seem to be breaking any new ground on the IPC front, it wouldn't make sense to bring it to the desktop where its relatively low frequencies would hinder it against the competition.

 

[Squared]

Since 14 nm, older Intel processes have allowed processors to clock much higher than the first iteration of their new process. Broadwell didn't clock as high as Haswell and Ice Lake failed to approach the clocks of the later Skylake iterations. It took until Tiger Lake to approach those frequencies. It's rumoured that Intel 4 can't deliver clock frequencies as high as the final variant of Intel 7. Since Meteor Lake doesn't seem to be breaking any new ground on the IPC front, it wouldn't make sense to bring it to the desktop where its relatively low frequencies would hinder it against the competition.

Broadwell might not be the best example of a new node struggling to reach high clock speeds. Skylake was released just a few weeks later on the same node and beat Haswell Refresh in clock speed. Moreover the top Broadwell model (5775C) had a low TDP, huge integrated graphics, and an embedded DRAM cable, so there were a lot of things limiting its clock speed but it still came within 200MHz of the 4770K.

The Intel 6th Gen Skylake Review: Core i7-6700K and i5-6600K Tested

www.anandtech.com

 

[DavidC1]

Arrow Lake should be 20A in Q4 2024.

Raptor Lake refresh is a gap closer but will probably last a full generation.

If Arrow Lake is delayed, I think Intel will be in trouble vs Zen 5 @ 3nm (Q2 2024) and later Zen 5 with 3D Cache (Q4 2024).

Do you remember the Arrowlake performance leak a little while ago? It was disappointing.

Apparently not only it'll have N3 and 20A versions, but Redwood Cove + Crestmont version and Lion Cove + Skymont version. The ISA documents also support this with -S and non-S models supporting different instructions. Credible leakers pointed to -P mobile having 20A and -S being on N3. Incredibly, some even say low power "Arrowlake" is potentially Meteorlake but on Intel 3. But that won't really matter as Lunarlake is the Real Thing for low power anyway.

So where are the Adamantine Meteor Lake models? They're coming later?
Is Intel sandbagging until Strix Point and Strix Halo arrive?

No ADM in Meteorlake. Kepler_L2 has said ALL ADM models are cancelled. Suffice to say extra cache module layer would be a waste of money on 8 Xe core part since AMD is also doing fine without it.

Any stated IPC and clock increase in P-cores. TPU says it but not the slides.

The only thing that seems noticeable for P is the doubled L1i cache to 64KB. It's definitely a departure from the server P core in Granite Rapids: https://images.anandtech.com/doci/20034/HotChips 2023 Press Briefing Final__07.png

The E cores are different too.
Crestmont client: https://images.anandtech.com/doci/2... Power Efficient Processor_FINAL CLEAN-13.png

Sierra Glen server: https://images.anandtech.com/doci/20034/HotChips 2023 Press Briefing Final__08.png

Most notable, Crestmont gets Allocate/Rename/Move Elimination/Zero Idiom increased to 6-wide while Sierra Glen stays at 5. Chipsandcheese overview shows that at longer loop length, the decode throughput is limited by the renamer: https://chipsandcheese.com/2021/12/21/gracemont-revenge-of-the-atom-cores/

So Crestmont in Meteorlake should get full 6-wide throughput. Slides also mention branch prediction on Crestmont but not on Sierra Glen. Apparently Sierra Glen prioritizes higher clocks and Crestmont, higher perf/clock.

 

[AnotherReader]

Broadwell might not be the best example of a new node struggling to reach high clock speeds. Skylake was released just a few weeks later on the same node and beat Haswell Refresh in clock speed. Moreover the top Broadwell model (5775C) had a low TDP, huge integrated graphics, and an embedded DRAM cable, so there were a lot of things limiting its clock speed but it still came within 200MHz of the 4770K.

The Intel 6th Gen Skylake Review: Core i7-6700K and i5-6600K Tested

www.anandtech.com

It isn't the best example, but in those days, Intel didn't get to spend much time optimizing for high clock speeds. Also, Skylake was nearly a year after the first 14 nm processor: Broadwell-Y.

 

[ToTTenTranz]

No ADM in Meteorlake. Kepler_L2 has said ALL ADM models are cancelled. Suffice to say extra cache module layer would be a waste of money on 8 Xe core part since AMD is also doing fine without it.

Phoenix's 780M has a large 2MB L2 cache (for its size) and it's still bandwidth limited, though.
I wasn't aware that Adamantine Meteor Lake was cancelled. Wouldn't it also increase single-threaded performance in games, though? It did so in Broadwell.

 

[DavidC1]

Phoenix's 780M has a large 2MB L2 cache (for its size) and it's still bandwidth limited, though.
I wasn't aware that Adamantine Meteor Lake was cancelled. Wouldn't it also increase single-threaded performance in games, though? It did so in Broadwell.

2MB is small. Iris Xe had 3.8MB L3 cache since Tigerlake(of course nothing changed in Raptorlake). Alchemist A380 has 4MB L3.

I am saying that Phoenix already performs ballpark in the range of Meteorlake's GPU and it doesn't have dedicated cache chip of any kind.

Also, what do you mean by "bandwidth-limited" for Phoenix? Cause that definition varies among people. For example, older well-balanced GPUs had a performance to bandwidth ratio of 0.3:1, meaning for every 1% increase in bandwidth, you gained 0.3%. So if it's less than that it's unbalanced because the memory subsystem is overbuilt and if the gain is greater than the memory subsystem is underbuilt.

Going from DDR5-5600 to LPDDR5x-7500 is about 34% increase in bandwidth, and if we take the 1:3 rule than you'd gain 11%. I do remember on older AMD desktop APUs, the gain would be way greater than that(like a 1:2 or even 3:5 ratio) and after it reaches a certain point, the requirements drop off drastically, like to 1:6, which is a bit strange.

The ADM cache might be justified on even higher end setups, but since we need to wait until Arrowlake to get real big, the BoM increase is not worth it. The improvement was rarely about CPU performance, since even eDRAM is still too high of a latency to benefit uniformly. We know HBM is useless on the client CPU side, and that's entirely about the GPU, and you need at least eDRAM of some kind to have lower latency. Based on rumored size ADM sounds like it's continuation of the eDRAM but on the base Foveros die.

Laptop reviews with eDRAM parts didn't really show noticeable advantage except in graphics. I remember the battery life suffered as well. They minimized the loss with Broadwell parts but earlier versions were a terrible loss on battery life. That's cause the 1st Gen eDRAM had a 1W standby power while second generation reduced that to 1/4th, and the reason was because the 2nd gen reduced the refresh rate of the eDRAM cells thus it needed less power to maintain data.

 

[Borc]

Disappointing that they neutered the AV1 encoder to 4:2:0. I was expecting them to make AV1 ubiquitous but they left room for improvement for future generations in classic Apple style.

Why neutered? Alchemist dGPU is also 4:2:0.

 

[Squared]

It isn't the best example, but in those days, Intel didn't get to spend much time optimizing for high clock speeds. Also, Skylake was nearly a year after the first 14 nm processor: Broadwell-Y.

Right I forgot all about Broadwell-Y, I think I only remembered it as Core M.