Discussion Intel Meteor, Arrow, Lunar & Panther Lakes Discussion Threads

[Tigerick]

As Hot Chips 34 starting this week, Intel will unveil technical information of upcoming Meteor Lake (MTL) and Arrow Lake (ARL), new generation platform after Raptor Lake. Both MTL and ARL represent new direction which Intel will move to multiple chiplets and combine as one SoC platform.

MTL also represents new compute tile that based on Intel 4 process which is based on EUV lithography, a first from Intel. Intel expects to ship MTL mobile SoC in 2023.

ARL will come after MTL so Intel should be shipping it in 2024, that is what Intel roadmap is telling us. ARL compute tile will be manufactured by Intel 20A process, a first from Intel to use GAA transistors called RibbonFET.





As mentioned by Tomshardware, TSMC will manufacture the I/O, SoC, and GPU tiles. That means Intel will manufacture only the CPU and Foveros tiles. (Notably, Intel calls the I/O tile an 'I/O Expander,' hence the IOE moniker.)

 

[Geddagod]

Source

It’s around the 1:04 mark in the video. Some of the press materials also explicitly say same performance at half thr power consumption of RPL.

you just posted a dot lol

 

[H433x0n]

you just posted a dot lol

Yeah, it was an accident. Dont know how to delete posts (if we can at all).

 

[Abwx]

If Intel had no E-cores, they a 6+8 chip would be a 8 big core chip. A 6+4 would be a 6 big core chip. Simply from the floorplan of the CPU itself.
Comparing a 6+4 CPU vs a 8 core CPU is just unfair to the Intel CPU, because it's "8 core competitor" is literally a 6+8 product.

They compare the 13900K to the 7950X, 8 + 16 vs 16/32T, so that s straightforward that 4 + 8 is comparable to 8/16T.

It s just that they are late on the process efficency, it can work for DT chips where they can inflate the TDP but not for mobile where efficency is the metric of choice.

 

[Geddagod]

As for how much Intel is behind on power efficiency vs AMD in mobile, this is a neat chart from notebookcheck

Looks to be anywhere from 15-25% less efficient than Phoenix, so ye, I think MTL at the very least can get close Phoenix

 

[Geddagod]

They compare the 13900K to the 7950X, 8 + 16 vs 16/32T, so that s straightforward that 4 + 8 is comparable to 8/16T.

8+16 isn't comparable to 16C, it's just that Intel does not want to spend the extra die space to actually compete on core counts in desktop, since you know, no one really cares that much about power consumption in desktop.
I mean literally, go look at a die shot of ADL/RPL. One 4 core cluster is around the same area as 1 big core cluster. Comparing 2 e-cores to 1 P core is just trying to make Intel look worse than it is. For Christ's sake, 4 e-cores literally take up one ringstop as well on the ringbus. If Intel had no E-cores, then they would just insert a P core into the slot where those 4 e-core used to go too and it would fit pretty well.
Actually, you don't even have to look. Took the 10 secs to pull up a die shot for you:

And here's a die shot of MTL. Would you look at that, the ratio of 4 e-cores being around the same size of a P-core STILL holds true-

 

[Geddagod]

I am of the same opinion that the same RedwoodCove(x86) is in GraniteRapids and MeteorLake.The difference in the presentation is that in the case of GraniteRapids Intel announced changes in the microarchitecture, and in the case of MeteorLake only that it is a new P Core.I believe that since Intel calls RedwoodCove a new core, it means that it is a generational jump and I assume an IPC jump in the range of 8-15%.

I am sure that more details, including numbers, will be provided on or before December 14.

Nope, in the hothardware tech interview with Intel, Intel themselves calls RWC esentially a ported GLC. There is no substantial IPC increase.

 

[Abwx]

As for how much Intel is behind on power efficiency vs AMD in mobile, this is a neat chart from notebookcheck
View attachment 86177
Looks to be anywhere from 15-25% less efficient than Phoenix, so ye, I think MTL at the very least can get close Phoenix

Even on those charts, 13700H@55W is matched by the 7940H@35W, not what i would call 15-25%, 7940H is 57% more efficient at this score..

That s in line with my previous number that the 1360P is half as efficient than the 7840U since the 13700H has more cores and has indeed better perf/watt at same score than the 4 + 8.

You are comparing orange and apples to get your flawed numbers, efficency should be compared at same throughput.

 

[HoveringStyle]

Slightly faster than 780M on synthetics at least.

If this slide is accurate enough to be measured


...it can clock up to 2.4GHz, for 1024 shaders that'll be 4.9 TFLOPs. Time Spy's graphic score against flops is very linear on Iris Xe, so with the help of a simple regression equation, it looks like Meteor Lake has a high chance of breaking the 4000 mark on it (780M tops out at ~3100).

 

[lightisgood]

Nope, in the hothardware tech interview with Intel, Intel themselves calls RWC esentially a ported GLC. There is no substantial IPC increase.

Agree.
Intel says that Redwood Cove is focused on efficiency, not IPC.
BTW, Golden Cove is high performance but power hungry core especially in over 4.2GHz zone.
I think that Intel's decision, not IPC but more efficiency, is make sense.

 

[H433x0n]

Even on those charts, 13700H@55W is matched by the 7940H@35W, not what i would call 15-25%, 7940H is 57% more efficient at this score..

That s in line with my previous number that the 1360P is half as efficient than the 7840U since the 13700H has more cores and has indeed better perf/watt at same score than the 4 + 8.

You are comparing orange and apples to get your flawed numbers, efficency should be compared at same throughput.

So when MTL comes out with better perf/watt than Phoenix it won’t count since it’s using a 6P+8E configuration?

 

[jpiniero]

Better iGPUs remove the need for dGPUs at the bottom of the stack. Intel iGPUs were very basic, now they can do light gaming. Not saying they are going to replace a 4050 but anything below seems now redundant, and we used to have the MX series and 1030s as dGPUs on mobile. Not arguing with the rest of your point.

But it's going to be slower than the 3050 6 GB... and while the GDDR6 memory cost is an issue, the chip being on a cheap node helps even it out and then some. nVidia is running out of cheap nodes though... so who knows... They might be at the end of the line.

 

[Abwx]

So when MTL comes out with better perf/watt than Phoenix it won’t count since it’s using a 6P+8E configuration?

According to the NBC curves above it need to be about 60% more efficient than the 13700H to match a 7940H efficency at 30W, that s the perf/watt improvement required by the process they ll use.

 

[Geddagod]

Even on those charts, 13700H@55W is matched by the 7940H@35W, not what i would call 15-25%, 7940H is 57% more efficient at this score..

That s in line with my previous number that the 1360P is half as efficient than the 7840U since the 13700H has more cores and has indeed better perf/watt at same score than the 4 + 8.

You are comparing orange and apples to get your flawed numbers, efficency should be compared at same throughput.

Oops.
Intel needs 15-25% more performance per watt than AMD to catch up. Something pretty doable, considering AMD did a similar jump between Zen 4 and Zen 3+.
But comparing the 1360p to the 7840u is a unfair comparison, one should compare whatever the 6+8 Intel sku is vs an 8 core AMD one.

 

[Geddagod]

If this slide is accurate enough to be measured
View attachment 86180

...it can clock up to 2.4GHz, for 1024 shaders that'll be 4.9 TFLOPs. Time Spy's graphic score against flops is very linear on Iris Xe, so with the help of a simple regression equation, it looks like Meteor Lake has a high chance of breaking the 4000 mark on it (780M tops out at ~3100).

AFAIK Intel always overperforms in synthetics vs gaming (perhaps due to poor drivers, maybe due to some other reasons). But still interesting

 

[H433x0n]

According to the curves above it need to be about 60% more efficient than the 13700H to match a 7940H efficency at 30W, that s the perf/watt improvement required by the process they ll use.

Or - it could have +20% better performance at iso power (which is exactly what Intel 4 is advertised with) then voila it has superior perf/watt across entire curve.

 

[Abwx]

Oops.
Intel needs 15-25% more performance per watt than AMD to catch up. Something pretty doable, considering AMD did a similar jump between Zen 4 and Zen 3+.
But comparing the 1360p to the 7840u is a unfair comparison, one should compare whatever the 6+8 Intel sku is vs an 8 core AMD one.

WIth 25% better perf/watt their 13700H will need 55/1.25 = 44W to match the 13700H@55W.

We are still far from 35W, it s 57% better perf/watt that is required to match the 7970H@35W, namely 55/1.57 = 35W.

Is the math clear now..?..

 

[Abwx]

Or - it could have +20% better performance at iso power (which is exactly what Intel 4 is advertised with) then voila it has superior perf/watt across entire curve.

20% better perf/isowatt allow for about 50% better perf/watt at isoperf, quite better but it wont beat Phoenix, notice that Intel stated that it will be "their most efficent mobile SKU ever", not "the most efficient mobile SKU ever", notice the nuance in their wording...

 

[Geddagod]

WIth 25% better perf/watt their 13700H will need 55/1.25 = 44W to match the 13700H@55W.

We are still far from 35W, it s 57% better perf/watt that is required to match the 7970H@35W, namely 55/1.57 = 35W.

Is the math clear now..?..

Umm, the 13700h scores 12000 CB R23 MT points at 35 watts. 25% better perf/watt would be 12000 x 1.25 = 15,000 points.
The 7940hs scores 14000 CB R23 MT points at 35 watts.
a 25% increase in perf/watt should be enough for MTL to reach Phoenix levels of efficiency.

 

[jpiniero]

Of course Meteor Lake is coming to the desktop in 2024. But it's gonna get a new name... Arrow Lake.

Nah it's probably similar to the B line that Tiger Lake got. Leaky stuff for 65+W AIOs.

 

[H433x0n]

WIth 25% better perf/watt their 13700H will need 55/1.25 = 44W to match the 13700H@55W.

We are still far from 35W, it s 57% better perf/watt that is required to match the 7970H@35W, namely 55/1.57 = 35W.

Is the math clear now..?..

That’s not how it works, we’ve had this exact same conversation before. I guess it doesn’t matter, we’ll find out in December!

 

[Geddagod]

20% better perf/isowatt allow for about 50% better perf/watt at isoperf, quite better but it wont beat Phoenix, notice that Intel stated that it will be "their most efficent mobile SKU ever", not "the most efficient mobile SKU ever", notice the nuance in their wording...

25% better perf/watt indicates a 25% increase in performance iso wattage. What you're looking at is % lower wattage iso performance. Two different things.

 

[lightisgood]

Nah it's probably similar to the B line that Tiger Lake got. Leaky stuff for 65+W AIOs.

Intel already sold their NUC business.
They have no motivation for releasing like 11900KB product.

I think that MTL-S is simply aliving.
Rumors are failure.
We saw same thing when Gelsinger declared that 288C Sierra Forest is real product.

Granite Rapids (Redwood Cove) is made by Intel 3 and coming in mid24 time frame.
So, I expect that MTL-S is also made by Intel 3 too and possibly coming earlier than GNR.

To accomplish the insane strategy, 5 nodes in 4 years, is providing insane harvest now.

 

[jpiniero]

Intel already sold their NUC business.
They have no motivation for releasing like 11900KB product.

AIOs, not NUC. While they can (and do) use straight up mobile parts in AIOs... there's room for leaky stuff too if the OEM desires and Intel offers a nice discount.

Also they could make a BGA package with the 'desktop' 10 nm SoC instead of the TSMC ones.

 

[Hulk]

I'm very skeptical that Meteor will be released for the desktop in a meaningful way. If it is released it will be like Broadwell. Looking back I don't think anyone really considers Broadwell a desktop release. Haswell was simply clocking too high, think Devil's Canyon for the immature 14nm Broadwell node to compete.

We are in a very similar situation now except that Meteor Lake isn't competing with the 2nd gen Alder Lake (Raptor), it's competing with the 3rd gen (Raptor Refresh), which has pushed clocks even higher. Meteor, with minimal IPC gains, if any, on a new node will not be able to compete from a performance point of view. In addition, Arrow Lake is expected next year and would release concurrently or there about. It doesn't make sense.

I will say this right now, there will be no Meteor Lake part that will compete with the 14900K from a pure performance point of view on the desktop.

I'm not going to "un-write" anything I wrote above but if anything I guess it would be possible for MTL-S to fill in some of the low end of the 14th gen stack but I just don't see how it makes sense.

Finally, the remark that Intel VP said was very nebulous. When prodded to talk about keeping mobile and desktop releases on the same core she said something like "That's the goal." I read that meaning for 15th gen and/or beyond.

 

[Hulk]

WIth 25% better perf/watt their 13700H will need 55/1.25 = 44W to match the 13700H@55W.

We are still far from 35W, it s 57% better perf/watt that is required to match the 7970H@35W, namely 55/1.57 = 35W.

Is the math clear now..?..

It depends on which part of the performance spectrum you cherry pick your power figures. If you look at a score of 16,000 then Intel's only behind by about 18%. Either way, I agree Intel needs a giant leap in efficiency to catch up never mind pull ahead.