Discussion Intel current and future Lakes & Rapids thread

[Markfw]

I am sorry that you are sick of it and yet I still will comment. But, I think there is a fundamental point that keeps getting missed. Intel E-cores (when compared to Intel P-cores) perform their best in low power per core situations. The E-cores perform terribly in higher power per core situations. What power level will each core in Clearwater Forest have? Roughly 1 W to 2 W each (give or take depending on model)? Now look at E-core vs P-core performance near that 1 W to 2 W power per core:
View attachment 94357

https://i0.wp.com/chipsandcheese.com/wp-content/uploads/2022/01/image-16-1.png?ssl=1

Yes, you are correct that E-cores have been over-defended and often over-hyped. But large core count situations is where E-cores really can shine. If AVX-512 is needed, the E-cores are the wrong way to go. But, otherwise I don't think you can compare what you think of high power per E-core performance to what will happen at low power per E-core. In other words, forget whatever you think you know about E-cores until these chips come out.

I don't have any stats to back this up, but my GUESS based on what we know about P-cores and C cores is that 1 C-core ~= 3 E-cores , and that does not include avx-512. C-cores are also very power efficient and space efficient. So 288 E-cores ~= 96 C-cores. With technology at current levels.

As I said, I will wait to argue any more until an actual product comes out and is benchmarked against the current competition.

Edit: performance wise. And even at 2.5 to 1 thats 115 cores. power wise we need to wait until chips are out.

 

[Tigerick]

Physically only has 288 cores

https://twitter.com/x/status/1760987514754789413

Hmm, so it is 24 e cores per compute tile, i.e.. 96 e cores per base tile (with memory controllers?) total 288 e cores. That's rather odd numbers, I am sure Intel will cut certain numbers of core to ensure better yield?

BTW, based on timeline, we should be expecting CWF by end of 2025..

 

[dullard]

I don't have any stats to back this up, but my GUESS based on what we know about P-cores and C cores is that 1 C-core ~= 3 E-cores , and that does not include avx-512....With technology at current levels.

The very point of my post is that ratios like you are using are only valid at certain power levels per core. They are invalid at lower power levels. Technology won't be ran at current levels. So any form of rough math that ignores that difference of lower power per core will be vastly inaccurate.

 

[SiliconFly]

The very point of my post is that ratios like you are using are only valid at certain power levels per core. They are invalid at lower power levels. Technology won't be ran at current levels. So any form of rough math that ignores that difference of lower power per core will be vastly inaccurate.

And the ratios he's using is limited only to certain generation of cores.

 

[igor_kavinski]

In other words, forget whatever you think you know about E-cores until these chips come out.

Hopefully, ChipsandCheese can do a similar graph comparing Zen 4 and Zen 4c soon.

 

[DavidC1]

Interesting how the compute tile will be so small. At 24 cores at most it might end under 50mm2. Actually 24x 1mm2 core might mean 35mm2 tile adding generously to non core components.

I don't have any stats to back this up, but my GUESS based on what we know about P-cores and C cores is that 1 C-core ~= 3 E-cores , and that does not include avx-512. C-cores are also very power efficient and space efficient. So 288 E-cores ~= 96 C-cores. With technology at current levels.

Where did you get 1:3 from? The performance gap of P and E is 2. And that 2x is broken down into:
-40-50% combined PPC
-Hyperthreading
-Clock gap

Further, SpecCPU tests show the combined 40% is split into 25% in Integer and 50-60% in FP. A cloud CPU won't need powerful FP. 288 core SKU is also the top chip succeeding the 288 core SRF. There will be a smaller core version. 176 is the current rumor if I remember correctly.

 

[trivik12]

Rumor is CWF uses Darkmont. So that is 2 gen jump from crestmont+ that SRF will use. That said we have to wait until we see retail benchmarks for SRF. That should be sometime this year for sure as Intel specifically mentioned H1 2024.

 

[DavidC1]

Rumor is CWF uses Darkmont. So that is 2 gen jump from crestmont+ that SRF will use. That said we have to wait until we see retail benchmarks for SRF. That should be sometime this year for sure as Intel specifically mentioned H1 2024.

Darkmont sounds like a Crestmont like small revision, so it's really 1 gen.

 

[DrMrLordX]

That said we have to wait until we see retail benchmarks for SRF.

Hopefully we'll see plenty of samples reviewed (including the 288c machines), but Sierra Forest may not be a very high-volume product.

 

[jpiniero]

Hopefully we'll see plenty of samples reviewed (including the 288c machines), but Sierra Forest may not be a very high-volume product.

I still think it's nothing more than a test vehicle to see if YOLOing Granite Rapids is even plausible.

 

[DrMrLordX]

I still think it's nothing more than a test vehicle to see if YOLOing Granite Rapids is even plausible.

If that's true then Granite Rapids is going to be a pretty late product to the table.

 

[SiliconFly]

If that's true then Granite Rapids is going to be a pretty late product to the table.

Since Granite Rapids is still based on the seriously outdated power hungry redwood cove cores, I don't think it'll stand up to competition. Mostly DOA like previous 2 gens.

Diamond Rapids is a exciting product and might stand a chance as it's based on a newer node with a newer architecture.

 

[DrMrLordX]

Diamond Rapids is a exciting product and might stand a chance as it's based on a newer node with a newer architecture.

It's hard to get excited after yet another lackluster datacentre product.

 

[SiliconFly]

It's hard to get excited after yet another lackluster datacentre product.

Yep. Intel has a serious problem with its data center offerings.

 

[ashFTW]

Yep. Intel has a serious problem with its data center offerings.

GNR AP will likely double the cores compared to EMR (5th gen). It may still not match EPYC in ST or MT at the very high end, but it’s still a huge step forward for Intel DCG. GNR also has nice on chip accelerators. SRF with 144/288 cores should compete very well against Bergamo and various ARM offerings. Also note that most sales don’t happen at the highest end SKUs. Things look quite promising to me, provided they don’t have massive delays like they did with SPR.

 

[SiliconFly]

GNR AP will likely double the cores compared to EMR (5th gen). It may still not match EPYC in ST or MT at the very high end, but it’s still a huge step forward for Intel DCG. GNR also has nice on chip accelerators. SRF with 144/288 cores should compete very well against Bergamo and various ARM offerings. Also note that most sales don’t happen at the highest end SKUs. Things look quite promising to me, provided they don’t have massive delays like they did with SPR.

TBH, I was only mentioning about how GNR stands up against competition. Intel needs something a lot better imho.

On the other hand, SRF 288 sounds good.

 

[Markfw]

Yep. Intel has a serious problem with its data center offerings.

I can't believe I am agreeing with you, but yes, there has been no decent data center products in like 8 years !

I just got rid of my last 2 Xeons in 2023, that at the time were great. 2699 v4's.

 

[SarahKerrigan]

Interesting how the compute tile will be so small. At 24 cores at most it might end under 50mm2. Actually 24x 1mm2 core might mean 35mm2 tile adding generously to non core components.

Where did you get 1:3 from? The performance gap of P and E is 2. And that 2x is broken down into:
-40-50% combined PPC
-Hyperthreading
-Clock gap

Further, SpecCPU tests show the combined 40% is split into 25% in Integer and 50-60% in FP. A cloud CPU won't need powerful FP. 288 core SKU is also the top chip succeeding the 288 core SRF. There will be a smaller core version. 176 is the current rumor if I remember correctly.

While you're correct and I think 1:3 is a very bad take, a lot of things come down to what Sierra Forest clock speed looks like, and there's not a clear answer to that question yet AFAIK.

288c @ 1.9GHz would not necessarily be particularly whelming, while 288c @ 3GHz would be solid.

 

[SiliconFly]

While you're correct and I think 1:3 is a very bad take, a lot of things come down to what Sierra Forest clock speed looks like, and there's not a clear answer to that question yet AFAIK.

288c @ 1.9GHz would not necessarily be particularly whelming, while 288c @ 3GHz would be solid.

While 288C sounds exciting, they're still crestmont cores (which in turn are rehashed gracemont cores). So, performance wise, it's not gonna be much.

On the other hand, clearwater forest looks good. Has nextgen e-cores, more advanced packaging, and also has adm cache on base die. That should be competitive both in performance & efficiency.

 

[DavidC1]

288c @ 1.9GHz would not necessarily be particularly whelming, while 288c @ 3GHz would be solid.

While 288C sounds exciting, they're still crestmont cores (which in turn are rehashed gracemont cores). So, performance wise, it's not gonna be much.

Reminder. Intel's Hot Chips presentation showed that they expect Sierra Forest's performance/watt to exceed Granite Rapids, while some variants will have single threaded performance equalling Sapphire Rapids.

At 2.4x perf/watt over SPR, you can expect a chip that's 40% faster while only at 205W TDP.

GNR leaked slides show Intel expects 40% perf and perf/watt improvement over EMR for the -SP version with the 80 core, while they expect a 500W, 120 core -AP to beat 350W EMR by 2x - a similar 40% perf/watt improvement.

Since EMR is 20% over SPR, GNR is ~1.7x over SPR. For the intended workload(Cloud) it sounds pretty damn good. Raichu also claimed a while back Sierra Glen lacking the 6-wide rename/allocate(unlike Crestmont) is because they are focusing on clocks, so I don't expect it to be that low either.

 

[SiliconFly]

...while some variants will have single threaded performance equalling Sapphire Rapids.

Thats a bit hard to believe. On a related note, a techpowerup article says SRF 288 uses a "derivative" of Crestmont. Any ideas what they actually mean? Is it an upgraded version (Crestmont+) or is it exactly the same Crestmont cores found in MTL just adapted to SRF on Intel 3? Cos' some of the leaked benchmarks were a bit appalling.

 

[Dayman1225]

Thats a bit hard to believe. On a related note, a techpowerup article says SRF 288 uses a "derivative" of Crestmont. Any ideas what they actually mean? Is it an upgraded version (Crestmont+) or is it exactly the same Crestmont cores found in MTL just adapted to SRF on Intel 3? Cos' some of the leaked benchmarks were a bit appalling.

It’s just Crestmont with additional features/instructions only useful for servers AFAIK.

 

[semiman]

It’s just Crestmont with additional features/instructions only useful for servers AFAIK.

In Knight's landing, Intel used AVX512 stuffed hoards of atoms. Should be similar to how it was implemented.

 

[dullard]

Considering there are already announced products and benchmarks with MTL-S (both posted on this page of the thread), I don't think it is cancelled. @jpiniero is likely correct: MTL-S will only be the low-end chips (65 W and under). It looks more and more like the high-end 125 W overclockable K chips will not be part of Meteor Lake.

Depends on your definition of desktop. I think Meteor Lake as HTPC or SFF will do quite well. The company that I work for only buys SFF desktop computers now (not that I have any say in it). My HTPC is less than 2 years old, so I'm not really in the market for one now.

And yet another mini-ITX on LGA1851 (that means not a laptop) Meteor Lake rumor is out: https://www.tomshardware.com/pc-com...-the-first-chips-to-use-intels-lga1851-socket

 

[uzzi38]

And yet another mini-ITX on LGA1851 (that means not a laptop) Meteor Lake rumor is out: https://www.tomshardware.com/pc-com...-the-first-chips-to-use-intels-lga1851-socket

Given the naming scheme (-PS) this is just gonna be a socketed version of MTL 6+8 I'd bet.