Discussion Intel current and future Lakes & Rapids thread

[IntelUser2000]

What are the advantages of using compute tiles for desktop with lower core counts that would fit on a monolithic chip?

If Raptor desktop is 8+16 and Meteor is 8+16 what's the advantage of the compute tiles? It is simply production cost?

Mind you, if they go tiles, they'll most likely go all-in on tiles, because creating a monolithic die will be additional work and resources required. So they will not create a monolithic small die if they have a tile strategy.

Also interesting.

Raptorlake goes from 8 to 16 E cores for additional performance. So Meteorlake might stay at 8+16, but use new cores. Arrowlake uses same cores, but goes to 8+32. Lunarlake would be 8+32 but new cores? Can't guess about Nova.

 

[Hulk]

1) Higher yields with tiles. Instead of needing to throw away an entire monolithic chip (total loss of that silicon area), you might have several good tiles and one bad tile in the same area (mostly usable silicon area).

2) One design problem doesn't hold up the whole generation. Suppose there is a problem with a new iGPU but the new CPU is performing great. With a monolithic chip, the thing can't ship. With tiles, you can use a previous iGPU combined with your new CPU. This eliminates the need for long delays and the expenses of backporting (like the 26.5 months between Comet Lake and Alder Lake and the necessary costs of creating Rocket Lake).

3) Flexibility. This is related to #2, but you essentially can ship products incrementally when they are ready rather than waiting for all new concepts to be perfected. The CPU team doesn't need to wait for the GPU team to be ready (and vise versa). They can launch what they have and move on to the next project. Being able to have your people work more independently gives your designs, planning, etc far more flexibility. Heck, Intel can even outsource some of the tiles to other companies for even more flexibility.

The drawbacks are of course higher latency, higher power, and/or more costly packaging.

Excellent explanation. Those reasons make perfect sense. Thanks.

 

[Hulk]

Mind you, if they go tiles, they'll most likely go all-in on tiles, because creating a monolithic die will be additional work and resources required. So they will not create a monolithic small die if they have a tile strategy.

Also interesting.

Raptorlake goes from 8 to 16 E cores for additional performance. So Meteorlake might stay at 8+16, but use new cores. Arrowlake uses same cores, but goes to 8+32. Lunarlake would be 8+32 but new cores? Can't guess about Nova.

Having a 12700K currently I know for my usage 8+4 is not enough. The "8" is fine, but not the 4. I need more background app compute. I don't think 8+8 would get me where I want to be performance-wise either. 8+16, or even 8+12 would be fantastic. Don't get me wrong I love this 12700K, my 4770K was basically 0+4 in terms of ADL performance!

Right now I'm rendering some video in the background at 3-4fps. When I bring that app to the front the P's make it happen at 13-14fps.

 

[coercitiv]

Stockholm syndrome at it's finest. Disable the E-cores already, have your system assign threads like it should to all the P-cores and enjoy one of the fastest consumer CPUs.

It's amazing how much performance you're leaving under the table just to maintain this illusion that Intel's approach for the scheduler was correct.

 

[LightningZ71]

Having a 12700K currently I know for my usage 8+4 is not enough. The "8" is fine, but not the 4. I need more background app compute. I don't think 8+8 would get me where I want to be performance-wise either. 8+16, or even 8+12 would be fantastic. Don't get me wrong I love this 12700K, my 4770K was basically 0+4 in terms of ADL performance!

Right now I'm rendering some video in the background at 3-4fps. When I bring that app to the front the P's make it happen at 13-14fps.

You'd be lucky to get 7fps in rendering with a +8 processor. That's not really going to improve your workflow much. You'd do better manually assigning priority or affinity to the rendering/processing task and ketting the e cores handle your foreground tasks.

 

[Hulk]

Stockholm syndrome at it's finest. Disable the E-cores already, have your system assign threads like it should to all the P-cores and enjoy one of the fastest consumer CPUs.

It's amazing how much performance you're leaving under the table just to maintain this illusion that Intel's approach for the scheduler was correct.

I'll try it. Perhaps I have been brainwashed. For some reason I have this nagging thought that I'd be leaving some additional compute behind by disabling the E's.

 

[eek2121]

I'll try it. Perhaps I have been brainwashed. For some reason I have this nagging thought that I'd be leaving some additional compute behind by disabling the E's.

try leaving it on the E cores, then do some other intense workload on the P cores.

 

[jpiniero]

Been meaning to mention this but Icelake W is pretty much nonexistant, even several months after the "launch". There does seem to be retail places you can special order it and they will ship it indirectly from Intel but that doesn't sound exactly like high volume. From what I looked you can't really buy a OEM workstation with the regular Icelake Server either. Guess the demand is all Threadripper Pro now.

Of note on this, Apple was expected to release a minor update of the Mac Pro with Icelake W as the last x86 release. I wonder if Apple will just not bother if they can't get the chips.

 

[DrMrLordX]

I wonder if Apple will just not bother if they can't get the chips.

They could be holding out for Sapphire Rapids W instead.

 

[IntelUser2000]

Man, Alderlake mobile feels like more of a disappointment than I initially expected. The silence around battery life was deafening. Seriously not even one mention of it? While the competitor AMD gave some really, really good power numbers.

The E cores this generation is all about performance and performance/watt. Not battery life, don't get them confused. The problem is the -U Rembrandt is 25-30% faster than the predecessor. ADL will probably catch up to Cezanne, but not Rembrandt. So it'll have the MT performance, iGPU, and battery life leadership! Granted MT performance gap is closer than Cezanne vs. Tigerlake. But based on marketshare none really cared about Cezanne being pretty much an 8 core HT less version of Tigerlake in MT.

14K Cinebench R23 and 5K+ R20 in 28W is what Rembrandt is.

Raptorlake is what puts the focus back on battery life and fully closing the CPU performance gap. Not the iGPU part.

I'd really want a proper Lakefield successor though. Not because Lakefield was good, because it was piss poor. But because it had potential. Meteorlake-M is probably the proper Lakefield, and since ADL-M is nonexistent, that's one product that can come in H1 2023. MTL-M done properly is what can also allow them to enter the proper Tablet market, one they exited after Airmont in 2016.

 

[JoeRambo]

The E cores this generation is all about performance and performance/watt. Not battery life, don't get them confused.

Important part of battery life equation is waking cores and supporting uncore up ( L3 cache slices, ring etc) then doing the task required and going back to sleep. Race to the idle if You'd like.
So while hybrid is good at power efficiency @ peak performance, full 100% of cpu load rendering and compiling something, it falls flat in actual real world tasks for laptops like browsing web, office jobs, video, typing in some IDE. If you are not idle, you are keeping your ring/L3 active, your IMC busy and DRAM chips not in power saving mode. Need massive difference between joules spent on big/marketing cores to "win" here.

There is also software side, hybrid scheduling decisions cost extra power and rescheduling tasks between big/small cores is not free either. If task starts life on small cores and is moved to big core it is disaster from HW point of view. L1/L2 caches are obviuos, but modern CPUs have massive BPUs that train over millions of cycles and TLBs that also start empty on context switch to fresh core.
This does not matter when workload will run for seconds 100%, but in bursty workload it is wasting energy.

 

[mikk]

The problem is the -U Rembrandt is 25-30% faster than the predecessor.

Do you really compare 15W Cezanne against 28W Rembrandt on the CPU side?

 

[insertcarehere]

14K Cinebench R23 and 5K+ R20 in 28W is what Rembrandt is.

Is it? Because that's a whole lot more than '25-30%' faster that AMD themselves claim over the 5800u:

 

[uzzi38]

Yeah that R23 number in specific is definitely much too high.

 

[Hulk]

Is it? Because that's a whole lot more than '25-30%' faster that AMD themselves claim over the 5800u:

I think the big question is what frequencies ADL mobile will be able to sustain within various power/thermal envelopes?

If 2+8 can sustain 3GHz for P's and E's that would provide CB23 of ~8,850.
If that configuration can sustain 3.5GHz for P/E then that would attain ~10,300.
Even 2.5GHz P/E would score ~7,400, which substantially higher than TGL.

Another bit of information to keep in mind is that at equal frequencies a Gracemont cluster outperforms a Golden Cove core by nearly a 2:1 factor. Assuming approximate area, the E's can be clocked much lower than the P's and still be more efficient.

I think we are going to have to wait for actual tests. We just don't know how this silicon optimized for mobile is going to do when optimized for power at the extreme U mobile envelope. Intel has it work cut out for it determine optimum frequency and core distribution for various workloads. It's a much more nuanced problem than for the desktop since power and heat dissipation are much more limited.

I think we are going to be surprised by how well 2+8 performs in the ultra mobile space. I have a feeling that 2 P's, with 4 strong logical threads are going to be able to carry the bulk of the work for most applications with the E's handling the lighter thread work loads.

Furthermore Intel knew (knows) that mobile is ultimately the most important category so I believe they designed ADL around that in the first place.

We shall see soon enough.

 

[IntelUser2000]

The Cinebench numbers are bit high but I believe the 30% gen-on-gen is going to be true, because the base clock increases greatly, which suggests power handling has improved greatly.

Their iGPU will not be greatly behind, but still significant. Their presentation compared it to a Dell XPS 13 9310, which has a PL1 of 18W and underperforms in games a fair bit. That 3x gain in Doom Eternal makes no sense, since Iris Xe G7 is getting 40 and 60 fps right now. I believe the 9310 is getting somewhere around 25-30 fps.

The Witcher 3 gains are muted since the 9310 happens to perform exceptionally in that game.

Overall it'll probably end up being 30-50% faster in iGPU over a properly configured 28W, LPDDR4x-4266 Tigerlake.

@Hulk They'll probably do well on the 9W segment. And the 6+8 might be competitive.

 

[mikk]

The Cinebench numbers are bit high but I believe the 30% gen-on-gen is going to be true, because the base clock increases greatly, which suggests power handling has improved greatly.

Once again, do you really compare 15W Cezanne against 28W Rembrandt? Really? If you think 30% is tested versus the same power your are wrong once again. AMD even admitted there is no 28W Cezanne U device in the market (no idea if true or a silly excuse) and apparently there is a 15W slide with less gains.

 

[jpiniero]

AMD even admitted there is no 28W Cezanne U device in the market (no idea if true or a silly excuse) and apparently there is a 15W slide with less gains.

Just the opposite - 15W Cezanne U is very rare.

NBC samples they have of the 5800U:

25/30
25/30
22/27
25/30

 

[uzzi38]

Once again, do you really compare 15W Cezanne against 28W Rembrandt? Really? If you think 30% is tested versus the same power your are wrong once again. AMD even admitted there is no 28W Cezanne U device in the market (no idea if true or a silly excuse) and apparently there is a 15W slide with less gains.

It's funny that you say that, when AMD's reference numbers last year on launch are 25W figures, not 15W figures.

 

[Abwx]

Just the opposite - 15W Cezanne U is very rare.

NBC samples they have of the 5800U:

25/30
25/30
22/27
25/30

Actually there s designs that get to 15W after some time, CB R15 perfs can be extracted from the graphs below.

1924 pts@30W
1820 pts@25W
1420 pts@14W

Lenovo ThinkBook 13s G3 AMD laptop in review: Subnotebook with fast Ryzen CPU

Notebookcheck tests the Lenovo ThinkBook 13s G3 with AMD Ryzen 7 5800U and 16:10 display.

www.notebookcheck.net

 

[IntelUser2000]

Once again, do you really compare 15W Cezanne against 28W Rembrandt? Really? If you think 30% is tested versus the same power your are wrong once again. AMD even admitted there is no 28W Cezanne U device in the market (no idea if true or a silly excuse) and apparently there is a 15W slide with less gains.

Dude the scores are from CPUMonkey. I don't know where they are getting that from and I might have been bit hasty on the claims but Alderlake isn't the magic bullet for mobile.

Even if the CPU is closer than it looks they'll be behind on the battery life and iGPU department, and not by a small bit. AMD is claiming 30% battery life improvement gen-on-gen? That's an insane lead! Right now Tigerlake leads a bit on video playback and uses little more power in browsing. Video playback they were claiming 30-40% reduction or something which will get ahead of TGL.

The web browsing gains are really, really good. TGL is not that efficient. There's a notebookreview thread about optimizing battery life and you can get Cometlake-U to be lower power.

The 20-30% gain in battery life for RMB cuts the fundamental battery life differences between x86 and ARM laptops by about half. Great job on them! I thought I'd see that with Lakefield!

 

[Thala]

Actually there s designs that get to 15W after some time, CB R15 perfs can be extracted from the graphs below.

1924 pts@30W
1820 pts@25W
1420 pts@14W

Lenovo ThinkBook 13s G3 AMD laptop in review: Subnotebook with fast Ryzen CPU

Notebookcheck tests the Lenovo ThinkBook 13s G3 with AMD Ryzen 7 5800U and 16:10 display.

www.notebookcheck.net

Thanks for the link! Reading a bit more carefully it seems that at 14W the score drops to 1244 while for the 1402 score it is looking closer to 19-20W.

 

[gdansk]

I find it very unlikely Rembrandt will be ahead in MT. 28W 6+8 should win in most multi-thread benchmarks. In usability and perceived performance I have no idea. But it's basically designed for benchmarks and given the exact right PL2 that Intel knows it needs.

 

[LightningZ71]

I hope that Intel has optimized their process tweaks for mobile with these parts. In other reviews, we've been seeing Intel not fare quite as well when held to hard power limits, winning some and loosing some. However, I think that their cache advantage in mobile will be the deciding factor in many cases.

 

[Khato]

Even if the CPU is closer than it looks they'll be behind on the battery life and iGPU department, and not by a small bit. AMD is claiming 30% battery life improvement gen-on-gen? That's an insane lead! Right now Tigerlake leads a bit on video playback and uses little more power in browsing. Video playback they were claiming 30-40% reduction or something which will get ahead of TGL.

The web browsing gains are really, really good. TGL is not that efficient. There's a notebookreview thread about optimizing battery life and you can get Cometlake-U to be lower power.

The 20-30% gain in battery life for RMB cuts the fundamental battery life differences between x86 and ARM laptops by about half. Great job on them! I thought I'd see that with Lakefield!

Where exactly is a 30% battery life improvement being claimed? From AMD's own press release there are two specific claims:

Compared to last-generation AMD Ryzen processors, the AMD Ryzen 6000 Series processors can use up to 30% less power for video conferencing.4 This results in incredible battery life with Ryzen 6000 Series processors achieving up to 24 hours of movie playback on a single charge5, all in a cool, quiet laptop.

(From https://www.amd.com/en/press-releases/2022-01-04-amd-unveils-new-ryzen-mobile-processors-uniting-zen-3-core-amd-rdna-2 )

The first one is the 30% figure being referenced I presume? In which case note that the claim is that the processors can use up to 30% less power, not the system. Page 8 of this AMD presentation comparing the 4000 series mobile to 5000 series illustrates the difference, and in that case how battery life can increase more than the power reduction in the processor - https://www.amd.com/system/files/documents/battery-life-whitepaper.pdf

The second important claim there is up to 24 hours of movie playback on a single charge. They don't provide a direct comparison to the 5000 series mobile here, and for good reason. Their 5000 series mobile announcement from a year ago - https://www.amd.com/en/press-releases/2021-01-12-amd-announces-world-s-best-mobile-processors-ces-2021-keynote - claimed 21 hours. A 14% improvement is still good no question, but it's not in line with the rest of their narrative. And again, not all of that improvement may be due to reduced processor power consumption as evidenced by their 4000 series to 5000 series battery life comparison.

What I'm currently more interested in is how ADL will compare to TGL in battery life. Substantive claims on that aspect were suspiciously absent in Intel's announcements. There are obviously going to be improvements compared to TGL, but to what extent will have to wait for reviews I guess.


Oh, and no question that AMD is going to be enjoying integrated graphics superiority for quite some time. I'm sure Intel will eventually catch up, but in the interim it's a notable advantage for those markets which care about such. (Which is itself a good point - AMD's single die strategy puts them at a notable cost disadvantage for quite possibly a majority of the laptop market that doesn't care about graphics performance/8 CPU cores.)