Discussion Intel current and future Lakes & Rapids thread

Page 178 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
Hmm, thought they used EMIB on one of the chip layers instead of stacking monolithic dice. Maybe I misinterpreted their presentation.

The die size is 60-70mm2 at most. Yes they stack the 10nm CPU/GPU die on top of the 22FFL process die which includes all the chipset functions.

Isn't latency also in play here? Though nobody's said anything about how EMIB connections affect latency versus "traditional" MCM configurations. At least not that I've seen.

If you look at package shots for Icelake, the two chips are extremely close, and it uses OPIO. You can also compare that to Kaby-G, which uses EMIB.
 

Hitman928

Diamond Member
Apr 15, 2012
5,182
7,633
136
So both Icelake and Tigerlake are build on 10+, really? I wonder what products Intel is planning for 10++ and 10+++. On the client side there is only Alder Lake coming after Tigerlake as far as we know.

The write-up I read said that Tiger Lake uses 10+ which is a revision on Ice Lake's 10 (no plus). I think intel is basically just pretending that Cannon Lake and their original 10 nm process never existed at this point.
 

jpiniero

Lifer
Oct 1, 2010
14,510
5,159
136
People here in the past weeks tried to tell me it's coming in Q1.

Figure they will ship product to OEMs in Q1, not actual devices being available.

Why is that weird? Aside from 5-6 people with very hard-driven goalposts here was able to draw the not-so-difficult conclusion that Charlie Demerjian is right and ICL is practically some marketing SKUs with just enough volume to make every major vendor to create a couple of designs and then that's it.
Until somebody can provide me with the number actual shipped ICL chips, I will believe the one with the best track record yet regarding intel's 10nm: Charlie.

My estimate is something like 65% Whiskey 25% Comet 5% Ice. 5% is still a pretty big number... but if you are willing to burn wafers it's easily achievable even with abysmal yield. It's not like Intel is fabbing much other than Icelake and Lakefield at this point.
 

majord

Senior member
Jul 26, 2015
433
523
136
She specifically said shipping to OEMs summer this year

Make of that what you will of that vague statement regarding when devices will be available
 

jpiniero

Lifer
Oct 1, 2010
14,510
5,159
136

I am talking Tiger Lake. My guess it that it will just be the 28 W U and 9 W Y models that replace the MIA Icelake models and branded 10th Gen.

Betcha it will be announced roughly the same time as the 10610U and 10810U.
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
I am talking Tiger Lake. My guess it that it will just be the 28 W U and 9 W Y models that replace the MIA Icelake models and branded 10th Gen.

Betcha it will be announced roughly the same time as the 10610U and 10810U.

The 28W 1068G7 at least exists in some format. The -Y parts are absolutely nonexistent. One Netbook talk about how Ys are only for MS and Apple is an indication they'll skip ICL-Y for the most part. Maybe we'll see a Surface Go with it?

I don't believe Apple will use ICL-Y.

I read somewhere that they they'll start shipping the 1068G7 soon. Considering how they are still introducing Whiskeylake systems, we might still see it, but it could be in one obscure model sold to a small country or something.

Why they bother with 1068G7 I have no clue. The HP Spectre x360's base mode is 18W, and performance mode is 28W.
 
Last edited:

lobz

Platinum Member
Feb 10, 2017
2,057
2,856
136
I am talking Tiger Lake. My guess it that it will just be the 28 W U and 9 W Y models that replace the MIA Icelake models and branded 10th Gen.

Betcha it will be announced roughly the same time as the 10610U and 10810U.
Let's make a deal then: you won't shave till intel ship TGL to OEMs. Not fair if you're a hobo or a member of ZZ Top!!! :D :D :D
 
  • Haha
Reactions: lightmanek

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
So I would say Icelake's perf/clock gain is 15% than 18%. With the countless number of applications out there, its extremely strange for CPU manufacturers to be specific with the numbers. With Ryzen it was "52%". I mean come on! Just say 50%.

They used to be much more sane and claimed 15-20% instead.

15% seems to be the Integer gain which is what we want to see as a uarch advance. Higher gains are in FP.
 
  • Like
Reactions: mikk

csbin

Senior member
Feb 4, 2013
838
351
136

4 Cores, 8 Threads Tiger Lake U DDR4 SODIMM RVP

Ubuntu 16.04.6 LTS 4.15.0-45-generic x86_64


Single-Core Score: 1309
Multi-Core Score: 3132



"frequencies": [ 3825, 3845, 3828, 3873, 3869, 3817, 3831, 3882, 3881, 3808, 3812, 3874, 3504, 3667, 3719, 3772, 3842, 3788, 3723, 3759, 3735, 3640, 3560, 3792, 3094, 3828, 3647, 3652, 3716, 3741, 3193, 3708, 3770, 3802, 2779, 3757, 3768 ]
 

Ajay

Lifer
Jan 8, 2001
15,332
7,792
136
The write-up I read said that Tiger Lake uses 10+ which is a revision on Ice Lake's 10 (no plus). I think intel is basically just pretending that Cannon Lake and their original 10 nm process never existed at this point.
IIRC, cannon Lake was 10, ICL was 10+. If Tiger Lake is still 10+, it just means Intel hasn’t made any changes to the xtor geometry or materials.

I’m sure Intel is still trying to reduce defect density. Amongst other issues, SAQP just sucks. It requires more masks, multiple lighting techniques, more litho steps and creates too much geometric variation that results in too many defective circuits to salvage the CPUs.
 

uzzi38

Platinum Member
Oct 16, 2019
2,565
5,575
146

4 Cores, 8 Threads Tiger Lake U DDR4 SODIMM RVP

Ubuntu 16.04.6 LTS 4.15.0-45-generic x86_64


Single-Core Score: 1309
Multi-Core Score: 3132



"frequencies": [ 3825, 3845, 3828, 3873, 3869, 3817, 3831, 3882, 3881, 3808, 3812, 3874, 3504, 3667, 3719, 3772, 3842, 3788, 3723, 3759, 3735, 3640, 3560, 3792, 3094, 3828, 3647, 3652, 3716, 3741, 3193, 3708, 3770, 3802, 2779, 3757, 3768 ]

Effectively same IPC as ICL-U. Use Linux scores like this: https://browser.geekbench.com/v5/cpu/941452.gb5 for comparison.

Also, GB5 on Linux reports ST Turbo as base frequency, the .gb5 shows the true base of 1.8GHz. This is probably 15W config.
 

DrMrLordX

Lifer
Apr 27, 2000
21,583
10,785
136
Effectively same IPC as ICL-U.

. . . hmm. Didn't Intel already hint at that? Something about Sunny Cove and Golden Cove being the uarch steps that brought IPC gains while Willow Cove was more focused on restructing cache?
 

uzzi38

Platinum Member
Oct 16, 2019
2,565
5,575
146
. . . hmm. Didn't Intel already hint at that? Something about Sunny Cove and Golden Cove being the uarch steps that brought IPC gains while Willow Cove was more focused on restructing cache?
Yes, but from what we've seen that cache restructure has been, for the most part, a very notable cache bump. So many people were expecting upto a 10% IPC uplift.
 

DrMrLordX

Lifer
Apr 27, 2000
21,583
10,785
136
Yes, but from what we've seen that cache restructure has been, for the most part, a very notable cache bump. So many people were expecting upto a 10% IPC uplift.

Might be that Geekbench just isn't responding well to the larger cache.
 

csbin

Senior member
Feb 4, 2013
838
351
136

TGL vs ICL

IJKrDf3Xn5ZOh9L.jpg



BDYHmgCx9qVovQd.jpg
 
  • Like
Reactions: lightmanek

uzzi38

Platinum Member
Oct 16, 2019
2,565
5,575
146
Might be that Geekbench just isn't responding well to the larger cache.
GeekBench traditionally does benefit from better memory latency afaik, though with 5 it doesn't have it's own test. By extension, the effect of the cache should be present too.
 

DrMrLordX

Lifer
Apr 27, 2000
21,583
10,785
136
GeekBench traditionally does benefit from better memory latency afaik, though with 5 it doesn't have it's own test. By extension, the effect of the cache should be present too.

Maybe we'll see more effects in AVX512 benchmarks?

@csbin

That second comparison is weirdly bad for the reference Tiger Lake system. I wonder if the Dell system is running faster RAM? Or is it a case of dual channel vs single channel?
 

csbin

Senior member
Feb 4, 2013
838
351
136
Maybe we'll see more effects in AVX512 benchmarks?

@csbin

That second comparison is weirdly bad for the reference Tiger Lake system. I wonder if the Dell system is running faster RAM? Or is it a case of dual channel vs single channel?


Dell running LPDDR4X 3733Mhz
reference Tiger Lake system 2666/3200?
 

mikk

Diamond Member
May 15, 2012
4,112
2,108
136
. . . hmm. Didn't Intel already hint at that? Something about Sunny Cove and Golden Cove being the uarch steps that brought IPC gains while Willow Cove was more focused on restructing cache?


It's a stepping 0 Tigerlake and only Geekbench, way too early for such analysis. In this test the MT score is very low, there are better TGL scores on Geekbench.
In fact all of the leaks are stepping 0 based beside the zhihu ES2 leak with the 5-8% IPC uplift claim. 1.8 Ghz base is really good for stepping 0 if it's a 15W model.
 

tamz_msc

Diamond Member
Jan 5, 2017
3,726
3,554
136
It's a stepping 0 Tigerlake and only Geekbench, way too early for such analysis. In this test the MT score is very low, there are better TGL scores on Geekbench.
In fact all of the leaks are stepping 0 based beside the zhihu ES2 leak with the 5-8% IPC uplift claim. 1.8 Ghz base is really good for stepping 0 if it's a 15W model.
I've never heard about PPC increasing with a CPU stepping, unless there were errata on the older stepping which is fixed by the newer stepping. Based on the leaks till now ICL-U PPC = TGL-U PPC.
 
  • Like
Reactions: lobz and uzzi38

mikk

Diamond Member
May 15, 2012
4,112
2,108
136
I've never heard about PPC increasing with a CPU stepping, unless there were errata on the older stepping which is fixed by the newer stepping. Based on the leaks till now ICL-U PPC = TGL-U PPC.


I'm talking about bugs, bios and software optimizations, misconfigurations, different TDP etc, everything is unknown. The MT score of this entry is a good example, there is no explanation why this sample scores 1000 points higher. We don't even have an accurate frequency log.


So here Tigerlake has a better IPC, there is only one big outlier for Tigerlake. This Tigerlake sample scores 81% higher in the HDR subtest! This alone proves my point.
 

tamz_msc

Diamond Member
Jan 5, 2017
3,726
3,554
136
I'm talking about bugs, bios and software optimizations, misconfigurations, different TDP etc, everything is unknown. The MT score of this entry is a good example, there is no explanation why this sample scores 1000 points higher. We don't even have an accurate frequency log.


So here Tigerlake has a better IPC, there is only one big outlier for Tigerlake. This Tigerlake sample scores 81% higher in the HDR subtest! This alone proves my point.
Any discrepancy between different runs in Geekbench is almost always due to whether or not it was a clean run. Background processes, Windows doing its own thing etc. all bring down the multi-thread score. The single-thread score however is more consistent across various runs. As for the HDR subtest being 81% faster, it could be down to missing instruction support, which could be a BIOS issue.

However, since all the TGL-U leaks we have right now are Geekbench runs, except for the one which leaked earlier with a 400 MHz clock, there is no indication till now that PPC has increased from ICL-U.
 

uzzi38

Platinum Member
Oct 16, 2019
2,565
5,575
146
I'm talking about bugs, bios and software optimizations, misconfigurations, different TDP etc, everything is unknown. The MT score of this entry is a good example, there is no explanation why this sample scores 1000 points higher. We don't even have an accurate frequency log.


So here Tigerlake has a better IPC, there is only one big outlier for Tigerlake. This Tigerlake sample scores 81% higher in the HDR subtest! This alone proves my point.
Comparing across different OSes is never a good idea with Geekbench. The results get weird.
Stick to same OS and also, apparently Geekbench 5.1 changed the scoring system ever so slightly, so don't use anything of the Geekbench 5.0.x versions either.

Also, use the ST scores. The clocks in the .gb5 are most definitely the ST clocks of this chip in particular.
 
  • Like
Reactions: spursindonesia