Discussion Intel current and future Lakes & Rapids thread

Page 308 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

jj109

Senior member
Dec 17, 2013
391
59
91
I got a relative to run MLC on a TGL 1135G7 Inspiron laptop he got for school.
30ns HITM i5-1135G7 latency at 4.2GHz

My system for reference
22.5ns HITM 9900K at 5GHz/4.7GHz

The L2 <-> L2 latency hasn't increased much in relative cycles with the switch to non-inclusive.

Other notes:
14.5W PL1
10.5W package power in 1080p Dolphin emulator
2W CPU, 4W GPU, 4W (uncore + other) power.
18W total platform power (package, PCH, display, etc)

As to how the iGPU XeLP can out-perform the dedicated version, I think the difference is the lack of the shared 12MB LLC.
Given Anandtech's die-size calculated at 72mm^2, I can't see how it would have additional L3 cache over the iGPU version.
 
  • Like
Reactions: lightmanek

Spartak

Senior member
Jul 4, 2015
353
266
136
Now that Rocket Lake based on Sunny Cove is officially confirmed I'd like to remind @jpiniero and @DrMrLordX to the discussion we had after the sunny/willow/golden cove introduction in december 2018 whether it would be sunny or willow cove that would be backported. I was wayyy to optimistic about the time frame of the backport (2019) but for reasons unclear to me both of you ruled out the possibility of a sunny cove backport when there was literally nothing in that interview suggesting such a thing, and actually pointed to the opposite, that with the new cove designs the architecture was decoupled from process.

Not that it is that relevant anymore, I was just surprised nobody saw a sunny cove backport to 14nm when it was almost literally announced.
 
  • Like
Reactions: lightmanek

DrMrLordX

Lifer
Apr 27, 2000
21,620
10,829
136
Fair enough, but we've all acknowledge that there was a backport effort going on since then. Much has changed since 2018.
 
  • Like
Reactions: Spartak

coercitiv

Diamond Member
Jan 24, 2014
6,187
11,855
136
Now that Rocket Lake based on Sunny Cove is officially confirmed I'd like to remind @jpiniero and @DrMrLordX to the discussion we had after the sunny/willow/golden cove introduction in december 2018 whether it would be sunny or willow cove that would be backported. I was wayyy to optimistic about the time frame of the backport (2019) but for reasons unclear to me both of you ruled out the possibility of a sunny cove backport when there was literally nothing in that interview suggesting such a thing, and actually pointed to the opposite, that with the new cove designs the architecture was decoupled from process.

Not that it is that relevant anymore, I was just surprised nobody saw a sunny cove backport to 14nm when it was almost literally announced.
And yet the truth ended up right in the middle: they barely had time to backport the Cove into a product with meaningful shelf life (assuming Alder Lake doesn;t end up being a 2022 product). Rocket Lake U will not see light of day, that's how close everything was.

That's the problem with taking what they said at face value, without asking for them to be specific with their wording: it can mean whatever they want, whatever they're able to deliver. I won't pretend to know what @jpiniero and @DrMrLordX were thinking, but my take on the situation back then was the interview did not provide sufficient information to make 14nm backporting a certain event.
 

TheGiant

Senior member
Jun 12, 2017
748
353
106
And yet the truth ended up right in the middle: they barely had time to backport the Cove into a product with meaningful shelf life (assuming Alder Lake doesn;t end up being a 2022 product). Rocket Lake U will not see light of day, that's how close everything was.

That's the problem with taking what they said at face value, without asking for them to be specific with their wording: it can mean whatever they want, whatever they're able to deliver. I won't pretend to know what @jpiniero and @DrMrLordX were thinking, but my take on the situation back then was the interview did not provide sufficient information to make 14nm backporting a certain event.
Rcl u doesnt have any sense when tiger lake is out imo
 
  • Love
Reactions: Zucker2k

Zucker2k

Golden Member
Feb 15, 2006
1,810
1,159
136
There's got to be more performance left on the table between Skylake and Icelake. I suggest we eliminate the penalty of comparing a mobile chip to a desktop chip by comparing mobile to mobile. Something like the 6700HQ to a 1065G7, for example.
 

coercitiv

Diamond Member
Jan 24, 2014
6,187
11,855
136
Rcl u doesnt have any sense when tiger lake is out imo
@Spartak 's argument back in 2018 was based on the assumption we'd see RKL S in 2019, so around 1 year after his prediction. This made a lot of sense since it molded perfectly with a common accepted scenario of Intel being able to fix 10nm in 12-24 months (which they almost did). So RKL was a gap filler in 2020. It made perfect sense and yet Intel was very careful not to speak about 14nm backports, only about developing tools and not repeating mistakes of the past.

Fast forward to the present: RKL S is a gap filler in 2021. Half the lineup is canceled. In hindsight Intel talking about backporting in 2018 was done with little confidence in 14nm backport schedule or alternatively with little confidence of being able to fix 10nm in 12+ months.
 

SAAA

Senior member
May 14, 2014
541
126
116
I sense déjà vu:
Prescott → Skylake and friends, up to the super hot, super -clocked Comet Lake
Yonah → Tiger Lake, also Rocket lake
"Core" 2 cores, 65 nm → Alder Lake, 8 big cores (+8 little ones…) on 10 nm
Penryn 4 cores, 45 nm→ Meteor Lake, 16 big cores, 10 nm or maybe just 7nm EMIB tiles?

Given Alder has little cores supposed to be as fast as Skylake the comparison makes even more sense: big cores must be amped up.
Hopefully with 10nm ++ they keep the clocks up (maybe just slightly less than Rocket variant) so all of 40-50% IPC gains are felt even if it's just 8 cores.

Let's see how accurate it turns out by 2022-2023.
 

Spartak

Senior member
Jul 4, 2015
353
266
136
And yet the truth ended up right in the middle: they barely had time to backport the Cove into a product with meaningful shelf life (assuming Alder Lake doesn;t end up being a 2022 product). Rocket Lake U will not see light of day, that's how close everything was.

That's the problem with taking what they said at face value, without asking for them to be specific with their wording: it can mean whatever they want, whatever they're able to deliver. I won't pretend to know what @jpiniero and @DrMrLordX were thinking, but my take on the situation back then was the interview did not provide sufficient information to make 14nm backporting a certain event.

Absolutely and the fact backporting was such an incredible effort was the reason they were very careful in their wording not to raise any expectations. Still to me it was clear such a backport effort was indeed already started and sunny cove would be the most likely candidate for that.
 

DrMrLordX

Lifer
Apr 27, 2000
21,620
10,829
136
Kinda makes me wonder why it took Intel so long to get Rocket Lake ready for market? Or was it more that they didn't want to release it at all when they saw a competitive 8c Rocket Lake was going to be a 250W part? Doesn't seem like backporting worked out as well as planned.

@SAAA

You seem awfully optimistic about how well Alder Lake is going to turn out. Yonah was launched without much fanfare in 2c form, which was the same core count as future Conroe as well as Pentium D. Tiger Lake-U, in contrast, still hasn't gone beyond 4c, which is half the core count of Intel's prevailing Skylake-based mobile chips (8c). Intel has gone through two generations of 10nm CPUs with half the core count of Skylake, and 40% of desktop Comet Lake. It only looks worse if you include Cannonlake. We'll learn more when Intel finally launches Tiger Lake-H so we can see what the price/availability of 10SF/10SFE is looking like in the future.
 

SAAA

Senior member
May 14, 2014
541
126
116
Kinda makes me wonder why it took Intel so long to get Rocket Lake ready for market? Or was it more that they didn't want to release it at all when they saw a competitive 8c Rocket Lake was going to be a 250W part? Doesn't seem like backporting worked out as well as planned.

They had three options for desktop 2021, aside from doing nothing: Comet refresh (low chance), Rocket backport or 10 nm working so Alder.
10nm... we know how it goes so 14 nm new arch it is.
They did cancel Rocket U/H parts so they must be confident of Tiger H delivering on time, some quarters before Alder at least. That or they don't want big.little on mobile yet, for some reason.
 

DrMrLordX

Lifer
Apr 27, 2000
21,620
10,829
136
They had three options for desktop 2021, aside from doing nothing: Comet refresh (low chance), Rocket backport or 10 nm working so Alder.

Rocket Lake-S shouldn't have taken this long, though. Assuming Intel started work in late 2018, why would it take them over two years to get this thing to market?
 

Spartak

Senior member
Jul 4, 2015
353
266
136
They had three options for desktop 2021, aside from doing nothing: Comet refresh (low chance), Rocket backport or 10 nm working so Alder.
10nm... we know how it goes so 14 nm new arch it is.
They did cancel Rocket U/H parts so they must be confident of Tiger H delivering on time, some quarters before Alder at least. That or they don't want big.little on mobile yet, for some reason.

Why do you leave out Tiger Lake S? Comet refresh is out of the question but I don't understand why they didnt invest more in an 8c Tiger Lake S part. Tiger Lake has the frequencies for it, if they can do 4.8 on mobile they should be able to hit 5 on the desktop.
 
  • Like
Reactions: Tlh97 and coercitiv

Spartak

Senior member
Jul 4, 2015
353
266
136
Rocket Lake-S shouldn't have taken this long, though. Assuming Intel started work in late 2018, why would it take them over two years to get this thing to market?

That's assuming they started the backport effort around the announcement but it should have been at least a half year in the works by then. So with a launch in Q1-Q2 '21 it looks more like a 3 year timeframe to me.
 

Spartak

Senior member
Jul 4, 2015
353
266
136
Why do you leave out Tiger Lake S? Comet refresh is out of the question but I don't understand why they didnt invest more in an 8c Tiger Lake S part. Tiger Lake has the frequencies for it, if they can do 4.8 on mobile they should be able to hit 5 on the desktop.

To answer my own question, the reason we don't have Tiger Lake S for 2021 IMO was the panic mode they were in. Rocket Lake, in spite of the massive effort, was their most safe bet to have a somewhat competitive core available in '20/'21 with the alder lake core competitive enough to remain relevant if it got delayed to '22 and switched to 10nm++ (ESF) if the 10SF didnt pan out as hoped/expected*.

10SF turned out well performance wise (yields is another issue) but developing three desktop parts for one year based on three different cores and two different nodes was a bit much.

*I'm assuming from reports alder lake s is on 10SF but not sure where that info is based on. Alder Lake on 10ESF could still be a possibility?
 

DrMrLordX

Lifer
Apr 27, 2000
21,620
10,829
136
Why do you leave out Tiger Lake S? Comet refresh is out of the question but I don't understand why they didnt invest more in an 8c Tiger Lake S part. Tiger Lake has the frequencies for it, if they can do 4.8 on mobile they should be able to hit 5 on the desktop.

Another possibility beyond the one you mentioned yourself is that yields on 8c parts may be so bad that Intel just can't bring enough chips to market to serve everyone's needs.
 
  • Like
Reactions: Tlh97

yuri69

Senior member
Jul 16, 2013
387
617
136
For what it's worth I double-checked the results and found two additional Nehalem results, which I averaged with the one I had from Anand.

P4 to Conroe: 29.2%
Conroe to Nehalem: 8.6%
Nehalem to Sandy Bridge: 9.3%
Sandy Bridge to Ivy Bridge: 3.7%
Ivy Bridge to Haswell: 6.4%
Haswell to Broadwell: 1.8%
Broadwell to Skylake: 3.3%
Skylake to Ice Lake: 12.3%
Ice Lake to Tiger Lake: 0.8%
This is a weird interpretation. I've used your data but made all results based on Smithfield. So Smithfield is 100%.
Orig. order
Release
Clockspeed
Cinebench 11.5 ST
Weighted
Smithfield-based
Previous gen delta
Pentium 4​
Smithfield​
1​
05/25/05​
4.336​
0.56​
0.66​
100.00%​
0.00%​
Pentium 4​
Prescott​
2​
02/01/04​
3.2​
0.42​
0.67​
101.52%​
1.52%​
Pentium 4​
Dothan​
3​
03/12/03​
4.272​
0.6​
0.72​
109.09%​
7.58%​
Pentium 4​
Cedarmill​
4​
01/05/06​
3​
0.43​
0.73​
110.61%​
1.52%​
Pentium 4​
Presler​
5​
01/16/06​
4.007​
0.58​
0.74​
112.12%​
1.52%​
Conroe​
C2D​
6​
07/27/06​
2.53​
0.76​
1.53​
231.82%​
119.70%
Nehalem​
965x​
7​
11/11/08​
3.2​
1.14​
1.82​
275.76%​
43.94%
Sandy Bridge​
2700k​
8​
01/09/11​
3.8​
1.52​
2.04​
309.09%​
33.33%​
Ivy Bridge​
3770k​
9​
04/29/12​
3.9​
1.64​
2.14​
324.24%​
15.15%​
Haswell​
4770k​
10​
06/04/13​
3.9​
1.76​
2.3​
348.48%​
24.24%
Broadwell​
5200u​
11​
10/27/14​
2.7​
1.26​
2.38​
360.61%​
12.12%​
Skylake​
6700k​
12​
08/05/15​
4.2​
2.05​
2.49​
377.27%​
16.67%​
Kaby Lake​
7700k​
13​
08/30/16​
4.5​
2.19​
2.48​
375.76%​
-1.52%​
Coffee Lake​
8086k​
14​
10/05/17​
5​
2.4​
2.45​
371.21%​
-4.55%​
Coffee Lake R​
9900k​
15​
10/19/18​
5​
2.43​
2.48​
375.76%​
4.55%​
Comet Lake​
10900k​
16​
08/21/19​
5.1​
2.47​
2.47​
374.24%​
-1.52%​
Sunny Cove​
Ice Lake​
17​
08/01/19​
3.9​
2.16​
2.82​
427.27%​
53.03%
Willow Cove​
Tiger Lake​
18​
09/02/20​
4.8​
2.68​
2.85​
431.82%​
4.55%​

A chart showing the Ice Lake is more or less in-line with the gains since C2D/Nehalem era:
1604233480666.png
 

Hulk

Diamond Member
Oct 9, 1999
4,214
2,006
136
Wait, something is off: how are you measuring the jumps? Because I'm certain Conroe was 2x Prescott IPC. All the other look quite small too.
Still, it puts into perspective how large the change is with Icelake cores, biggest since Conroe, hopefully the next one is just as good.

You are absolutely right. When I first started this I wanted to have a look at relative performance increases by making the fastest CPU a "1.0" or 100% and all others a percentage of that. This of course as the effect of "compressing" all of the increases.

The way generation-to-generation increases are calculated is by only looking at the two generations being compared. Sorry about any confusion I caused here! I was trying to add clarity to the conversation but instead I muddled things up!

Here is the correct chart and results.
P4 to Conroe: 119.1% or 2.19x IPC increase
Conroe to Nehalem: 15.9%
Nehalem to Sandy Bridge: 14.9%
Sandy Bridge to Ivy Bridge: 5.1%
Ivy Bridge to Haswell: 8.5%
Haswell to Broadwell: 2.2%
Broadwell to Skylake: 3.9%
Skylake to Ice Lake: 14.2%
Ice Lake to Tiger Lake: 0.8%


ClockspeedCinebench 11.5 STWeightedAverage% Increase
Pentium 4Smithfield
4.336​
0.56​
0.66​
0.70
Pentium 4Prescott
3.2​
0.42​
0.67​
Pentium 4Cedarmill
3​
0.43​
0.73​
Pentium 4Presler
4.007​
0.58​
0.74​
ConroeC2D
2.53​
0.76​
1.53​
1.53119.1%
Nehalem965x
3.2​
1.14​
1.82​
1.7815.9%
Nehalem980x
4.101​
1.43​
1.78​
Nehalem965
3.737​
1.27​
1.73​
Sandy Bridge2600k
3.8​
1.52​
2.04​
2.0414.9%
Ivy Bridge3770k
3.9​
1.64​
2.14​
2.145.1%
Haswell4770k
3.9​
1.78​
2.33​
2.338.5%
Broadwell5200u
2.7​
1.26​
2.38​
2.382.2%
Skylake6700k
4.2​
2.05​
2.49​
2.473.9%
Kaby Lake7700k
4.5​
2.19​
2.48​
Coffee Lake 8086k
5​
2.4​
2.45​
Coffee Lake R9900k
5​
2.43​
2.48​
Comet Lake10900k
5.1​
2.47​
2.47​
Sunny CoveIce Lake
3.9​
2.16​
2.82​
2.8214.2%
Willow CoveTiger Lake
4.8​
2.68​
2.85​
2.850.8%
 

Hulk

Diamond Member
Oct 9, 1999
4,214
2,006
136
@yuri69

Ha! By "indexing" from the bottom you in effect expanded IPC increases from generation-to-generation. As you can see from my post before this one I recalculated by only looking at 2 generations at a time, which is generally how this is computed I believe.

Thanks for helping!
 

Hulk

Diamond Member
Oct 9, 1999
4,214
2,006
136
There's got to be more performance left on the table between Skylake and Icelake. I suggest we eliminate the penalty of comparing a mobile chip to a desktop chip by comparing mobile to mobile. Something like the 6700HQ to a 1065G7, for example.

I looked at a couple of "HQ" Skylake parts and the results fall right in range with the other Skylake desktop parts.
 
  • Like
Reactions: Zucker2k

Hulk

Diamond Member
Oct 9, 1999
4,214
2,006
136
Here is a chart I was making that prompted my Cinebench 11.5 ST anlysis. I was trying to pull together from data from various Anandtech (and other) sources. As always I defer to the experts here so please let me know where I have made errors and I'm sure there are plenty of them. I will make corrections and repost.
When reading from left to right new internal structures are in red font. I was trying to get a feel for the relationship between the new structures and Cinebench IPC results.
Updated.
 

Attachments

  • Untitled-1.jpg
    Untitled-1.jpg
    414 KB · Views: 13
Last edited:

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
As I was typing "Dothan" I was thinking to myself "I don't remember this as a P4?" Good catch.
That 965 score at 3.2GHz was recorded in the Haswell review written by Anand himself. He noted the frequency at 3.2GHz for that test. The 2700k was actually a 2600k score.

In broader applications, Nehalem was 10% ST gain at best. Xbitlabs did comparisons with HT and Turbo off and it came out to be 5-10% in applications which showcase architectural gains.

Sandy Bridge was 15-20% and this is based on my taking results from multiple reviews, and calculating per clock for each relevant benchmark(meaning ignoring synthetics and game benchmarks). Reviews using HDDs were closer to 10% than those using SSDs. X25-M, which started the client SSD revolution came out in 2009, so in 2011 it was still in a transition period even among enthusiasts.

Also Nehalem had the advantage of having a small gap between base/turbo plus a third memory controller. Lynnfield was the better comparison.

Given Anandtech's die-size calculated at 72mm^2, I can't see how it would have additional L3 cache over the iGPU version.

@jj109 Given that the Xe LP in Tigerlake is only slightly larger than 40mm2, 72mm2 die is a very large increase.

*I'm assuming from reports alder lake s is on 10SF but not sure where that info is based on. Alder Lake on 10ESF could still be a possibility?

I'd be surprised if Alderlake is not ESF. It features laptop CPUs which they prioritize like mad.

@stebler Goldwaterlake is Atom. They have many products outside of server/client so some of them could just be IoT variants based on Alderlake or even more niche applications most don't care about.

@mikk https://www.ultrabookreview.com/41841-asus-zenbook-14-ux425ea-review/

UX393 26W 1165G7. You can compare that with the 17W Acer in their other review.

Also while I agree about the CPU, the leader in GPU is clearly the Swift. I'm not sure why you think the cooling setup and the PL1 is the only indicator to performance. There are dozen or so settings related to this the manufacturer can optimize, plus there are finer details(with different firmware) that we don't know.
 
Last edited:
  • Like
Reactions: Tlh97 and mikk

Hulk

Diamond Member
Oct 9, 1999
4,214
2,006
136
@IntelUser2000,

Understood. I've been reading your posts around here for nearly 20 years and you know a LOT more about CPU architecture than I do. Absolutely no doubt about that! I enjoy the discussions and like to participate but generally I end up with my foot in my mouth;)

How accurate does this look?
 

Attachments

  • Untitled-1.jpg
    Untitled-1.jpg
    414 KB · Views: 34
Last edited:

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
@IntelUser2000,

Understood. I've been reading your posts around here for nearly 20 years and you know a LOT more about CPU architecture than I do. Absolutely non doubt about that! I enjoy the discussions and like to participate but generally I end up with my foot in my mouth;)

How accurate does this look?

Thank you.

That looks pretty good. I would switch the simple/complex around. Simple generally means the x86 instruction gets decoded into internal operations in a 1:1 ratio. Complex operations means it requires more micro ops(or internal operations) per x86, so it has to get sent to that decoder. Also even the "simple" ones are good enough to decode most/all of the SSE/AVX instructions and to top it off it has micro/macro ops fusion to further reduce the worst-case scenario of being 1-wide.

I don't know what you mean by L1 cache decoders and Decode. I'm assuming by the latter you mean the uop cache, and the former the decoders?

Another note(not necessarily asking you to put it in your chart). Skylake's L1 is 5-cycles except in a few rare cases. So Sunny Cove doesn't really lose in that regard.

Nehalem has some loop stream detector improvements and later generations even more. Gotta look it up again.