• Guest, The rules for the P & N subforum have been updated to prohibit "ad hominem" or personal attacks against other posters. See the full details in the post "Politics and News Rules & Guidelines."

Intel Cannonlake, Ice Lake, Tiger Lake & Sapphire rapid thread

Page 179 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

DrMrLordX

Lifer
Apr 27, 2000
14,181
3,297
136
Comparing across different OSes is never a good idea with Geekbench. The results get weird.
Just wanted to point out that that's less of an issue with GB5. GB4 used different compilers (I guess?) between OSes so it ran faster in Linux.
 

IntelUser2000

Elite Member
Oct 14, 2003
6,556
1,084
126
All Cometlake dies.

-U: 2/4/6
-H: 4/6/8
-S: 6/8/10

Just wanted to point out that that's less of an issue with GB5. GB4 used different compilers (I guess?) between OSes so it ran faster in Linux.
Different Linux versions also have different performance.
 

repoman27

Member
Dec 17, 2018
31
19
41
All Cometlake dies.

-U: 2/4/6
-H: 4/6/8
-S: 6/8/10
Totally being pedantic here, but the total number of new dies for Comet Lake will only be between 2 and 4.

For platforms you have the following:

CML-Y 2/4C
CML-U v1 (LP3) 2/4/6C
CML-S 2/4/6/8/10C
CML-H 8/10C (possibly lower core count as well)
CML-U v2 (LP4x) 2/4/6C

Intel decided to go with the Comet Lake Y name in their public announcements of AML-Y 4+2, which is actually built using the WHL-U 4+2 V0 die in conjunction with the SPT-LP PCH.

CML-U v1 launched as 6 SKUs, only one of which requires the new A0 6+2 die. The others can all be made using the WHL-U 4+2 V0 die, however they should all be paired with the 400 Series CMP-LP (a.k.a. ICP-LP) PCH. In fact, it appears that the Core i7-10510U, the highest clocked 4-core model, can only be built from the older V0 die. Obviously the V0 die does not have LPDDR4X support, and if A0 does then it has yet to be enabled for any available SKU.

The S and H platforms generally share dies. CML-S will clearly require a new 10+2 die, and Intel has indicated that the majority of the product stack will be built from 6+2 and 10+2 dies in LGA1200 packages. The cynical side of me thinks the original CFL-H/S 6+2 U0 die may make another appearance here, but being stepping 10 it is due for a refresh for security patches if nothing else. The PCH is a separate package for the H/S platforms, and H CPUs are sticking with BGA1440, so all of the lower core count CML-H SKUs can just be renamed Coffee Lake / Coffee Lake Refresh chips.

CML-U v2 should finally bring LPDDR4x support, and the A0 die may well be capable of that already, especially seeing as how it carries the entirely new 166 model number. However, Intel may do a second stepping of 6+2 or a proper CML-U 4+2 v2 die to replace the current WHL-U 4+2 die.

edit: clarify CML-H/S packaging situation.
 
Last edited:

mikk

Platinum Member
May 15, 2012
2,607
528
136
But why with tiger lake this huge? It's only 4 cores? Xe graphics seem to be pretty bad performance/size wise.

Tigerlake isn't huge, actually it's surprisingly small compared to Icelake with these added features. The Xe architecture seems to get a much improved area efficiency. Renoir has a smaller GPU and fewer features like no PCIe4, no LPDDR5 controller, no AVX512, smaller CPU cache for each core+ lower IPC and nevertheless it's slightly bigger.
 

IntelUser2000

Elite Member
Oct 14, 2003
6,556
1,084
126
For ST only the libc/libm should matter (and it does for some of the tests), that indirectly depends on the Linux distro/version. Just nitpicking :)
I looked at about 10 results. The 19 version is clearly faster than the other Linux versions in Geekbench.

But Linux distros are slower than Windows for multi-threading despite having better performance for single thread.

There's too many variables to reliably conclude which CPU is 5% faster.

But renoir has double the CPU cores.
The two are about the same die size.

AMD put more on CPU, Intel more on GPU.

I'm kinda surprised at the die size of Renoir. I expected it to be smaller if its 8 CUs. Maybe it shrunk 30% like with Vega VII.
 

Gideon

Senior member
Nov 27, 2007
628
773
136
But why with tiger lake this huge? It's only 4 cores? Xe graphics seem to be pretty bad performance/size wise.
Ice-Lake is 122mm^2. The 50% extra EUs are probably the bulk of the size-increase and the next one is the added cache. AVX-512 also takes up considerable space (compare server Skylake to desktop) though that's already included in Ice-Lake.

Renoir has a smaller GPU and fewer features like no PCIe4, no LPDDR5 controller, no AVX512, smaller CPU cache for each core+ lower IPC and nevertheless it's slightly bigger.
Come on now. Saying "Renoir is bigger" and listing all the differences you can find, expect the the "minor" one that it has twice the cores, isn't a fair comparison.

Besides, Ian Cutress estimated Renoir at 148mm^2 (~150mm2). The upper poster estimated Tiger Lake is 155mm^2. Overalll the die sizes are in flux enough to say that they are about the same. Especialy when taking into account what some have claimed previously, that Intel's 10nm is denser than TSMC 7nm.

As for PCIe4 - AMD's IF is built upon that and AMD uses it internally in their chips. I'm pretty sure Renoir has support for it internally and we'll have it in desktop versions. They probably cut it on laptops for power reasons. LPDDR5 controller shouldn't add much space compared to LPDDR4X.

All in all, I do agree that it's impressive that they managed to cram both the extra cache and EUs in ~150mm^2.
 
Last edited:

mikk

Platinum Member
May 15, 2012
2,607
528
136
Come on now. Saying "Renoir is bigger" and listing all the differences you can find, expect the the "minor" one that it has twice the cores, isn't a fair comparison.

Besides, Ian Cutress estimated Renoir at 148mm^2 (~150mm2). The upper poster estimated Tiger Lake is 155mm^2.

I told Renoir is slightly bigger which I'm sure is correct! Are you crazy or what? Furthermore the upper poster source removed his posting and corrected his calculation and now says 145 mm², so you are not well informed. However the die size based on the Tigerlake wafer appears to be smaller towards 140mm² or below. Come on now.


Based on this Tiger Lake wafer, there are 23 dies across and almost 28 dies vertically, meaning that 4C+96EU Tiger Lake has a ~140mm^2 die.

Assuming Intel is better at dicing up the wafer, and only has a 0.15mm scribe lane, and also matching it up more exactly to the partial wafers out to the edges, I can get the chips to 12.7mm x 10.4mm for 132 mm^2

If we take 1 unit = 10.82 mm Then : Tiger lake width = 10.82 mm Tiger lake height = 1.24 * 10.82 = 13.4168 mm Tiger lake die size = 145.17 mm
 

Nothingness

Platinum Member
Jul 3, 2013
2,027
309
126
I looked at about 10 results. The 19 version is clearly faster than the other Linux versions in Geekbench.
Ubuntu 19 uses glibc 2.29
Optimized generic exp, exp2, log, log2, pow, sinf, cosf, sincosf and tanf.
That's what I had in mind. Might be a reason... or not.

There's too many variables to reliably conclude which CPU is 5% faster.
Yes definitely. One should pick dozens of results to make a comparison within that range.
 

krumme

Diamond Member
Oct 9, 2009
5,865
1,444
136
Tigerlake isn't huge, actually it's surprisingly small compared to Icelake with these added features. The Xe architecture seems to get a much improved area efficiency. Renoir has a smaller GPU and fewer features like no PCIe4, no LPDDR5 controller, no AVX512, smaller CPU cache for each core+ lower IPC and nevertheless it's slightly bigger.
You might want to read you post again. What is it we dont understand?
"nevertheless it's slightly bigger."
Its nonsense. Or you simply forgot 4 cores and some io in your confirmation bias ...
 

IntelUser2000

Elite Member
Oct 14, 2003
6,556
1,084
126
I told Renoir is slightly bigger which I'm sure is correct! Are you crazy or what? Furthermore the upper poster source removed his posting and corrected his calculation and now says 145 mm², so you are not well informed. However the die size based on the Tigerlake wafer appears to be smaller towards 140mm² or below. Come on now.
I got 146mm2, and I used paint, and assumed the PCH is same at 53mm2.

This estimate also supports my theory that Tigerlake just became longer, and the long:short ratio went from 1.05 in ICL to 1.25, also getting us the 145mm2 figure.

Ian got his 150mm2 estimate of Renoir by basically eyeballing.

Exact measurements can often be more off(as the delete post used) as it does not account for the possible errors introduced by camera angles and the raised die section messing up things.

I told Renoir is slightly bigger which I'm sure is correct!
This is AT nerd form of a dick measuring contest. However they are not measuring the size of their own, but Intel/AMDs. With errors being plus or minus 2-3mm2 easily, we can also conclude they are the same.
 

Gideon

Senior member
Nov 27, 2007
628
773
136
I told Renoir is slightly bigger which I'm sure is correct!
Thank you for the updateded measurement links!
My gripe was mainly with the fact that you totally omitted the core count, yet listed every other difference (including some with minimal effect on die-size).

On top of that you ignored the ~53.76 mm² PCH chip that Intel has, and AMD doesn't. While it indeed also includes stuff AMD doesn't have at all (like WiFi6, CNVi) a lot of the I/O is also there, including a lot of the PCIe, USB, SATA stuff.

My point being that Renoir is still actually a really tightly packed chip (don't forget that the smaller GPU is old architecture and still 20% faster than Ice-Lake). All-in-all they are about equal, unless final measurements are indeed >mm² different.

From the forementioned a selection of facts it just seemed as if Renoir is total PoS that does everything badly while Tigeralke is the pinnacle of engineering achievement possible to man.
Ian got his 150mm2 estimate of Renoir by basically eyeballing.
Yes that's true, but that was published before pictures were taken. AFAIK measurments from the pictures tend to confirm it still tended to believe it.

This is AT nerd form of a dick measuring contest. However they are not measuring the size of their own, but Intel/AMDs. With errors being plus or minus 2-3mm2 easily, we can also conclude they are the same.
LOL, Agreed :D
 

jpiniero

Diamond Member
Oct 1, 2010
7,250
876
126
Thunderbolt and the IPU does take up a ton of space. AVX-512 and bfloat and all that does take up a little bit I imagine but it can't be that much... it's not like they added the second AVX-512 unit like they did with Skylake Server.
 

ASK THE COMMUNITY