Discussion Intel current and future Lakes & Rapids thread

Page 78 - 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
So it got 50% bigger L1D, 100% bigger L2 and 30% bigger L3. Plus they throw a lot of crypto units to the mix.

This is the way they spent the transistors the 10nm brought to them.

Caches are relatively small portion of the Intel chips.

It's too early. Also, independent benchmarks can vary wildly. You can see in Geekbench the otherwise identical setup can vary by 2-3x. I compared against the 7130U's highest score. It's a Core i3 with no Turbo so there's no better comparison. Icelake is 5-10% better.

Sapphire Rapids is server only right? Will there be a desktop version?

Supposedly this is when client and server will diverge.
 
Last edited:

LTC8K6

Lifer
Mar 10, 2004
28,520
1,575
126
If it does more IPC and can clock around 5GHz (what is IMO the biggest issue now) I don't care
I think he considers it cheating to get a faster chip that way... :)
We won't mention the cache the other guy has...

Anyway, Intel supposedly canceled their original small10nm node, so they supposedly don't have as many more transistors as you'd think...
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
Anyway, Intel supposedly canceled their original small10nm node, so they supposedly don't have as many more transistors as you'd think...

Doesn't matter. It'll likely scale just as well as it did with previous Core chips. The density only matters for other much smaller market chips like Atom-based cores.

That's the stupid thing. They bet the whole farm on such a small gain.

There's also no indication 14+ or 14++ made Kabylake/Coffeelake lost density, despite what Anandtech articles say.
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
Intel backported Single Dummy Gate to 14++ to largely make up for the reduction in gate pitch.

Yea.

Process technology is sometimes compared to cooking. You have the same tools and materials, but the results can end up being wildly different.
 

jpiniero

Lifer
Oct 1, 2010
14,509
5,159
136
You know, I looked at the 2017 "Icelake" Geekbench result... and that has a Identifier of "GenuineIntel Family 6 Model 126 Stepping 0" and the BIOS identifier looks similar. It's a older version of Geekbench so it doesn't have the clock speed, so you can't really compare the two even if you discount the memory scores.
 
  • Like
Reactions: Brunnis

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
Let's do some crude and probably horribly wrong numbers to compare these two

@csbin's 6200U @ 2.8GHz sample vs that Icelake 2.6GHz sample, both running Linux. It's GB4 4.2.2 vs 4.3.0 though...

You don't need to use CPUs with Turbo which messes up comparisons. Just compare against 7130U. It's clocked at 2.7GHz, and is a dual core with Hyperthreading without Turbo. The lack of Turbo is what makes it great for comparisons.

I'm seeing a 10% difference in ST Integer and FP. I would say that's actually a big disappointment, unless we get a huge surprise at least in mobile or something. We should see closer to 20% to make up for all the delays.
 

csbin

Senior member
Feb 4, 2013
838
351
136
You don't need to use CPUs with Turbo which messes up comparisons. Just compare against 7130U. It's clocked at 2.7GHz, and is a dual core with Hyperthreading without Turbo. The lack of Turbo is what makes it great for comparisons.

I'm seeing a 10% difference in ST Integer and FP. I would say that's actually a big disappointment, unless we get a huge surprise at least in mobile or something. We should see closer to 20% to make up for all the delays.

Yea.

SKL→CNL(5%)→ICL(5%)

 

mikk

Diamond Member
May 15, 2012
4,111
2,105
136
You are a bit late. Dayman posted that first. The last 10 posts or so are talking about that result.


Yes indeed, I completely missed the last page. I have to add that the best CPU for comparison is probably the i3-7130U which runs with 2.7 Ghz without Turbo. There are lots of Geekbench results available for this CPU. Singlecore score average is roughly in the range of 3200-3300, however there are outliners of up to 3700 points.

Afaik the IPC jump from Haswell to Skylake was mediocre in Geekbench, looks like the IPC jump in Geekench is way higher with Icelake.
 

TheGiant

Senior member
Jun 12, 2017
748
353
106
Well the results look very nice. Going from 3300 @2.7GHz to 4100@2.6GHz is excellent.

Almost too good to be true with the improvements made last 5 years.

Didn't Intel say that they somehow break compatibility because they can not improve their IPC with supporting the run of the oldest code?
Isnt the solution to run even the MMX age level of code with emulation now?
 

csbin

Senior member
Feb 4, 2013
838
351
136
I have found an m3-7Y30 that was also running Ubuntu albeit a slightly older release, updated graphs below:

pEfAjYc.png

rqudAEa.png

Average difference:

ST: 15%

MT: 15.5%

Link to compared results

M3 score by it self


AES up to 51.5-66.7%......:fearscream:

Ice Lake have VAES Extensions....


Intel® Architecture Instruction Set Extensions Programming Reference (PDF)

5bd04eb562d4e.jpg



 
Last edited:

JoeRambo

Golden Member
Jun 13, 2013
1,814
2,105
136
That enchanced REP MOV that works properly fast with short sequences is going to be big boon. Replacing all those crazy memcpy library implementations from hell, with 1 instruction is going to be huge win for everyone, hopefully AMD will jump with bandwagon.
 

Nothingness

Platinum Member
Jul 3, 2013
2,371
713
136
That enchanced REP MOV that works properly fast with short sequences is going to be big boon. Replacing all those crazy memcpy library implementations from hell, with 1 instruction is going to be huge win for everyone, hopefully AMD will jump with bandwagon.
Yeah in ten years when all Intel CPU have it, like AVX. Oh wait no, AVX still isn't available on all currently fabbed Intel CPU more than 10 years after being announced, and more than 7 years after being available in Sandy Bridge.
 
  • Like
Reactions: KompuKare

JoeRambo

Golden Member
Jun 13, 2013
1,814
2,105
136
Yeah in ten years when all Intel CPU have it, like AVX. Oh wait no, AVX still isn't available on all currently fabbed Intel CPU more than 10 years after being announced, and more than 7 years after being available in Sandy Bridge.

You don't get what it does. It is good old rep movX, but made faster. For example Linux Kernel already is using REP MOV to memcopy. All is great and dandy, but current implementations on Intel/AMD are microcoded and have insane setup times, meaning performance is hurt when memcopy is done for small sizes. Alternative is those "implementations" that do crazy stuff like size detection and run into kilobytes of instruction cache use. Icelake (supposedly) reduces setup time and everyone already using it benefits automatically. The rest get incentives to delete code and replace hundreds of lines and branches with 5 instructions.
 

Nothingness

Platinum Member
Jul 3, 2013
2,371
713
136
You don't get what it does. It is good old rep movX, but made faster. For example Linux Kernel already is using REP MOV to memcopy. All is great and dandy, but current implementations on Intel/AMD are microcoded and have insane setup times, meaning performance is hurt when memcopy is done for small sizes. Alternative is those "implementations" that do crazy stuff like size detection and run into kilobytes of instruction cache use. Icelake (supposedly) reduces setup time and everyone already using it benefits automatically. The rest get incentives to delete code and replace hundreds of lines and branches with 5 instructions.
Thanks but I perfectly know what it does and how to implement it (and also know Torvalds rants about the various brain dead implementations of it). My point is that before getting rid of existing implementations you will need a huge critical mass of CPU implementing that new instruction. And you don't have any guaranty that Intel will implement it in all new CPU after Ice Lake.

Regarding the use of REP MOV in Linux kernel, it depends on a flag, because REP MOV is bad on some existing CPU and these CPU will be there for years. So in the end we will have to live with implementations that check CPU features to decide which code to use; if anything happens it will indeed be more complex code in libraries. And perhaps in 10 years these checks won't be needed anymore... No that's here forever.
 
  • Like
Reactions: Arachnotronic

LTC8K6

Lifer
Mar 10, 2004
28,520
1,575
126
Yeah in ten years when all Intel CPU have it, like AVX. Oh wait no, AVX still isn't available on all currently fabbed Intel CPU more than 10 years after being announced, and more than 7 years after being available in Sandy Bridge.
You're seriously complaining because you have to buy an i3 to get AVX?
 

LTC8K6

Lifer
Mar 10, 2004
28,520
1,575
126
Well, you don't get navigation and a good stereo when you buy the low end model car...
 

LTC8K6

Lifer
Mar 10, 2004
28,520
1,575
126
Yeah we needed some silly car analogy. I guess you're not doing software dev or you'd not say something like that :p Or perhaps you are just an Intel fan.
Definitely always been partial to Intel, both in the thin times and the thick times.
When I build budget systems, I do go with AMD.

I like Intel's market segmentation.