Discussion Intel current and future Lakes & Rapids thread

Page 240 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

jpiniero

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

So Alder Lake does support AVX-512 but it gets disabled when running in hybrid mode.
 

DisEnchantment

Golden Member
Mar 3, 2017
1,590
5,722
136
From GCC changes submitted by Intel, it seems the compiler will not generate AVX512 instructions when the -march=alderlake.
They are probably targeting the common denominator. Which then raises some other questions for those hybrid core configs.
How Intel tackles this segmentation should be interesting to follow.
‘tigerlake’
Intel Tigerlake CPU with 64-bit extensions, MOVBE, MMX, SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2, POPCNT, PKU, AVX, AVX2, AES, PCLMUL, FSGSBASE, RDRND, FMA, BMI, BMI2, F16C, RDSEED, ADCX, PREFETCHW, CLFLUSHOPT, XSAVEC, XSAVES, AVX512F, AVX512VL, AVX512BW, AVX512DQ, AVX512CD, AVX512VBMI, AVX512IFMA, SHA, CLWB, UMIP, RDPID, GFNI, AVX512VBMI2, AVX512VPOPCNTDQ, AVX512BITALG, AVX512VNNI, VPCLMULQDQ, VAES, PCONFIG, WBNOINVD, MOVDIRI, MOVDIR64B and AVX512VP2INTERSECT instruction set support.

‘sapphirerapids’
Intel sapphirerapids CPU with 64-bit extensions, MOVBE, MMX, SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2, POPCNT, PKU, AVX, AVX2, AES, PCLMUL, FSGSBASE, RDRND, FMA, BMI, BMI2, F16C, RDSEED, ADCX, PREFETCHW, CLFLUSHOPT, XSAVEC, XSAVES, AVX512F, CLWB, AVX512VL, AVX512BW, AVX512DQ, AVX512CD, AVX512VNNI, AVX512BF16, MOVDIRI, MOVDIR64B, AVX512VP2INTERSECT, ENQCMD, CLDEMOTE, PTWRITE, WAITPKG, SERIALIZE and TSXLDTRK instruction set support.

‘alderlake’
Intel Alderlake CPU with 64-bit extensions, MOVBE, MMX, SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2, POPCNT, AVX, AVX2, AES, PCLMUL, FSGSBASE, RDRND, FMA, CLDEMOTE, PTWRITE, WAITPKG and SERIALIZE instruction set support.

There is going to be more dark silicon on the hybrid configs? and what is the gain then if they can't use these instructions in the first place.
This seems to be like a generic problem to be solved going forward. HSA has similar problems.

Makes me wonder if there is going to be disintegration of the frontend and backend. Would probably be messy.

On SW side, they could possibly have illegal opcode trap and migrate the thread context if it is on the smaller core, but not going to be good. Not thought through, just an idea.

EDIT:
Just some points to the post from @IntelUser2000
 
Last edited:
  • Like
Reactions: lightmanek

FriedMoose

Member
Dec 14, 2019
48
28
51

So Alder Lake does support AVX-512 but it gets disabled when running in hybrid mode.
It will be interesting to see if the small cores are another architecture or just Golden Cove without AVX-512 and TSX. If they're the latter then this would effectively be a 16C/32T CPU in the vast majority of workloads.
 
  • Haha
Reactions: spursindonesia

TheGiant

Senior member
Jun 12, 2017
748
353
106
I like that approach
it has only benefits
for gaming/tough workloads that don't scale linearly you have more first few powerful cores than everything equal and for the flow workloads you lose nothing, one part is faster and one not
but the windows scheduler waves hello !
 
Last edited:

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
It will be interesting to see if the small cores are another architecture or just Golden Cove without AVX-512 and TSX. If they're the latter then this would effectively be a 16C/32T CPU in the vast majority of workloads.

LOL, they are trying TSX again?
 

mikk

Diamond Member
May 15, 2012
4,112
2,108
136
ADL -S(881/601) -P(282/682) -M(282)


2 and 6 big core mobile variants, why not 4 big cores. They must be confident they can beat 4x Willow Cove with a 2+8 config unless 6+8+2 is also a low power variant. I mean it could be because it comes with GT2 graphics, TGL-H only comes with GT1 graphics.
 

uzzi38

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


2 and 6 big core mobile variants, why not 4 big cores. They must be confident they can beat 4x Willow Cove with a 2+8 config unless 6+8+2 is also a low power variant. I mean it could be because it comes with GT2 graphics, TGL-H only comes with GT1 graphics.
Personally I'm hoping that both varients will be used for -U honestly. Both the 6+8+2 and the 2+8+2 configs.
 

mikk

Diamond Member
May 15, 2012
4,112
2,108
136
Finally a confirmation for the launch date, even though I was hoping for August. Also there is Hotchips and GDC with TGL/Xe sessions in August, so we should get more informations by next month.
 
  • Like
Reactions: uzzi38 and vstar

jpiniero

Lifer
Oct 1, 2010
14,510
5,159
136
Personally I'm hoping that both varients will be used for -U honestly. Both the 6+8+2 and the 2+8+2 configs.

That's what P is, H+U. If I had to guess, I would say that M is a replacement for Lakefield. Could also be used for some Atomish products with the big cores disabled.

Surprised that they are doing this in September.

I believe the launch was supposed to be by now.
 

Exist50

Platinum Member
Aug 18, 2016
2,445
3,043
136


2 and 6 big core mobile variants, why not 4 big cores. They must be confident they can beat 4x Willow Cove with a 2+8 config unless 6+8+2 is also a low power variant. I mean it could be because it comes with GT2 graphics, TGL-H only comes with GT1 graphics.

Where are you seeing the ADL configs? Am I missing something on that page?
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
2 and 6 big core mobile variants, why not 4 big cores. They must be confident they can beat 4x Willow Cove with a 2+8 config unless 6+8+2 is also a low power variant. I mean it could be because it comes with GT2 graphics, TGL-H only comes with GT1 graphics.

Gracemont is supposed to have performance like Skylake. They'll have to all work in tandem, but that seems plausible.

N5000 R15: 284
+30% for Tremont
+30% for Gracemont
+85% for double the cores

That gets us to ~900, which is probably at the same level as Tigerlake. You add Golden Cove cores working(maybe not always, similar to Lakefield) but it'll get you above it.

The part about 7nm is very disappointing. And with Jim Keller pushing for TSMC. Some said with Icelake-SP we'll see lower-core count variants first with the 38 core count part coming later.

TSMC clearly has better process than Intel looking at the products its using.

I believe the launch was supposed to be by now.

It seems after the lockdowns, and lack of public venues to announce products(such as CES/Computex) they are basically announcing it at the same time as you can buy devices.

Which, I guess is not such a bad thing? Just keeps us waiting for details. :)
 
Last edited:

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
Can Alderlake 8+8 be competitive?

Well let's see.

This is assuming all 16 cores can work in tandem, Gracemont cores included.

In the U form factor, if you take my simple calculations above, 8 Gracemont cores should get 1600 in Cinebench R20. Perhaps with even higher TDP it can clock 10-15% higher.

1800 for 8 Gracemont

10 Skylake cores get 6300 in R20. Moving to 8 Willow Cove cores, the improved uarch cancels out the 2 extra cores. Assuming 20% further gain from Golden Cove, we get -

7600 for 8 Golden Cove

If you simply add the two, you get ~9400, which is on par with the Ryzen 9 3950X. By that metric it can get quite close. It does need few addendums though.

-If Hyperthreading can be active for Golden Cove for a total of 24 threads. On Lakefield, Sunny Cove's HT is inactive.
-If having two sets of cores work together does not create any contention.*
-If applications/OS support is good enough that combined performance is representative of performance for vast majority of applications.

*Cinebench could be close to the ideal case because it can simply be made to split up workloads into 24 sections. Not all threaded applications are going to be like this.

One could think of the hybrid state as "Hyperthreading 2.0". Where potential gains in perf/watt is there but with further complications in coding and support.
 

jpiniero

Lifer
Oct 1, 2010
14,510
5,159
136
The part about 7nm is very disappointing. And with Jim Keller pushing for TSMC. Some said with Icelake-SP we'll see lower-core count variants first with the 38 core count part coming later.

So is Alder Lake a 10 nm CPU and 14 nm GPU or 14 nm CPU and 10 nm GPU? Heh.

Charlie btw has an recent article claiming that the XCC core count isn't 38. I'm sure it's clickbaity in that it's around 38.
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
Charlie btw has an recent article claiming that the XCC core count isn't 38. I'm sure it's clickbaity in that it's around 38.

Welcome to news media 2020. Where clickbait is the norm.

I think I saw somewhere that it might have 1-2 cores disabled. So 36 cores.

So is Alder Lake a 10 nm CPU and 14 nm GPU or 14 nm CPU and 10 nm GPU? Heh.

I can't see how the -M version will be Lakefield. First of all, the die size will be ballooned by having twice the amount of cores on the same process.

Second, it pretty much needs the "14nm" section be 22FFL, and be the active interposer aka Foveros.

Third, why would a demanding compute section(CPU or GPU) be on a much larger process?

We know the successor to Lakefield-R has another code name. One is Ryefield.
 

DrMrLordX

Lifer
Apr 27, 2000
21,582
10,785
136
I believe the launch was supposed to be by now.

Yup. TigerLake won't be as late as Comet Lake-S but it's still going to be late. What it doesn't tell us is the difference between launch and availability. The previous assumptions I heard had Intel launching in July with widespread availability by Christmas. Now . .. ? Hard to say.
 

jpiniero

Lifer
Oct 1, 2010
14,510
5,159
136
I can't see how the -M version will be Lakefield. First of all, the die size will be ballooned by having twice the amount of cores on the same process.

Like I said there's probably some overlap with different segments. The actual details, well we will have to see.

Yup. TigerLake won't be as late as Comet Lake-S but it's still going to be late. What it doesn't tell us is the difference between launch and availability. The previous assumptions I heard had Intel launching in July with widespread availability by Christmas. Now . .. ? Hard to say.

I agree that it probably will be immediate availability.