Discussion Intel current and future Lakes & Rapids thread

Page 698 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

AMDK11

Senior member
Jul 15, 2019
204
136
116
I'll summarize as simply "yes".

I really wouldn't quote MLID on any of this. He's clueless.
So RoyalCore x86 will be much wider and deeper than eg LionCove and PantherCove with the new x86 approach? Are there any patents or leaks about what the first generation of x86 from the RoyalCore line may contain?
 

Exist50

Platinum Member
Aug 18, 2016
2,445
3,043
136
This is the second time I see you discredit MLID here. What make you or anyone here more believable than them? He has been posting pretty accurate info and has inside sources on both camp.
Edit: I had something else written, but honestly, just wait and see. You're perfectly free to doubt me. Time will clear all that up.
 
Last edited:

nicalandia

Diamond Member
Jan 10, 2019
3,330
5,281
136
What's his latest prediction on IPC for Redwood Cove? 12-21% over Raptor Cove? I think that range, even the low bound, is too high.
I have to agree with that, but you know that even the most accurate leakers are sometimes given mislead information on purpose.

Based on CPU Layout I don't see enough change to call for higher than 16% IPC boost.

1663006114526.png
 

Exist50

Platinum Member
Aug 18, 2016
2,445
3,043
136
I have to agree with that, but you know that even the most accurate leakers are sometimes given mislead information on purpose.

Based on CPU Layout I don't see enough change to call for higher than 16% IPC boost.

View attachment 67439
Ah, you snipped that before I deleted it. Really, at this point I say just sit back and enjoy the show. I'm no YouTuber, and I have nothing to prove, vs MLID or anyone else. It was my mistake even letting their idiocy annoy me.
 
Jul 27, 2020
15,738
9,790
106
It was my mistake even letting their idiocy annoy me.
It's OK. Truth warriors have trouble being calm when they see falsehood being spread. But ARE you a truth warrior (as in, is there some secret you want to keep to yourself and not sharing with us because it makes you feel empowered over the rest of us plebeians? :p)
 

Kocicak

Senior member
Jan 17, 2019
982
973
136
BTW MLID had recently an interview with somebody from Intel, and that person explained that the only way they can stay competitive is to have specific accelerators in their server CPUs.

He also said that the Sapphire rapids delay is caused by a lot of steppings of the silicone needed to fix all the bugs, which may be also related to those accelerators - this is my speculation.
 
Last edited:
  • Like
Reactions: igor_kavinski

moinmoin

Diamond Member
Jun 1, 2017
4,933
7,618
136
BTW MLID had recently an interview with somebody from Intel, and that person explained that the only way they can stay competitive is to have specific accelerators in their server CPUs.

He also said that the Sapphire rapids delay is caused by a lot of steppings of the silicone needed to fix all the bugs, which may be also related to those accelerators - this is my speculation.
If true that sounds like they should have accelerated the switch to tiles first so that those accelerators can settle as individual tiles first. Doing that as part of monolithic dies as big as SPR would be pretty dumb.
 
  • Like
Reactions: igor_kavinski

inf64

Diamond Member
Mar 11, 2011
3,684
3,954
136
An Article from Sisoftware

What Intel needs is SVE-like variable width AVX (AVX-V?) to solve hybrid

What intel needs is a normal 16 big cores with SMT chip that has no compromises... That's why AMD will have an upper hand with 7950X, it's a no compromise 16 big cores product that clocks super high with comparable (or better) IPC and AVX512 support. No ifs or buts, no silly thread scheduler conundrums , no half AVX2 throughput on some threads. And they are yet to release a Vcache 16 big cores part, it's going to get ugly for intel really fast. Add on top that AM5 has 3+ years of products to come while Raptor Lake is the end of the road platform and there we have a reason why AMD is pricing their chips higher - they just can as it's a superior platform.
 

jpiniero

Lifer
Oct 1, 2010
14,509
5,159
136
and there we have a reason why AMD is pricing their chips higher

They aren't though. The 7950X is $100 less than the original 5950X MSRP on a far more expensive node. Now we don't actually know what Intel will charge for Raptor Lake but the 13900K is probally going to be about or more than the 7950X.
 
  • Like
Reactions: Tlh97 and Zucker2k

inf64

Diamond Member
Mar 11, 2011
3,684
3,954
136
They aren't though. The 7950X is $100 less than the original 5950X MSRP on a far more expensive node. Now we don't actually know what Intel will charge for Raptor Lake but the 13900K is probally going to be about or more than the 7950X.
Well we don't know Raptor Lake pricing yet, it could be higher but it could be comparable (or even a bit lower). The mid range Raptor Lake parts will be hard to beat as they will offer more threads (so better MT performance) Vs 7700X and 7600. AMD parts will likely be better in gaming and ST/ lightly threaded tasks, so people will have to decide what matters most to them in mid-range segment. 7950X and 7900X will have no issues competing with 13900K and 13700K.
 
  • Like
Reactions: Tlh97 and Rigg

Doug S

Platinum Member
Feb 8, 2020
2,201
3,404
136
He also said that the Sapphire rapids delay is caused by a lot of steppings of the silicone needed to fix all the bugs, which may be also related to those accelerators - this is my speculation.


I don't buy that. If they had them working other than any 'accelerators' they would already be selling them minus those that feature, and could start selling the ones with the accelerators later. It isn't as if they don't already have a massive number of SKUs with various features like AVX-512, types of VM support, ECC memory, SMP, HT, etc. etc. randomly turned off in some model numbers.
 

Exist50

Platinum Member
Aug 18, 2016
2,445
3,043
136
So RoyalCore x86 will be much wider and deeper than eg LionCove and PantherCove with the new x86 approach? Are there any patents or leaks about what the first generation of x86 from the RoyalCore line may contain?
Those seem like reasonable assumptions, but I have no idea what exact features will go into Royal. I do expect an extremely large IPC increase, however. Certainly well beyond Golden Cove. I think a while back someone else in this thread named a potential architect? André Seznec, was it? Think that's the most we have to go on, and even that I cannot confirm 100%.
It's OK. Truth warriors have trouble being calm when they see falsehood being spread. But ARE you a truth warrior (as in, is there some secret you want to keep to yourself and not sharing with us because it makes you feel empowered over the rest of us plebeians? :p)
Regardless of what I may or may not know, it's much easier to identify obvious fabrications than it is to make accurate predictions of the future. That's why I try to make my language very clear about things I know vs things I expect vs things I hope for. And why I have zero interest in gussying up with a Twitter anime pfp or a YouTube channel, lol.
 

Exist50

Platinum Member
Aug 18, 2016
2,445
3,043
136
An Article from Sisoftware

What Intel needs is SVE-like variable width AVX (AVX-V?) to solve hybrid

I agree with the premise, but I think they chose possibly the worse possible argument for it. Hybrid is, as far as we know, exclusive to client products. Client does not benefit significantly from wide vector support. They even lampshade this fact by acknowledging that despite using a vector length agnostic ISA, most current ARM cores (including the big ones!) don't actually support large vectors. For the HPC/compute applications that really benefit from it, Intel will have the server chips with all the bells and whistles available.

Instead, I'd argue that the greatest advantage is minimizing fragmentation. The same executable should be compatible across Rapids, Forest, and Lakes. But this could even be achieved with a smaller standard for vector length.
What intel needs is a normal 16 big cores with SMT chip that has no compromises... That's why AMD will have an upper hand with 7950X, it's a no compromise 16 big cores product that clocks super high with comparable (or better) IPC and AVX512 support.
It's not hybrid that's holding Intel back there. The opposite, if anything. Across the multicore loads that would actually give you a reason to buy the 7950x, Alder/Raptor Lake are heavily reliant on Gracemont to keep them competitive. And clearly from the benchmarks we've seen, AVX512 isn't providing a large advantage for common workloads.

As a thought experiment, imagine a Lake product with 16 big cores worth of area, on a competitive node. How would, say, 12+16, 10+24, or even 8+32 GLC/GRT on TSMC N5P perform? Obviously that would be more than competitive with Raphael, AVX512 or not. But instead Intel's forced to stick with a node disadvantage and choose to invest less silicon. Though in an ideal world, big cores and little cores would be unique dies that could be easily mixed and matched to suit everyone's needs.

Oh, and AMD's probably going to adopt hybrid eventually as well. They've been rather openly considering it, and Zen 4c is just the start. I'm not sure if I believe the Zen 5 + Zen 4c rumors, but it's gonna happen eventually. I'm sure some of the griping will have stopped by then.
 

ondma

Platinum Member
Mar 18, 2018
2,718
1,278
136
Royal core aims for double single core performance of golden cove and will have four way hyperthreading.. so royal core cinebench prediction 4000 single core performance.. just guessing
I hope you are right about the big IPC increase. Do you have a source for the four way hyperthreading? The rumors in this forum is the first I have heard of it being used.
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
Best case scenario: 4 years, unless AI has really sped things up.

Intel said yesterday:
-6 months design
-15 months until power on
-12 months validation

They said due to having a similar design to Alderlake, Raptorlake cuts that time down by 6 months.

On a side note, "Speeding up using AI" predictions never pan out that way for two important reasons:
-AI is overhyped like everything else, but especially. It's basically little more than pattern recognition using massive, massive amounts of data(granted it's a very neat way of doing so). The ideas existed what, 40+ years ago?
-If there's any speed up, then that's yet another resource. You can use the speedup to do more work, which means no time savings. If you do the same amount of work, then you get faster. Or a mix of two where you do little more and save a little bit of time. But when the industry has basically the same set of tools as you, then that's no advantage.