Discussion Intel current and future Lakes & Rapids thread

[Markfw]

Yes of course. But anytime I'm waiting on my computer and I know all 32 of those threads aren't loaded I know there is still software work to be done.

You have quite the passion for DC. It's admirable.

Remember when there you to be DC competitions among the various tech site forums? That was the only time I did some DC.

I think I was also involved with that DC effort to brute force crack some 56 bit encryption way back?

Not to get off-topic too much, I will reply and not further sidetrack... I use DC now to help cure cancer, since I lost to it.

But, yes, those encoding/rendering tasks SOME users do it a lot and use all the cores, and games are using more than before. 16c/32t is not too high for desktop IMO, but, yes, its not for everybody.

 

[DrMrLordX]

@DrMrLordX The E core team is one area within Intel that has been executing fantastically.

As opposed to their Core teams? Yes. If Intel keeps hitting 20-30% IPC gains generation over generation, eventually Core will take a back seat, and all we'll get is Atoms which will be fine (or not; it would be really sad for Core). It'll also allow Intel to do a clockspeed reset and give them room for future headroom.

Also, the Networking division is far, far smaller than the general server division. In addition to the fact that Ridge products are basically replacing entrenched ASIC competitors. So @Exist50 has a point about the real server E core being Sierra Forest.

Unlike Sierra Forest, the Tremont "Ridge" products actually exist and have made it to market. Q2 2024 for Sierra Forest is extremely ambitious for a new server product on Intel 3 when Intel hasn't even launched anything on Intel 4 yet, and may not until late this year. Keep your fingers crossed people, this could get ugly.

You do know Ampere Altra is Icelake level in core performance? Graviton 2 is even lower, at Skylake level. Crestmont-class cores in Sierra Forest is very much in the ballpark category.

Yes, they're aging products (Graviton 3 is the latest gen from Amazon, and it's not clear if or when Ampere will release more product). I never said they were good, I only mentioned them as an example of what kind of core finds itself in ARM server products, e.g. they're homogeneous in their core arrangements and they tend to be composed of "big" cores from their respective generations (Graviton 2 and Altra are basically A76-gen, when the X cores didn't exist). I would expect future ARM server products to have cores similar to the X-line rather than their true "small" cores. Nobody is going to use A510/A515 in a server product, despite the fact that the core spam that would be possible from such cores would be insane.

 

[Timmah!]

Not to get off-topic too much, I will reply and not further sidetrack... I use DC now to help cure cancer, since I lost to it.

But, yes, those encoding/rendering tasks SOME users do it a lot and use all the cores, and games are using more than before. 16c/32t is not too high for desktop IMO, but, yes, its not for everybody.

Agreed. Especially when you can have that in notebook now as well.
I will be kinda disappointed if Zen 5 wont come with more than 16 cores.

 

[itsmydamnation]

@NostaSeronx
I guess you would highly enjoy this:

Bulldozer, AMD’s Crash Modernization: Frontend and Execution Engine

AMD’s K7 Athlon architecture formed the basis of the company’s CPU offerings for around a decade.

chipsandcheese.com

And this:

Bulldozer, AMD’s Crash Modernization: Caching and Conclusion

In Part 1, we looked at Bulldozer’s core architecture. But the architecture itself isn’t the full story.

chipsandcheese.com

Quite interesting findings and interpretations.

i haven't read them , but was it CMT isnt bad , yes the core was to narrow , but the thing that really killed it was the entire cache system ?

 

[BorisTheBlade82]

i haven't read them , but was it CMT isnt bad , yes the core was to narrow , but the thing that really killed it was the entire cache system ?

Yep, that pretty much sums it up. Also the branch predictor wasn't state of the art and there were some pretty heavy penalties involved for taken branches.

 

[NostaSeronx]

i haven't read them , but was it CMT isnt bad , yes the core was to narrow , but the thing that really killed it was the entire cache system ?

What had hurt all the Family 15h cores was taking a execution mid from a high-frequency design(reduce power from this) and smashing it with a front-end, core interface, fpu of a low-power mobile design(increase frequency from this). Which had caused clashing methodologies(K10(HP-focus)<-BD(LP-focus)) and balooned it's area from its 2/3rds of Greyhound/Husky target(Moore) to its double of Greyhound/Husky size(Butler).

TSMC 40nm => 162CPP/120Mx [4.9 mm2 w/o L2]
GloFo 32nm => 130CPP/104Mx [median size for this in low-power many-core config is 6.8 mm2 w/o L2 rather than the ~20mm2 we got]

The Bulldozer core that launched was based off K10 is thus part of Weber's 2003 roadmap:

2005(expected) => K9 Alsup
2007(expected) => K10 Glew

Q4 2004~2005 Fred Weber announced K9/K10 being dropped and leaves AMD.
The only core AMD has that isn't based on prior cores is Bobcat. So, when the IBM crew came in 2005 they opt for Bobcat to act like GPUL(PowerPC970) and Bulldozer to act like GP(POWER4).
Bobcat, >1 GHz:

Integer Scheduler/Cluster	Address Scheduler/Cluster	FPU Scheduler/Cluster
2x I-Pipes	1x Ld-pipe + 1x St-pipe	1x A-pipe + 1x M-pipe

Bulldozer(original), ~2 GHz:

Integer Scheduler0/Cluster	Integer Scheduler1/Cluster	Address Scheduler/Cluster	FPU Scheduler/Cluster
2x I-Pipes	2x I-pipes	3x Ld+St-pipes	4x FP-pipes *unsure of unit layout*

Clustered Microarchitecture(ST-focus first) -> Cluster-based Multithreading means each critical-path (general purpose/control) cluster gets a thread and everything else is SMT2.

K10 as Bulldozer(launched), ~3.5 GHz:

Unified Scheduler0/Core	Unified Scheduler1/Core	FPU Scheduler
2x EX-pipes + 2x AGLU-pipes	2x EX-pipes + 2x AGLU-pipes	2x FMA (2x~64b Hi/Lo) + 2x MMX (2x64b Hi/Lo)

Kept most of MCMT, but at launch converted towards Chip-level Multithreading.


Basically, have to track two different "CMT" implementations. One for the Moore's Bulldozer(Work up from Bobcat) and one for the Butler's Bulldozer(Work onto K10).

Weber(cores):


Hester(cores):


Weber[K10] -> Hester[Bulldozer] -> Meyer[K10 w/ Bulldozer bits]

 

[IntelUser2000]

Unlike Sierra Forest, the Tremont "Ridge" products actually exist and have made it to market. Q2 2024 for Sierra Forest is extremely ambitious for a new server product on Intel 3 when Intel hasn't even launched anything on Intel 4 yet, and may not until late this year. Keep your fingers crossed people, this could get ugly.

And most say that's due to product execution not process. I am way more worried about external factors than anything. I think the threat of nuclear war and/or China taking TSMC is way way bigger than all this. And we're sitting here arguing about what's basically nothing - computer chips. Imagine what happens when such things occur, when we worry about a mere 10-15% drop. If the whole economy drops 30% or even 50%!

Yes, they're aging products (Graviton 3 is the latest gen from Amazon, and it's not clear if or when Ampere will release more product).

And that's the point. That what ARM vendors are doing(starting to take enough share to make Intel/AMD notice) are using cores that are far worse in ST than on Sappphire Rapids/Genoa.

Gracemont is A780+10%. So if Crestmont is 20% faster per clock, then we're at X1 level. X2 is 10% faster and X3 is another 10% faster. And these server ARM cores are also quite low clocked, especially when it comes to ST as it doesn't have Turbo. That's why you are arguing what's basically semantics. If something like Tremont cores are in Sierra Forest, sure.

And they never just pure port the cores to server. Crestmont in SRF will have cloud server specific advancements over Crestmont in Meteorlake, however big it may be.

 

[moinmoin]

I think the threat of nuclear war and/or China taking TSMC is way way bigger than all this. And we're sitting here arguing about what's basically nothing - computer chips. Imagine what happens when such things occur, when we worry about a mere 10-15% drop. If the whole economy drops 30% or even 50%!

There is a reason nobody talks about that, it simply is off topic. I could mention climate change, it will eventually make the whole planet inhabitable for humanity, which means the whole economy drops 100%. So do we conclude we don't discuss anything anymore as everything is meaningless in face of that? Didn't think so.

Now back to Intel.

 

[nicalandia]

This Does Not look Good... 130% Higher than 14nm Skylake Xeon W-3275 in apps like Cinebench, V-Ray,

 

[igor_kavinski]

They are relieved and happy that they managed to finally ship!

Let them be ecstatic

 

[BorisTheBlade82]

This Does Not look Good... 130% Higher than 14nm Skylake Xeon W-3275 in apps like Cinebench, V-Ray,

View attachment 76617

View attachment 76618

View attachment 76619

Well, this is 2.3 times faster than the predecessor. Ignoring competition it doesn't sound that bad.

 

[nicalandia]

They are relieved and happy that they managed to finally ship!

Let them be ecstatic

I know, its a step in the right direction.. but beign 130% faster than a 28 core Skylake Xeon is really sad. I believe that the Xeon W-3175X is actually faster than the W-3275

 

[igor_kavinski]

I will be kinda disappointed if Zen 5 wont come with more than 16 cores.

I don't think it will unless Intel goes 8P+32E or even 6P+32E with Arrow Lake.

 

[igor_kavinski]

I know, its a step in the right direction.. but beign 130% faster than a 28 core Skylake Xeon is really sad. I believe that the Xeon W-3175X is actually faster than the W-3275

Progress is slow in the server world, especially on Intel's side. AMD had to really up their game to prove to big corporations that they could turboboost their workloads cost effectively and that's what has helped them gain marketshare. Intel isn't going to do anything revolutionary in this space unless they lose a lot of marketshare overnight. They can afford to be relatively complacent.

 

[nicalandia]

Well, this is 2.3 times faster than the predecessor. Ignoring competition it doesn't sound that bad.

No, the predecesor is Not a 14+ nm Skylake based CPU with half its cores. There is a 10 nm Ice Lake Xeon W-3375 with 38 Cores

 

[igor_kavinski]

There is a 10 nm Ice Lake Xeon W-3375 with 38 Cores

They will pretend it doesn't exist since it didn't ship in considerable volume. Typical Intel.

Also, they probably have a good idea of the installed base of Cascade Lake CPUs at their customer sites so this is an attempt to get them enticed to upgrade.

 

[jpiniero]

There is a 10 nm Ice Lake Xeon W-3375 with 38 Cores

That was basically cancelled, probably because of volume/yields.

 

[nicalandia]

..

That was basically cancelled, probably because of volume/yields.

Not at all, also there are other Icelake SKUs, like the much faster 32 Core Icelake Xeon W-3365.

Puget Systems did a Review on them.


Compare that to this..



The OG 3175X is actually faster than the 3275 as the X was Unlocked.

Besides AMX benchmarks, this does not look good. Doing the Math, this puts the Xeon W9 3495X between the 39995WX and The 5995WX in CBR23

 

[jpiniero]

Not at all, also there are other Icelake SKUs, like the much faster 32 Core Icelake Xeon W-3365.

I mean Icelake-W in general. They launched it but you can't really buy it.

 

[nicalandia]

I mean Icelake-W in general. They launched it but you can't really buy it.

But there were not Cancelled. So technically the Xeon W-3375 is the actual predecessor, even if there were not unlocked or not readily available.

 

[igor_kavinski]

So technically the Xeon W-3375 is the actual predecessor

You could try tweeting that to Pat...

 

[Timmah!]

I don't think it will unless Intel goes 8P+32E or even 6P+32E with Arrow Lake.

Yes, you are probably right.

Anyway, its rather sad how there is no way to have perfect product. At least, what i would consider to be at this time perfect for me personally, other people might be absolutely satisfied with what they have, or what is to be had, or opposite situation, maybe what i would want, would be actually insufficient to them... anyway, for me it would be 24C Zen4 part, clocking as high as 7950x (5,7ST/5,1~5,2MT) and ideally with v-cache on all dies. For price no higher than say 1000, ideally for the same price as 7950x3D. That would be ideal choice for me. I could get 24 big cores CPU now, but either TR or new SPR Xeon are significantly slower/lower clocked per core and most importantly prohibitively expensive. Paying 5x more for 1,5x more cores, that are 30 percent slower...how about no?

So there we have it. Its similar situation as with displays. Currently i have HP Z38c ultrawide-curved monitor. Its great all-around, though i already have it like 4-5 years, so would not mind something newer. There is 40 inch successor to my model with higher resolution (5120x2160, mine is 3640x1600). That would be great. But its still IPS 60Hz. Obviously i would want OLED at least 144Hz and with GSync/Freesync. It does not have that. If there are displays with those, they are either not ultrawide, or curved, or have the exact size, or have crap stand....or once again, are prohibitively expensive.

 

[igor_kavinski]

Anyway, its rather sad how there is no way to have perfect product.

I totally get what you mean. It's like these companies are slacking coz whatever they put out sells like hot cakes so they see no reason to push themselves harder to deliver revolutionary or even evolutionary products. Kinda like street sellers all selling the same food, just with different spices. It's a stale marketplace and companies aren't willing to take big risks to deliver something compelling that only a subset of the PC enthusiasts would truly appreciate. We are the minority.

 

[Tigerick]

I don't think it will unless Intel goes 8P+32E or even 6P+32E with Arrow Lake.

Actually, ARL was rumored to have 8+32 cores initially. Unfortunately Intel later reduced it to 8+16, but we are expecting it to be made by TSMC N3 process, so we might be seeing power efficient CPU from Intel at last.

As for Zen 5, we will see AMD handles the external L3 cache issue, we could see double cores if everything works. Turin also depends on it.

 

[Racan]

For what it's worth