Discussion Intel current and future Lakes & Rapids thread

[A///]

Threadripper is great for those type of workloads, but 32-48 *mont cores would consume less power and produce less heat.

The presumption here being those TR cores would be actively boosting on a consistant basis. The type of work you describe could be done on a much older computer without much trouble.

 

[dullard]

Are you insisting a 13900K is a HEDT processor? It isn't.

What is your personal definition of a high end desktop CPU? Most of the industry defines it as the region between a typical desktop CPU and a workstation CPU. But how do you define it?

 

[DrMrLordX]

At 1.8-2Ghz the P core (wich has SMT), has same throughput than the e core clocked at 3.2GHz.

Hence at 1.8-2Ghz the P core will use the same time as the e core to do the task while consuming about 3500 joules to do so, at the other end the e core will consume 4500 joules to do the same work in the same time but at 3.2GHz.

Right it's just that the e-cores are more area-efficient. Which is really why Intel put them on their consumer SKUs in the first place. They couldn't or simply didn't want to commit more silicon area to P-cores than what they already have.

 

[A///]

What is your personal definition of a high end desktop CPU? Most of the industry defines it as the region between a typical desktop CPU and a workstation CPU. But how do you define it?

High End Desktop was an Intel marketing term. It didn't exist before Intel artificially created the segmentation with the P4 Extreme. In the future if the typical core layout becomes more than 30 cores, are we to assume a midgrade processor then at say 50 cores is a HEDT when it neither supports true work features, doesn't offer anything less than a higher end model with more cores and shares the same feature set with the higher end processors as the lower end processors which may have 24-28 cores?

 

[dullard]

High End Desktop was an Intel marketing term. It didn't exist before Intel artificially created the segmentation with the P4 Extreme. In the future if the typical core layout becomes more than 30 cores, are we to assume a midgrade processor then at say 50 cores is a HEDT when it neither supports true work features, doesn't offer anything less than a higher end model with more cores and shares the same feature set with the higher end processors as the lower end processors which may have 24-28 cores?

Do you mind answering the question? What do you personally consider a HEDT? It seems like a battle you like to take on, but none of us has a real clue where you define it so it is hard to have a discussion about it with you.

 

[A///]

Do you mind answering the question? What do you personally consider a HEDT? It seems like a battle you like to take on, but none of us has a real clue where you define it so it is hard to have a discussion about it with you.

My apologies. It isn't intentional. I'm a rambler. It's what old people do. HEDT as Intel envisioned it was artificial segmentation, offering nothing more than higher speeds or a few more extra cores. It used very watered down features again for artificial segmentation. I would consider modern hardware like Xeon WS and TR to be the only types of platforms between regular computing and datacentre. HEDT as Intel hailed it has been dead in the water after having a spear shoved up its heiny and then roasted over a barrel fire and spat on. The days of that stupid segmentation is gone. Prior to AMD Intel would price their WS processors higher. This isn't the case now. Right now you can get a 24 core processor form Intel with a meandering mainboard chipset that doesn't offer a lot outside what you would expect and his hamstrung compared to what AMD offers on X670.

 

[dullard]

My apologies. It isn't intentional. I'm a rambler. It's what old people do. HEDT as Intel envisioned it was artificial segmentation, offering nothing more than higher speeds or a few more extra cores. It used very watered down features again for artificial segmentation. I would consider modern hardware like Xeon WS and TR to be the only types of platforms between regular computing and datacentre. HEDT as Intel hailed it has been dead in the water after having a spear shoved up its heiny and then roasted over a barrel fire and spat on. The days of that stupid segmentation is gone. Prior to AMD Intel would price their WS processors higher. This isn't the case now. Right now you can get a 24 core processor form Intel with a meandering mainboard chipset that doesn't offer a lot outside what you would expect and his hamstrung compared to what AMD offers on X670.

So, you are basically saying HEDT is essentially dead, so no processors are HEDT? Although I would disagree that there is nothing between desktop and datacenter. Since then you are arguing that workstations are dead as well.

 

[A///]

So, you are basically saying HEDT is essentially dead, so no processors are HEDT? Although I would disagree that there is nothing between desktop and datacenter. Since then you are arguing that workstations are dead as well.

I either misspoke or you misunderstood. HEDT as Intel artificially developed it is dead. The vast range of processor in an affordable price point now makes the concept dead. It's consumer, WS, datacentre.

Recall what Papermaster said of the 3950X. It was a gamble for them, they didn't expect to sell many at that price point. It was one of the most sought after processors for at least five months. Even the 7950X or 13900K if we must go there are incredible for the their relatively low price. Recall a decade and change ago when six cores cost just as much?

What AMD did to Intel's BS segmentation was great, but in a manner of speaker AMD also destroyed market segment pricing on a range of products. Intel had the wits to offer up low end processors that are still good for gaming for a good price undercutting AMD, because it was clear AMD didn't care much about the little guys anymore. Zen 4's low end ignoring the X3D variants was nothing to write home about this generation, where Intel candidly swooped in and charmed buyers. Big win for Intel, big dumb move for AMD, who only hurt themselves by pricing games.

I've discussed their choices for the platform and their launch timing at length before. Very much a damned if you do, and damned if you don't situation.

 

[Abwx]

Recall what Papermaster said of the 3950X. It was a gamble for them, they didn't expect to sell many at that price point. It was one of the most sought after processors for at least five months.

The 3950X was some exception in CPU evolution, comparatively to a 2700X it piled a healthy IPC improvement, 2x the core count and a full node transition from 12nm to 7nm, overall it was more impressive than Zen 1 in its time.

All this brought 2x the perfs at same TDP while largely outmatching the Zen+ based 16C 2950X SKU, only things missing to make it a complete HEDT were the PCI lanes and memory channel counts.

Guess that we wont ever see such an uplift in one gen, comparatively the 5950X and 7950X got back in usual gen to gen evolutions even if the 7950X was a quite good improvement.

 

[A///]

The 3950X was some exception in CPU evolution, comparatively to a 2700X it piled a healthy IPC improvement, 2x the core count and a full node transition from 12nm to 7nm, overall it was more impressive than Zen 1 in its time.

All this brought 2x the perfs at same TDP while largely outmatching the Zen+ based 16C 2950X SKU, only things missing to make it a complete HEDT were the PCI lanes and memory channel counts.

Guess that we wont ever see such an uplift in one gen, comparatively the 5950X and 7950X got back in usual gen to gen evolutions even if the 7950X was a quite good improvement.

Rumours of an AMD 16 cores had been swirling for a long time. This was before the internet got the traction it did and brought in more people into the hobby. Zen+ was a mild refresh with a slight IPC increase and fixed some bugs. Zen 3 was a good generations uplift, and it was known from the start that Zen 4 was not a major upgrade. AMD voiced this constantly. It was meant to establish the am5 environment. There was still a nice uplift, but not the one people were grasping at stars for. AMD's presentation with the geomean was accurate on release day tests.

Mike Clark, I want to say that's the person I constantly bring up, has a quote in an an an interview with Anand Tech where says "I wish I could go to bed today and wake up to Zen 5 being out." This was 2 years ago, I believe. It's been well known even before Zen 4 came out that Zen 5 was a big change for the company. However Clark went on further to address some poignant issues with the current consumer platform.

I would presume by your join date that you're not some one who latches onto silly rumors and have a sound mind, which I know you do. There is this phrase in English which you may hav ein your native language if you're not a native English speaker that is "It's a chicken and egg problem."

The whole more cores argument was always made when Intel were in their 10 year stride pumping out quad cores beginning with their i3, i5 and i7 marketing campaign. Most people would buy the i5 or the i7, though these became applicable to the mainstream end user with Sandy Bridge. The latter of which supported HT, the former did not. Unimportant as you'd know that but the argument began a few years later. Intel could have made it cheaper to send out more cores, but the reason wasn't there. What consumer software would be taking advantage of it? Very few if I had to guess without reading old benchmarks. When AMD did their core dance with Zen 1, as in the original Zen, Intel scoffed at the idea. Consumers did not. Sure they weren't as fast as their top end intel counterparts but there were 8 cores with the 1800X and the price was nice. The take rate was so good Intel quickly changed gears and began upping their core counts. We're not 5...6 years into the Ryzen series and we're still in that strange phase where most games do not take advantage of 8 cores let alone 16 or 24. Some do, but they're not terribly popular games.

What do come to now? This discussion. Increasing memory channels and capacity for mainstream is a logical step, it would increase consumer costs for a mainboard, but the reasoning behind it is very good, obviously. Performance would be better. But neither AMD nor Intel want to do that. I can't tell you why other than they want to maintain a clear distinction between it, work station and full fledged datacentre. Even AMD has followed Intel's prior moves and now segregates between regular workstation and pro workstation.

Clark did mention at some point about the memory channels for normal consumers, extra cores and needed bandwidth that simply isn't possible on current flagship hardware.

The chicken egg problem is the bane of computing. Intel and AMD can continue what they're doing now and simply offer 24 cores or 16 cores depending on the brand you gravitate toward and make changes or increase cores, memory channels, and pray consumers buy into the more costly hardware. It's a delicate issue and it can be healthily debated by friends at a pub for days on end.

I couldn't tell you anything about AM6 or Z870. Does it make a major change. Is it a minor change. What is it? Is there a future turning point in computing? Who bloody knows.

What bugs me a lot is that there's always a forward step available for both companies to go to push the limits for consumer hardware. But there is no financial gusto for them to go down such a path when nothing requires that kind of performance for now. Hence the chicken and egg problem. The risk of if you build it they will come is too great. They'll derisk changes and upgrades as much as possible to limit exposure to consumer griping or poor design choices that may not pan out even if it sounds good on paper.

 

[Mopetar]

I'm not sure what you're arguing for in the first line of this reply. Are you insisting a 13900K is a HEDT processor? It isn't.

No, I was saying they could put a lot of e-cores into a die that's not much larger than their current desktop die. Trying to make even a 16-core die with their regular cores wasn't feasible for them.

They could, but it's gonna be lousy performance. You're comparing actual full cores to e cores that lack SMT and that cannot run AVX512.

There are plenty of workloads where AVX512 isn't used and having more actual cores is better than a few more p-cores at high clock speeds. Not having to run the e-cores beyond their peek efficiency is going to be more important in any kind of chip where you have 40 cores.

Intel doesn't even need to beat AMD in all out performance. They just need to offer a compelling product and when Intel abandoned HEDT, AMD increased their prices. Intel could definitely make a 48c chip with their e-cores that uses no more die space than their mainline desktop parts. Sell it for $1,000 and clean up.

As you point out the only limitation is memory bandwidth. Intel would need to make a new socket for HEDT that offers more memory channels. The other benefit of anything with a ridiculous number of computational elements is that even if the process node isn't mature, you can still disable a fair bit of it while still having a lot performance left available.

 

[A///]

No, I was saying they could put a lot of e-cores into a die that's not much larger than their current desktop die. Trying to make even a 16-core die with their regular cores wasn't feasible for them.



There are plenty of workloads where AVX512 isn't used and having more actual cores is better than a few more p-cores at high clock speeds. Not having to run the e-cores beyond their peek efficiency is going to be more important in any kind of chip where you have 40 cores.

Intel doesn't even need to beat AMD in all out performance. They just need to offer a compelling product and when Intel abandoned HEDT, AMD increased their prices. Intel could definitely make a 48c chip with their e-cores that uses no more die space than their mainline desktop parts. Sell it for $1,000 and clean up.

As you point out the only limitation is memory bandwidth. Intel would need to make a new socket for HEDT that offers more memory channels. The other benefit of anything with a ridiculous number of computational elements is that even if the process node isn't mature, you can still disable a fair bit of it while still having a lot performance left available.

It's not the only limititations mopar. It isn't a laxative buffet for Intel to squeeze cores non stop. You've got bandwidth, routing, thermals, and likely lowering clocks to remain within a performance and heat balanced envelope. You can't have a free lunch without thinking about consequences.

If Intel is smart they'll design their small e cores like AMD and not self cripple it and bite the silver bullet and incorporate more channels. Just like AMD crapped in their porridge they can do the same by pulling their own AMD. It's a damn big gamble for Intel but it would pay off. This is why I don't believe their big/little as it's designed will remain in perpetuity.

Someone's gotta take their hands out of their pocket and punch the sky proclaiming themselves to be the ultimate leader.

 

[Henry swagger]

https://twitter.com/x/status/1702629569834938549

Intel golden c is more efficient with smt than amd zen 4 but amd get 7% more performance from it.. i think thats why arrow lake will have no smt

 

[aigomorla]

The 3950X was some exception in CPU evolution, comparatively to a 2700X it piled a healthy IPC improvement, 2x the core count and a full node transition from 12nm to 7nm, overall it was more impressive than Zen 1 in its time.

All this brought 2x the perfs at same TDP while largely outmatching the Zen+ based 16C 2950X SKU, only things missing to make it a complete HEDT were the PCI lanes and memory channel counts.

Guess that we wont ever see such an uplift in one gen, comparatively the 5950X and 7950X got back in usual gen to gen evolutions even if the 7950X was a quite good improvement.

And how does this apply to Raptor Lake thread?
Didn't I state if your going to pull AMD in this thread you better have a valid reason with comparison?
Where is your comparison?
No one in this thread cares about a history lesson in regards to AMD.
ITS NOT AN AMD THREAD.

All i see is a wall of text on AMD.
If I see stuff like this again, I'll show you all a wall of infractions.

Moderator Aigo.

 

[JoeRambo]

Intel golden c is more efficient with smt than amd zen 4 but amd get 7% more performance from it.. i think thats why arrow lake will have no smt

How valid those numbers are, depends on methodology. Esp on AMD side, if it was not done with static VCore ( that is enough for stable operation with HT ), it won't measure correctly.
Intel's numbers are much more near truth, as HT is basically extracting extra performance from chip resources that are already powered on anyway. So 5 to 10% extra power for -10 to 45% performance, most of it spent in execution ports and mem subsystem's extra work.

I think the real killer for HT is the combination of two facts:

1) On desktop and mobile there are enough cores now and HT comes with scheduling headaches and can hurt "smoothness", made even more complex by bunch of marketing cores.
2) In cloud, virtualization stuff and similar due to massive security issues vendors no longer want HT or disable it outright. And for Intel due to those same security issues the cost is rising real fast both is validation, but also in transistors needed to "separate" the inter sibling thread interactions that leak information in various ways. I think Intel might have done research and extrapolated into future and realised that probably even just sharing ALU's or some scheduling/load/store queues is dangerous enough from information leakage PoV.

So that leaves workstations and some sizable part of servers to care for HT.
For workstations the question is: in post CPU rendering market, how many workloads are still able to scale when thread count is in hundreds even without HT?

AMD EPYC 9754 Benchmarks For The 128-Core Bergamo Review - Phoronix

www.phoronix.com

Code compiles are a poster child of developer workstations and compile nodes, but even here we are hitting limits beyond just bunch of cores, easy to imagine some 64C Zen4 Threadripper having enough already even without HT.
For "non-cloud servers i know some of my company's workloads love HT, so we will be sad to see it go, HT surely helps extract value out of those "license optimal" boxes, not sure how relevant for others nor how long it will stay relevant. Cloud is coming for everyone after all.

 

[Fjodor2001]

I just wonder if there is any consensus on what the max core count will be on ARL-S? I've seen anything from 8P+16E to 8P+32E mentioned.

 

[Geddagod]

I just wonder if there is any consensus on what the max core count will be on ARL-S? I've seen anything from 8P+16E to 8P+32E mentioned.

There is. 8+16 for initial launch, 8+32 maybe launching like a year later.

 

[SiliconFly]

There is. 8+16 for initial launch, 8+32 maybe launching like a year later.

8+32 is a big maybe

 

[eek2121]

I don't see Intel going to 8+32. Windows has issues with high numbers of cores. Also, there is only so much power to go around.

 

[SiliconFly]

I don't see Intel going to 8+32. Windows has issues with high numbers of cores. Also, there is only so much power to go around.

It's on 20A. Power won't be the main issue. Also, Windows can handle a lot more cores easily.

 

[JoeRambo]

I want Intel to release 8+32 for sole reason => 8 partitions of E-Cores have 8 slices of L3, and that's a chip that has 33% more L3 than 8+16 config?.
I will disable them just like always.

 

[eek2121]

It's on 20A. Power won't be the main issue. Also, Windows can handle a lot more cores easily.

You are overestimating the amount of power savings 20A will give. It does not give enough savings for Intel to double up on area efficient “little” cores. Adding 67% more cores while efficiency is only 30-40% higher at best and quite possibly lower when other factors are considered ?

Don’t get me wrong, Such a product would be intriguing, but I doubt we will see one unless Intel finds a need for one to exist. After all, adding that many cores can actually hurt system performance…

 

[TESKATLIPOKA]

You are overestimating the amount of power savings 20A will give. It does not give enough savings for Intel to double up on area efficient “little” cores. Adding 67% more cores while efficiency is only 30-40% higher at best and quite possibly lower when other factors are considered ?

Don’t get me wrong, Such a product would be intriguing, but I doubt we will see one unless Intel finds a need for one to exist. After all, adding that many cores can actually hurt system performance…

If efficiency is 30-40% then It should be good enough, at worst just lower the clocks.
16 E-cores are in my opinion drawing less power than 8 P-cores.

 

[rtxtwt]

I just wonder if there is any consensus on what the max core count will be on ARL-S? I've seen anything from 8P+16E to 8P+32E mentioned.

No 32E being seen among the initial batch of ARL ES datasheet.

and E core seems didn't boost sells, IMO it's not worth it

 

[S'renne]

No 32E being seen among the initial batch of ARL ES datasheet.

and E core seems didn't boost sells, IMO it's not worth it

I don't think that there's any general consumer software(idk the actual term you guys should get what I'm trying to say) that would run optimally with 40 threads if it has HT off for RU..