GPUs & CPUs & Enthusiast hardware: Questions, Discussion and fanboy slap-fights - Nvidia & AMD & Intel - Separe but Equal. Intel rides in the back of the bus.

[The Ugly One]

E-cores exist primarily for market segmentation and selling OEMs new CPU SKUs when they have no new silicon.

Yes, Intel, Apple, Qualcomm, Samsung, and now AMD have spent billions of dollars developing heterogeneous CPU cores for no other reason than to trick their resellers into having more inventory SKUs than they need. There's no technical reason for it. It's a billion dollar psyop, and everyone in the industry was hoping nobody would be clever enough to see through the ruse.

 

[Susanna]

Yes, Intel, Apple, Qualcomm, Samsung, and now AMD have spent billions of dollars developing heterogeneous CPU cores for no other reason than to trick their resellers into having more inventory SKUs than they need. There's no technical reason for it. It's a billion dollar psyop, and everyone in the industry was hoping nobody would be clever enough to see through the ruse.

If heterogenous architectures are so great, explain why servers aren’t using them. Datacentres are all about cutting down on energy use, its almost as important to them as it is to cellphones, where these architectures are long well established.

The answer, it seems to me, is that the entire software ecosystem needs to be developed with this in mind. That’s doable on a phone, apple or google control the OS and can schedule whatever wherever, but on a PC that’s not true. Microsoft has too much software around itself to keep track of where it should go, and Linux doesn’t even try out of philosophy. That leaves it up to Intel to find algorithmic ways to assign cores, which demonstrably doesn’t work very well even after years of development. And while MacBooks are brilliant with Apple Silicon, there’s no indication that’s because the architecture is heterogeneous. With the highly integrated memory, I’m sure a homogeneous architecture would be as efficient.

 

[Misericordia]

Yes, Intel, Apple, Qualcomm, Samsung, and now AMD have spent billions of dollars developing heterogeneous CPU cores for no other reason than to trick their resellers into having more inventory SKUs than they need. There's no technical reason for it. It's a billion dollar psyop, and everyone in the industry was hoping nobody would be clever enough to see through the ruse.

I never said they were "tricking" their partners. I said their partners want "new" parts for the new PCs they're selling this year. That's why we got a "Raptor Lake Refresh" that did nothing but shuffle around the allocation of E-cores.

 

[Susanna]

I never said they were "tricking" their partners. I said their partners want "new" parts for the new PCs they're selling this year. That's why we got a "Raptor Lake Refresh" that did nothing but shuffle around the allocation of E-cores.

Do OEMs even care? If you go to a store or online to buy a computer, chances are it’ll only be advertised as “Core i7” or “Ryzen 7” processor, and you have to search the model number to find some weird laptop review site that actually says which processor is inside, which you can then look up at Intel’s or AMD’s website.
As long as Intel can continue supplying them with silicon in each segment, I don’t think the OEMs care whether that’s actually good silicon or not. Just look at how AMD continues getting away with selling zen 2 as “Ryzen 7030U”.

 

[Misericordia]

Do OEMs even care?

I assume they care because I've only started hearing Intel's branding "#th" generation relatively recently. I don't think I ever heard that terminology used before Kaby Lake as "7th gen" at the earliest. There's also the fact that E-cores allow them to sell PCs and say they have a "10 core CPU", when it's really a 6P+4E.

 

[The Ugly One]

If heterogenous architectures are so great, explain why servers aren’t using them. Datacentres are all about cutting down on energy use, its almost as important to them as it is to cellphones, where these architectures are long well established.

There is datacenter heterogeneity, but it's at the node level, not the chip level. You just spin down the entire node when it's not busy.

They are also very different use cases. The typical desktop/laptop use case is a single user, with a single CPU, running lots and lots of different applications, often with many applications open and in the background. Lots of datacenter use cases are a single application running in hundreds or thousands of homogeneous instances. Heterogeneity at the chip level doesn't really buy me anything when every single core is being used by SAP HANA or Ansys Fluent.

Also, in the server, if I need more compute power, I buy more nodes. E-cores aren't just about saving energy, they're also about economizing on die space. An 8E+8P config can do a lot more overall work than a 10P config.

That leaves it up to Intel to find algorithmic ways to assign cores, which demonstrably doesn’t work very well even after years of development.

Disagree - everyone's been pretty successful at deprioritizing background tasks to E-cores. The one thing they're not very good at is homogeneous workloads that just count the number of hardware threads a CPU has and blindly spin out that many software threads, but the programming guides for all these chips, not just Intel's, tell you not to do that.

That leaves it up to Intel to find algorithmic ways to assign cores, which demonstrably doesn’t work very well even after years of development. And while MacBooks are brilliant with Apple Silicon, there’s no indication that’s because the architecture is heterogeneous.

This is an M3 die map. 4 E-cores consume a lot less power than 4 P-cores. There are just fewer transistors to keep powered, besides the fact they're clocked lower. It really does save quite a bit of energy to move background tasks to small cores, which is why everyone does it now.

That's why we got a "Raptor Lake Refresh" that did nothing but shuffle around the allocation of E-cores.

Raptor Lake Refresh exists solely because Intel 4 had yield issues that couldn't be fixed in time to have a desktop version of Meteor Lake. It isn't the culmination of a secret plot hatched in 2018 to trick you.

 

[WelperHelper99]

There is datacenter heterogeneity, but it's at the node level, not the chip level. You just spin down the entire node when it's not busy.

They are also very different use cases. The typical desktop/laptop use case is a single user, with a single CPU, running lots and lots of different applications, often with many applications open and in the background. Lots of datacenter use cases are a single application running in hundreds or thousands of homogeneous instances. Heterogeneity at the chip level doesn't really buy me anything when every single core is being used by SAP HANA or Ansys Fluent.

Also, in the server, if I need more compute power, I buy more nodes. E-cores aren't just about saving energy, they're also about economizing on die space. An 8E+8P config can do a lot more overall work than a 10P config.

I think the use cases is a important distinction. With my laptop, I'm opening shit up, browsers, games, word processors, etc. Those different cores suddenly make sense.

A server is doing the same thing, over and over. Having the same core type in that instance makes sense, at least in my mind.

 

[Misericordia]

Raptor Lake Refresh exists solely because Intel 4 had yield issues that couldn't be fixed in time to have a desktop version of Meteor Lake.

Which is a marketing problem. They needed something to be "14th gen" for whatever reason, and since it couldn't be Meteor Lake, it was Raptor Lake with the E-cores shuffled around.

I acknowledge that E-cores have some legitimate use cases. I personally don't like them, and bought a 12700K specifically because I intended to disable the E-cores. I accept the value of these heterogeneous architectures particularly in laptops. Something like a 14900K might be a perfect product for somebody's use case, but would I rather have a 10 Raptor Cove core ringbus CPU? Yeah, sue me.

 

[geckogoy]

I don't like E-cores either, total dead weight. Complicates scheduling and can cause certain applications to run worse if not designed with non-uniform cores in mind. Supposedly the upside is I can possibly save power some of the time. Okay, I guess? Has never been a problem for me, servers can have fun with them or whatever.

 

[The Ugly One]

Which is a marketing problem. They needed something to be "14th gen" for whatever reason, and since it couldn't be Meteor Lake, it was Raptor Lake with the E-cores shuffled around.

Right, every year is a new "generation" regardless of how little the design changed from last year, something Intel has been doing since 2016's "7th gen" was just Skylake+. That doesn't imply the massive R&D expense put into heterogeneous compute--something literally every consumer-grade CPU company is doing--was merely a marketing trick with zero technical justification. And they didn't change E-core counts from 13th to 14th gen merely to manipulate you. It's because yields of the 7nm process improved, so they can sell bigger, faster CPUs at the same price point.

There's a reason everyone is doing E-cores now, and the reason is pretty simple. This is how the power demand of a CPU grows when you increase its compute throughput by increasing its clock speed, vs increasing its core count. No numbers because it's an abstraction.



From the programming perspective, a single 5 GHz CPU with 10 MB of cache is going to beat five 1 GHz CPUs with 2 MB of cache each 100% of the time. However, it will also consume 5x the power and generate 5x the heat. That's why everyone's gone to heterogeneous cores for desktops, phones, tablets, and laptops.

A server is doing the same thing, over and over. Having the same core type in that instance makes sense, at least in my mind.

Right. I don't sit in the server room with a mouse and keyboard and dick around with stuff. I load it up and walk away.

@snov One huge difference between how Apple approaches E-cores and how Intel does is that "hw.perflevel0.logicalcpu" on a mac returns the number of P-cores. E-cores are excluded. By contrast, in Windows, most system calls are going to lump E-cores and P-cores together, and you have to go out of your way to exclude E-cores if you're polling the system and launching threads based on that.

 

[Misericordia]

That doesn't imply the massive R&D expense put into heterogeneous compute--something literally every consumer-grade CPU company is doing--was merely a marketing trick with zero technical justification.

As a matter of personal preference, I am unhappy with the sacrifices that had to be made on Alder Lake for the sake of the E-cores. Saying they exist "primarily" for market segmentation was my own prejudice coming through. I never meant to suggest there was no technical justification for it.

And they didn't change E-core counts from 13th to 14th gen merely to manipulate you. It's because yields of the 7nm process improved, so they can sell bigger, faster CPUs at the same price point.

I'm not complaining about Intel manipulating me. Improved yields of Raptor Lake could have just led to price cuts for the preexisting parts. Instead they used it to start a new "generation", which is purely a marketing concept. I'm not mad about that, it's just the way it is.

That's why everyone's gone to heterogeneous cores for desktops. . .

Desktops are a tiny part of the market and R&D decisions aren't made for the sake of making the best desktop CPU anywhere, which was really my point.

 

[The Ugly One]

Desktops are a tiny part of the market and R&D decisions aren't made for the sake of making the best desktop CPU anywhere, which was really my point.

This is a breakdown of Intel's revenue. The dark blue is desktops and laptops (which, as markets, are often aggregated as just "desktops").

 

[Misericordia]

This is a breakdown of Intel's revenue. The dark blue is desktops and laptops (which, as markets, are often aggregated as just "desktops").

View attachment 5897201

Alright man, I relent. I was talking out of my ass and spun my personal taste as business analysis. They aren't making the desktop CPUs I want them to make.

 

[George Lucas]

Niggers calm down. I’m literally just making fun of the fact that the new scheduler Incel and Microcock teamed up to make Windows 11 a killer app was a failure and they deserve to be laughed at for it.

E-core design may be the future but we just don’t know yet. There’s always the possibility of another breakthrough making your old research moot. Companies spend billions to develop products that fail all the time. It can go either way.

 

[Susanna]

I’d still rather just save power by parking unused chiplets, than by having some of the cores be less powerful. AMD demonstrated that this can be done, and that it works rather well, with the 7900X3D and the 7950X3D.

 

[Just Some Other Guy]

I’d still rather just save power by parking unused chiplets, than by having some of the cores be less powerful. AMD demonstrated that this can be done, and that it works rather well, with the 7900X3D and the 7950X3D.

Agreed. E cores can screw off. Bonus points for marketing getting away with calling the processors "X amount of cores" by combining the E and P core count.

 

[AmpleApricots]

The secret is that one day you're gonna have to buy keys to unlock cores. That's what they're prepping you for.

 

[Just Some Other Guy]

The secret is that one day you're gonna have to buy keys to unlock cores. That's what they're prepping you for.

Sure, why not. Cars and shit are doing it. "Didn't pay your $50 subscription? No features for you, even though we already spent the money on implementing them anyways".

And the cucks will eat it up.

 

[Susanna]

I’d still rather just save power by parking unused chiplets, than by having some of the cores be less powerful. AMD demonstrated that this can be done, and that it works rather well, with the 7900X3D and the 7950X3D.

Adding on to that, you can have heterogenous chiplets. Like, first chiplet only carries four cores, so when the computer is idle and the extra chiplets power down, it'll use little power. Launch something that requires more cores and it'll start the two parked 8-core chiplets. The penalty from parking/unparking is a burst of initial cache misses for those threads that are moved either to or from the affected core, but that's not really noticeable to the user as long as it only happens while the new application is loading.

The secret is that one day you're gonna have to buy keys to unlock cores. That's what they're prepping you for.

Intel were pretty open that this is what their Software Defined Silicon will be, but they've not announced any plans for that to be a thing on the consumer platforms yet, since those lack most of the accelerators you'd pay to unlock anyway. It's been a thing in Enterprise for a long while though, but those customers are fine with paying for it as long as it can be budgeted for. Intel just needs to make a competitive server platform first.

 

[Ibanez RG 350EX]

E-core design may be the future but we just don’t know yet. There’s always the possibility of another breakthrough making your old research moot.

I don't see how it could be anything but different implementations of that same concept. No amount of R&D is going to change the electrical/thermodynamic reality that power consumption has an exponential relationship with frequency.

I'm sure breakthroughs will be made, but then new designs will just implement those breakthroughs IN ADDITION TO heterogeneous cores.