Intel Plans Return to Unified Core Design, No More Performance and Efficiency Core Split

[erek]

“Intel is assembling a team of engineers for its "Unified Core" design group that will deliver the new microarchitecture to power the next generation of processors.

Separating the core design into P-Cores and E-Cores has yielded Intel the desired results that the company hoped for. This includes product separation and many goals across platforms. For example, in the consumer sector, E-Cores run a lot of side and background tasks in the operating system, while P-Cores power main applications like games. To extract maximum performance, Intel has a dedicated Thread Director that makes the entire process work and dictates just what application goes to which core, in tandem with the operating system. Intel also provides P-Core-only and E-Core-only Xeon server processors that serve either performance sectors like HPC and AI or the cloud sector that needs many cores with somewhat lower performance, but in a dense 100-core+ package.

With a unification of the P-Core and E-Cores, Intel would have to look into other techniques for separating its product offerings. That can simply be done by using smaller cache capacities, as L2 and L3 cache occupy a huge percentage of the CPU die, just like AMD does with Zen 5 and Zen 5c. We can hope to see other ways, as Intel's Unified Core design team could bring surprises and some interesting new technologies. The timeline for the launch of this "Unified Core" is still unknown, as the design is probably in the early stages. These processes take years, and it is likely that this new core design will be launching by the end of the decade. Until then, we are looking into the hybrid core design like we have been using in recent years.“

Source: https://www.techpowerup.com/346645/...no-more-performance-and-efficiency-core-split

 

[Marees]

which fab
which node

hardware drives software

 

[uOpt]

What a novel idea.

 

[LukeTbk]

e-core becoming so good and gen on gen improvement so much higher than on the p-core side, that a little beefed up version of the 2028 version of them with access to the large cache could make p-core irrelevant.

 

[erek]

which fab
which node

hardware drives software

might be tsmc

e-core becoming so good and gen on gen improvement so much higher than on the p-core side, that a little beefed up version of the 2028 version of them with access to the large cache could make p-core irrelevant.

what if the e-cores are the new p-cores but just rebranded going forward

 

[LukeTbk]

might be tsmc

what if the e-cores are the new p-cores but just rebranded going forward

would not be surprising if the single type of cores used come from the e-core team yes, that what the rumors-internal leak point too.

 

[erek]

would not be surprising if the single type of cores used come from the e-core team yes, that what the rumors-internal leak point too.

any opinions on this approach? just curious

 

[LukeTbk]

any opinions on this approach? just curious

i am not sure unified here would mean they will go with a single type of core, the latest third type (meteor lake-panther lake) of ther low power core island could continue, scheduling challenge here could be simpler (only on battery, only for clear idle/low power task and they are kind of their own small independant cpu that live there, the but they do not support avx-512 type of problems do not exist for what they are asked to do)

Now that the e in e-core finally became real, the panther lake e-core are not just die space efficient but energy efficient using that team work as the best for the new core and retiring the very old by now p-core lineage seem a good choice.

also in general unified core does not mean they will all be the same, AMD zen5 and zen5c are unified, exact same isa and everything, but one is smaller than the other, Intet could do similar (smaller similar core with less cache and denser that do not reach as high as clock), it is more that the design will be a more direct successor of the e-core teams work than the old one, we should not overread unified here, Apple has unifed core (each can do everything the other can do) their small and big are still quite different.

you can have just same core (an 9700x)
you can have unified while very different (apple)
you can have unified and very similar (zen5-zen5c)
you can have non-unified (e-core that cannot do everything a p-core can do ISA wise)

Not sure how much we can read from a unified core teams existing, but I would really not assume Intel would all of the sudden change their way and be a single all the same core (while AMD-Apple-Qualcomm-Mediatek-Samsung-Google-Nvidia on the N1X/spark and others continue their different cores models)

 

[MrGuvernment]

What a novel idea.

I hated the P/E cores as I had a laptop when they first came out and VMware workstation would only get stuck to the E-Cores. I think there were plenty of other apps that could not use the P/E cores properly. So after that for work, any laptops I bought for people, I stuck with AMD, I want all my cores, that can all do the same workload, none of this under the hood switching back and forth and Intel deciding what needs what.

 

[erek]

I hated the P/E cores as I had a laptop when they first came out and VMware workstation would only get stuck to the E-Cores. I think there were plenty of other apps that could not use the P/E cores properly. So after that for work, any laptops I bought for people, I stuck with AMD, I want all my cores, that can all do the same workload, none of this under the hood switching back and forth and Intel deciding what needs what.

isn’t HTT Hyper-Threading back
too?

 

[LukeTbk]

isn’t HTT Hyper-Threading back
too?

that new core could be made small enough they will able to have 16 and more of them all over the stack and not go back to hyper threading.

they will still be able to scale them at different size/power on the same chip like AMD do and will go with Intel rentable unit instead of previous HT
https://www.tomshardware.com/pc-com...ing-ultra-wide-execution-using-multiple-cores

It is even a bit of the reverse, instead of a single core appearing to be 2 cores to the OS, multiple cores can combine together to appear as one (unified core will make that way easier to manage, if the small core participate with the big cores being able to do all the things make that always more possible) and with 20+ cores on small laptop that make more sense to go into that direction for consumer device.

 

[jardows]

The hybrid core approach introduced some significant additional complexity in software/OS design to make it work right. I think it would take a while (if ever) for Microsoft to simplify the OS code and no longer support the hybrid approach, but hybrid support could quickly languish in the OSS kernels. The kernels are getting large enough, not needing the advanced schedulers would be a great benefit in my opinion.

 

[OFaceSIG]

Jeez man.... I'm getting tired of these companies making monumental architecture decisions that then involve Microsoft, Linux devs, and everyone else.

 

[kram182]

Jeez man.... I'm getting tired of these companies making monumental architecture decisions that then involve Microsoft, Linux devs, and everyone else.

Chicken/egg problem - can't run HW without SW can't run SW without HW. Things don't remain stagnant whatever course they take - affect one, affect the other.

 

[LukeTbk]

Jeez man.... I'm getting tired of these companies making monumental architecture decisions that then involve Microsoft, Linux devs, and everyone else.

this is not like they are not a lot the same people (linux dev, those are intel employees quite a bit):

https://commandlinux.com/statistics/linux-kernel-contributors-lines-of-code-statistics/#:~:text=Intel maintained its position as,for approximately 12.5% of patches.

Same for google and AMD for the others company that push core design choices and according to rumors microsoft went with a zenC (small)-regular (big) zen core mix for their next Xbox, I am sure working with intel scheduler back and fort all the time.

OS being made to make hardware run they had some big IBM so they coded an OS for it and os evolving with hardware change also make sense, i am sure there is back and forth, specially at apple but even with Intels with the major OS that run on them, but it feel like the OS jobs to make the hardware work well for programs and users the best they can.

Here at least that seem to be something that would just simplify all those people life without much downside.

 

[DanNeely]

Big question is this for both CPU design teams or just one? IOW will we have ~2 years of unified cores and then go back to a P/E design for ~2 years?

 

[Lakados]

The P/E core config was something Intel needed to do so they could meet EU and California low power requirements.

Intel's fab process, combined with advancements in power delivery to the chips, makes it so the process is likely no longer necessary.

Go Intel Go!

 

[Sycraft]

Jeez man.... I'm getting tired of these companies making monumental architecture decisions that then involve Microsoft, Linux devs, and everyone else.

I mean for Linux devs ARM already made that decision a long time ago. Basically every Android phone is running one of their big.LITTLE core setups which are the same kind of hybrid core designs.

 

[sunruh]

Jeez man.... I'm getting tired of these companies making monumental architecture decisions that then involve Microsoft, Linux devs, and everyone else.

on every new version of the chip set i worked on ... the speed increase alone would change how the code behaved and we would have to implement more checks to make sure things happened in the correct order.

out of order execution is NOT always that great. sometimes you remove the bus before you remove the wheels and then it all crashes (ie deallocate memory while it is still in use)

big hugs to my mutex friends

 

[erek]

The P/E core config was something Intel needed to do so they could meet EU and California low power requirements.

Intel's fab process, combined with advancements in power delivery to the chips, makes it so the process is likely no longer necessary.

Go Intel Go!

very unfortunate and wild

amd was able to maintain compliance without this approach?

 

[Lakados]

very unfortunate and wild

amd was able to maintain compliance without this approach?

AMD didn't release a product meeting the low power requirements until the introduction of the 7000 series, and you need to enable Eco-Mode and set the limit to 65w to achieve it.

It's why you see so few OEM business machines with AMD CPUs.

Servers and Workstations follow a different set of rules, but AMD didn't correctly meet them for the EPYC and Threadripper lineup until the 4'th Gen Genoa lineup, but they were granted some exceptions because they were able to successfully argue that Servers and Workstations rarely exist in a sleep state.

 

[LukeTbk]

The P/E core config was something Intel needed to do so they could meet EU and California low power requirements.

P/E is not specially interesting, Big-small core concept that apple-amd evreyone else is the big one, and they will probably still do it (and even keep a big/small/low power island/new compute type on other tiles approach), just using unified cores that have the same ISA for the big and small cores the cpu will have (exactly what apple and amd has been doing) could be what they will do.