True nature of E-cores and how effective are they?

[phanbuey]

In my view, the strategy of small cores is directly linked to intel's slowness in advancing in chip manufacturing. With no extra space from a denser lithograph, they needed effective cores per area to match the massive MT performance of the Ryzen line up

This strategy is limited by TDP and heating etc...

This is exactly right.

Conversely it shows that they CAN match the massive MT performance using this design on a less dense node (at the cost of TDP).

The chip with e-core is far superior to one without given the same die space and lithography being held constant -- basically intel is squeezing the most performance they can through design/innovation because their node sucks. Clearly it works they took ST performance crown and matched MT performance on an inferior node. So to answer the question of "do e cores work/ make sense" based on the raw numbers answer is clearly yes -- ADL s would not be in the same league without them.

 

[R-T-B]

Some applications spawn one thread per core even if it's for HT, like 7zip for example probably winRAR as well. If threading & handling multiple cores was so easy we would've had 1000c/2000t (E?) cores by now.

None of that matters to the scheduler which works on a per thread basis.

 

[dragontamer5788]

Its not the OS that needs to be rewritten for E-cores or P-cores, its software in general. That's what people don't understand. If Blizzard / Activision decided to optimize for E-cores, then the next release of Overwatch or whatever will be faster on E-cores.

Microsoft can create libraries / algorithms to try and figure things out at the OS level. But its not like the OS is an oracle that knows how programs work, it just collects dumb statistics and makes broad general guesses. Its not Microsoft's responsibility to make more efficient use of E-cores.

----------

And there-in lies the problem. Zen4 cores are more similar to P-cores, so it will make more sense for developers to optimize for Zen4/P-cores, rather than spending efforts making things better for E-cores.

 

[Assimilator]

From an engineering perspective, while Intel's cores are big and hot they make up for this with raw performance, and since the vast majority of consumer workloads today are (still) single-threaded, 12 of Intel's cores can probably beat 16 of AMD's.

The problem there is simple marketing, because 12 < 16, and consumers are all about meaningless numbers, and therefore marketing departments are all about meaningless numbers, and therefore it became impossible for Intel to launch products with inferior core counts, even though their cores are arguably better than AMD's.

But Intel has one thing that AMD doesn't: heterogenous x86 CPU lines, with powerful Core and efficient Atom. So why not take those 12 powerful (P) cores, add 4 Atom (E) cores, and boom you have your 16-core CPU and you can now compete again in meaningless numbers? And that's exactly what they did.

So the E-core concept was somewhat of a desperation move, driven by necessity rather than any desire for efficiency, but it is the kind of innovative thinking I honestly didn't believe Intel was capable of anymore. And going forward, it's ultimately a good thing for Intel:
* they can still claim core-count leads over AMD
* they don't have to throw their big/hot core design out the window and start again immediately (it is quite obvious that the Skylake-derivative architecture is at the end of its life and needs replacing, but the addition of E-cores have allowed Intel to eke another generation or two out of it)
* instead of Atom being the unloved red-haired stepchild, since it is now in all Intel's CPUs as E-cores it should get significantly more resources to become a better architecture

and the industry/consumers, because scheduling improvements to be aware of heterogenous core counts should make Arm's second attempt at breaking into the Windows desktop significantly easier than the last try, and competition is a good thing.

Of course, the scheduling is the biggest stumbling block because (a) it seems it still isn't very good, even after significant development (b) Microsoft has chosen to gate it behind Windows 11.

 

[dragontamer5788]

But Intel has one thing that AMD doesn't: heterogenous x86 CPU lines, with powerful Core and efficient Atom. So why not take those 12 powerful (P) cores, add 4 Atom (E) cores, and boom you have your 16-core CPU and you can now compete again in meaningless numbers? And that's exactly what they did.

I disagree.

The iPhone / Android market has been swept up by big.LITTLE designs for the last decade. Its a proven design that physically makes sense: efficient cores really use grossly less power for only a little bit loss of performance. (Ex: 12% power usage for 50% performance, which is exactly what cell-phone users want). With the Apple M2 chip being released with TWO DAYS worth of battery life, Intel (who makes most of their money from laptops) is running scared, for good reason. If Intel's laptops are to compete against Apple laptops, they must use the big.LITTLE design, and P vs E-cores is an adequate version of the concept.

The issue is that Microsoft is likely not going to rewrite all their software (ex: Word, Office, etc. etc.) to efficiently use E-cores. And video gamers almost certainly won't use those E-cores either.

So we're in a situation where people won't get the huge battery lives that the Apple-crew is getting, but AMD is striking them from the other end with high-performance Zen4 cores.

Of course, the scheduling is the biggest stumbling block because (a) it seems it still isn't very good, even after significant development (b) Microsoft has chosen to gate it behind Windows 11.

I mean, if Microsoft had to spend a few million bucks on developers developing a new scheduler in the OS to handle P-cores / E-cores, it makes sense for Microsoft to try to make money off of it to recoup the costs. Schedulers are hard, and expert computer programmers who know how to write schedulers are expensive.

This is a totally different issue going on. Even in the Android / iPhone world, the proper scheduling of big vs LITTLE cores is very much a dark art and no one does it well yet. The decades of OS theory was built upon the assumption of homogenous (all the same) cores. When you have "some cores are faster than other cores", there's very weird and non-intuitive math that occurs. OS design barely understands power-efficient scheduling (invented for this era of cell-phone apps), let alone this new era of "sometimes power-efficient, sometimes high-performance, based on the User's expectations".

 

[P4-630]

And Qualcomm says they want to enter the desktop market in 2024. It will be interesting to see what design they (along with Nuvia which they purchased) may bring.

Or:
Surface Pro "9" ARM variant being powered by a custom Snapdragon 8cx Gen3 SoC, dubbed the Microsoft SQ3.

First Windows on ARM desktop PC in the form of a developer kit dubbed “Project Volterra,” which I’m told features the Snapdragon 8cx Gen3 SoC and includes the same neural processing unit (NPU) AI features and power that is expected to ship in the Surface Pro 9 with ARM.

Microsoft to merge Surface Pro X ARM and Surface Pro 9 Intel versions under one product line

Windows on ARM may be getting a promotion to the main Surface Pro line.

www.windowscentral.com

 

[R0H1T]

None of that matters to the scheduler which works on a per thread basis.

Nope, applications can override the built-in performance counters to set for instance priority themselves. Sandra's SiSoft IIRC fixes it's service priority to high even when you manually lower it. It mainly depends on the application. While the software can't randomly choose which core (P or E) it runs on it can definitely be programmed to not allow core parking/unparking or the system entering lower power states et al. I'm thinking E cores won't activate unless the OS/application wants the system to enter lower power state, or whatever the thread director says it to do.

Its not the OS that needs to be rewritten for E-cores or P-cores, its software in general. That's what people don't understand. If Blizzard / Activision decided to optimize for E-cores, then the next release of Overwatch or whatever will be faster on E-cores.

You can mitigate that to a certain extent using the registry.

 

[dragontamer5788]

You can mitigate that to a certain extent using the registry.

Yes, Windows is surprisingly flexible.

But I assume most users won't know how to use NUMA settings or P vs E-cores, or thread affinity. They'll rely upon the developers to "set sane settings". Registry and/or application startup settings through PowerShell (and other ways of achieving this) are always possible, but "users are dumb" (or at least, users won't spend the time learning these features).

 

[Vayra86]

I swear some of you people love to make something harmless into something controversial. E-cores are fantastic at managing background processes, leaving more room for the P-cores to be open for heavier tasks.

This isn't about power usage. It's not about the Skylake equal of these smaller cores being, well, not as good. It's just support cores in a sense.

I agree. The problem isn't E-cores, its the high peak power limits to get benchmark results. And that goes for both camps, but Intel is the bigger offender.

That said, they do their job regardless, both in getting top bench results and in munching on MT loads. Though I'm pretty far away from calling it a 'design win'. It shows itself purely as a desperate attempt to keep building monolithic CPUs when its clear as day a chiplet concept is the future. It works, for now, and until the moment people really can present heavy loads on these CPUs while gaming. That's when the ugly rears it head as you bump into power limits and perhaps even throttling to base clock. What Intel does do for x86, is pave the road for chiplet based CPUs with different core configurations. 'A chiplet of E cores with your P's sir?'

The real question is, will a gamer run into those limits. Not extremely likely, even a 4 year old CPU is still major overkill. But if you run heavier parallelized loads, E cores directly incur a penalty on P core performance or vice versa, while Ryzen does not have this issue, while Ryzen also carries a much lower TDP ceiling for peak performance.

Burst vs consistency, in a nutshell, when put under heavy stress.

 

[cvaldes]

I mean, if Microsoft had to spend a few million bucks on developers developing a new scheduler in the OS to handle P-cores / E-cores, it makes sense for Microsoft to try to make money off of it to recoup the costs. Schedulers are hard, and expert computer programmers who know how to write schedulers are expensive.

Microsoft has the money to hire expert computer programmers. So does Intel. And AMD. And Nvidia.

If they fumble the ball, it won't be because of money. And being first out of the gate doesn't guarantee a win. Go ask former Windows Mobile programmers about their perspective.

So does Apple and their operating system software has supported E-cores for years, first with A-series SoCs and now M-series SoCs. Of course, they were supremely motivated to do so due to battery life constraints. Over 85% of Mac unit sales are notebook models; nothing new, it has been like this for over a decade.

It's harder for Microsoft since they are under pressure to support legacy software and they don't control the hardware.

In the end, Microsoft will figure this out. Today's Microsoft is much different than the Microsoft of the '90s. While they have Xbox, their primary focus is enterprise/datacenter/cloud today. They don't even make much money from Office. I bought a cheap lifetime key for Office Home & Student 2019 for less than $35.

My guess is that some of their datacenter customers are pressing for better functionality from CPUs that have E-cores. Some people here don't seem to understand that the datacenter business is substantial and growing far faster than the gaming PC business (which itself is a subset of the overall desktop PC market).

A lot of this comes down to power efficiency. Institutional and government customers also have strong interest in the performance-per-watt metric, something that many PC gamers don't care about.

 

[Vayra86]

I know right? Progress? Is that what they call what those pesky developers do?

One could argue Win 11 is a major step back, rather than progress.

The user base sure does.

Developers exist to get mistreated for such concepts by management.

 

[cvaldes]

One could argue Win 11 is a major step back, rather than progress.

The user base sure does.

Developers exist to get mistreated for such concepts by management.

Which is why Microsoft needs to put effort into improving the task scheduler and other idiosyncrasies with Windows 11.

Again, it's not just about little ol' Joe Gamer living in his mom's basement.

It's more about that 5,000 Alder Lake (Windows 11 pre-installed) desktop PC order that Dell received from the General Accounting Office's purchasing department.

Microsoft will make more of a widespread impact by optimizing Windows 11 and Teams for some Alder Lake business PC than making sure all the graphical elements on Start Menu have rounded corners on Joe Gamer's RGB discotheque PC.

Hell, from a PC gaming perspective, if Microsoft can get OBS to run better on a CPU with E-cores in tandem with a AAA game versus a comparable CPU that only has P-cores, the Windows 11 task scheduler programmers can high five each other.

 

[dgianstefani]

One could argue Win 11 is a major step back, rather than progress.

The user base sure does.

Developers exist to get mistreated for such concepts by management.

The UI is trash, so is the further erosion of privacy, the underlying codebase is an evolution of 10, and superior in many regards.

 

[ThrashZone]

One could argue Win 11 is a major step back, rather than progress.

The user base sure does.

Developers exist to get mistreated for such concepts by management.

Hi,
Yep striped to the bone and rebuilt 11 isn't bad but you now realize it's 10 with some settings moved around so no biggie use 10 and be done.

 

[openbox1980]

I disagree.

The iPhone / Android market has been swept up by big.LITTLE designs for the last decade. Its a proven design that physically makes sense: efficient cores really use grossly less power for only a little bit loss of performance. (Ex: 12% power usage for 50% performance, which is exactly what cell-phone users want). With the Apple M2 chip being released with TWO DAYS worth of battery life, Intel (who makes most of their money from laptops) is running scared, for good reason. If Intel's laptops are to compete against Apple laptops, they must use the big.LITTLE design, and P vs E-cores is an adequate version of the concept.

The issue is that Microsoft is likely not going to rewrite all their software (ex: Word, Office, etc. etc.) to efficiently use E-cores. And video gamers almost certainly won't use those E-cores either.

So we're in a situation where people won't get the huge battery lives that the Apple-crew is getting, but AMD is striking them from the other end with high-performance Zen4 cores.



I mean, if Microsoft had to spend a few million bucks on developers developing a new scheduler in the OS to handle P-cores / E-cores, it makes sense for Microsoft to try to make money off of it to recoup the costs. Schedulers are hard, and expert computer programmers who know how to write schedulers are expensive.

This is a totally different issue going on. Even in the Android / iPhone world, the proper scheduling of big vs LITTLE cores is very much a dark art and no one does it well yet. The decades of OS theory was built upon the assumption of homogenous (all the same) cores. When you have "some cores are faster than other cores", there's very weird and non-intuitive math that occurs. OS design barely understands power-efficient scheduling (invented for this era of cell-phone apps), let alone this new era of "sometimes power-efficient, sometimes high-performance, based on the User's expectations".

Intel still needs to do something with the power draw of their processors in the laptop market. Most of their laptops are getting no more than 5hrs on battery, while AMD apus are normally getting over 8hrs.

Its why I think e-cores arent really that good, even for laptops.

 

[Assimilator]

Intel (who makes most of their money from laptops)

Completely incorrect.

 

[Dr. Dro]

AMD apus are normally getting over 8hrs.

It's really dependent on the laptop type, build, SKU... Intel's U- and Y- processor SKUs are some of the lowest power SoCs on the market. But perhaps most importantly, battery capacity... my particular laptop has that problem, the battery is so small that it feels closer to a battery backup (such as an UPS) than something you were actually intended to use as a power source for the computer. Buuut it's a gaming laptop, so there's that.

 

[phanbuey]

Intel still needs to do something with the power draw of their processors in the laptop market. Most of their laptops are getting no more than 5hrs on battery, while AMD apus are normally getting over 8hrs.

Its why I think e-cores arent really that good, even for laptops.

e cores are actually less efficient than P cores - they are made for die space - but because they are so slow compared to P/Zen3/4 cores, even though they are 'low power', they actually have to work harder to finish given tasks and therefore use slightly more power. In laptops this is obvious -- the alder lake laptop chips are aimed at performance at the cost of battery life. I think the future iterations will get better at this but agree here -- in their current state they aren't good for laptops.

 

[dragontamer5788]

Completely incorrect.

Unfortunately for you, I know how to read a 10k document.

https://www.intc.com/filings-reports/annual-reports/content/0000050863-22-000007/0000050863-22-000007.pdf

Page 86, breakdown of revenue by Intel's sectors. Client computing group is by far the largest. Of the CCG portion of Intel, the laptop group makes far more revenue than the desktop group.

Intel's #1 segment, by volume, is the laptop portion of their "CCG" / Client Computing Group. Desktop and Datacenter (and other portions) are smaller.

EDIT: That's 25 Billion to Laptops, and 23 Billion to Data-center, and only 11 Billion to Desktop.

 

[R-T-B]

Nope

I mean... you just went on a tangent about process priority and how it can be requested, which has nothing to do with the fact the scheduler still operates on threads, sorry.

The user base sure does.

Meh. It ain't perfect I'll be the first to admit, but I still think most of the whining is from people with no intention to use it longterm. Not really the "user base."

 

[cvaldes]

Meh. It ain't perfect I'll be the first to admit, but I still think most of the whining is from people with no intention to use it longterm. Not really the "user base."

At some point most of them may not have a choice if both Intel and AMD commit to the E-core/P-core route for their consumer CPU stacks. Once properly implemented (hardware, manufacturing node, software, etc.), the pros will far outweigh the cons. We've already seen this in the mobile space.

You can't buy an M-series powered Mac with only E-cores or P-cores. If you want a Mac with only P-cores, you can choose between the Mac mini 2018 (Intel Core i5 or i7) or the Mac Pro (a selection of Intel Xeon). Soon those will go away.

I can't buy an iPhone with only Avalanche cores either.

These whiners may end up eventually being part of the user base whether they "like" it or not. Sure they have alternatives today if they don't want Alder Lake. Someday they probably won't.

My guess is that AMD have been running CPUs with E-cores for several years in some lab in Santa Clara. And half of the people in that building probably have iPhones. This concept of differentiated silicon isn't new anymore. AMD added raytracing cores to RDNA2 GPUs.

Maybe AMD waiting on the sidelines letting Intel stumble through with early Windows support until they can debut something more polished. Maybe it's more difficult getting E-cores and P-cores to play nicely on 2 CCDs. Who knows?

It's not like Dr. Lisa Su is suddenly going to slap her forehead while she's eating lunch at her desk and say "Hey, why don't we have different CPU cores for different tasks?!?" and it'll magically materialize a few months later.

 

[wheresmycar]

Unfortunately for you, I know how to read a 10k document.

https://www.intc.com/filings-reports/annual-reports/content/0000050863-22-000007/0000050863-22-000007.pdf

Page 86, breakdown of revenue by Intel's sectors. Client computing group is by far the largest. Of the CCG portion of Intel, the laptop group makes far more revenue than the desktop group.

Intel's #1 segment, by volume, is the laptop portion of their "CCG" / Client Computing Group. Desktop and Datacenter (and other portions) are smaller.

EDIT: That's 25 Billion to Laptops, and 23 Billion to Data-center, and only 11 Billion to Desktop.

interesting info...

data center revenue... does that include commercial desktops/small business desktop type servers or are these included in the Client Computing Group?

In my view, the strategy of small cores is directly linked to intel's slowness in advancing in chip manufacturing. With no extra space from a denser lithograph, they needed effective cores per area to match the massive MT performance of the Ryzen line up

This strategy is limited by TDP and heating etc...

Yeah this is the sort of stuff discussed when I initially added "Sometime ago I met a self-confessed AMD jock suggesting intels E-cores are just a poor attempt to overshadow AMD's "core count" which can't be achieved with performance cores alone". He did mention limitations in the chips make-up due to higher temps or power consumption hence e-cores are somewhat just filling the gap.

Can't complain though, with or without e-cores ADL's done a fantastic job... it would be interesting to see if e-cores play a more refined role in the long run especially if Intel sticks with monolithic designs... although not sure if this is correct, are earlier rumours of Meteor Lake moving to multi chip modules (MCM) now official or is intel playing DIE HARD with mono-bono-4-life?

 

[AusWolf]

I mean sure, but I also felt the same way when i upgraded to my ryzen 1400 back in the day


You can't be sure it's the E cores that makes it smoother, and not just being a faster CPU overall with more cores/threads

My Ryzen 3 3100 HTPC feels more responsive than my Core i7 11700 main desktop simply because it doesn't have half as many automatically starting background programs installed.

There's many things that make a PC responsive.

 

[mama]

Its not the OS that needs to be rewritten for E-cores or P-cores, its software in general. That's what people don't understand. If Blizzard / Activision decided to optimize for E-cores, then the next release of Overwatch or whatever will be faster on E-cores.

Microsoft can create libraries / algorithms to try and figure things out at the OS level. But its not like the OS is an oracle that knows how programs work, it just collects dumb statistics and makes broad general guesses. Its not Microsoft's responsibility to make more efficient use of E-cores.

----------

And there-in lies the problem. Zen4 cores are more similar to P-cores, so it will make more sense for developers to optimize for Zen4/P-cores, rather than spending efforts making things better for E-cores.

e-cores do not help in gaming.

 

[nguyen]

e-cores do not help in gaming.

Having more than 8C/16T do not help in gaming either though, just look at 10900K/5900X/5950X

IMO assuming same price category

8Cores+3DVcache > 8P+8E > 8Cores