E-cores still evolve. But is there a reason for it?

[Macro Device]

Let's pretend marketing, business and all that economy stuff are completely irrelevant. I'm about to ONLY talk engineering aspects of this phenomenon.

From what I've gathered so far (and I might be totally wrong. Correct me if I am):
• Hybrid structure cries for an impeccable prediction mechanism which can never be invented. At least with our current state of knowledge.
• E-cores are mocked by last gen architectures in gaming even if the game is coded so well E-cores actually improve the experience in all aspects.
• E-cores are mocked by P-cores in terms of performance per watt if you downclock the latters to around 4.3 (Alder Lake) or 4.6 (Raptor Lake) GHz.
• Software development is currently in a state that promotes fast releases but doesn't tolerate actual bug fixing if it takes more than a manhour to deploy. Which means scheduling is virtually thrown outta window.
• It's not impossible to land 16ish properly working P-cores on one die and make them feel at home, likely cutting about a half or two GHz all-core turbo so it actually doesn't go kaboom.
• Average Joes and Janes (and attack helicopters for that matter, too) don't have any idea what these cores are actually good at. They render confused at best.
• There's no evidence that heterogenous architecture helps alleviating background loads any better than just throwing more P-cores.
• It seems it's also more complex and failure prone than a good ol' technique of just having X cores of the same arch.

Why did Intel abandon HT (which I don't mind at all and it's not to be discussed in this thread) and not E-cores since they already implemented segmental layout? Is there anything real engineers can see going wrong that I don't? Once again, if it's all only limited to cash and marketing then I don't even know what to say.

 

[ir_cow]

E-Cores are great for low power and background tasks. Hyper-Treading is a old tech for P4 days that is a powerhog.

 

[Macro Device]

E-Cores are great for low power and background tasks.

I've never seen an E-core enabled CPU consuming less power on idle than an equally clocked E-core disabled CPU of the exact same architecture. Was quite the opposite. Why?

 

[_roman_]

Why did Intel abandon HT (which I don't mind at all)

I speculate it's about all those security issues with Hyperthreading.

 

[ir_cow]

Why are we focusing on idle here?

Also have you tested P vs E in workloads? E-Cores comes in clusters of 4, but you can still turn the cluster on and off and see how much lower the power draw is (not much) and how much performance is lost without them or even just a HT replacement for background windows tasks.

 

[Macro Device]

Stop discussing HT please. I know it's had to go. Let's keep E-core topic only.

 

[ir_cow]

Stop discussing HT please. I know it's had to go. Let's keep E-core topic only.

You asked why they removed it in your original question. Edit your post if HT isn't what you want to discuss lol.

 

[Macro Device]

Also have you tested P vs E in workloads?

Yes. I was testing a 13900K and at P-core speeds exceeding 5.3 GHz it was apparent that E-cores (at stock) are more watt efficient. Downclocked to 4.6 GHz, however, P-cores just ridiculed the snot out of these E-cores, sometimes being 2.5 (!) times more productive per watt.

 

[ir_cow]

RTX 4090 & 53 Games: Core i9-13900K E-Cores Enabled vs Disabled Review

In this week's TPU50 Megabench we're testing whether you can unlock additional gaming performance by disabling the E-Cores on a Raptor Lake Core i9-13900K processor. For our benchmarks we used the mighty GeForce RTX 4090, with 53 games at three resolutions.

www.techpowerup.com

 

[Macro Device]

You asked why they removed it in your original question.

Probably me being sleepy prevents from being non-ambiguous. I was asking why not both.

 

[ir_cow]

Probably me being sleepy prevents from being non-ambiguous. I was asking why not both.

Because HT is old tech and a powerhog. It was a way to get more threads when everyone still had single core CPUs

 

[Macro Device]

RTX 4090 & 53 Games: Core i9-13900K E-Cores Enabled vs Disabled Review

In this week's TPU50 Megabench we're testing whether you can unlock additional gaming performance by disabling the E-Cores on a Raptor Lake Core i9-13900K processor. For our benchmarks we used the mighty GeForce RTX 4090, with 53 games at three resolutions.

www.techpowerup.com

This test is great but lacks one mighty important thing. There's no LGA1700 SKU with more than 8 P-cores.

And I'm willing to bet everything I can that slotting just four more thereof would amount to a total and irrecoverable murder of any E-core config in gaming.

 

[ir_cow]

So you would trade 4E for 1P? Give up 2 threads (since E doesn't have HT).

 

[Macro Device]

Because HT is old tech and a powerhog.

Please tell me something that's not known since, like, 2015.

So you would trade 4E for 1P?

In a gaming-only rig? Absolutely. Zero questions asked.
In a productivity rig? Probably I'd rather only trade 3 to 1, 4 to 1 seems a bit too stretchy.

 

[ir_cow]

Please tell me something that's not known since, like, 2015.

Alright well no point in continuing this thread since you know all and have come to a conclusion that can't be presaded.

 

[Macro Device]

since you know all

Okay, let me translate to even more basic English.

1. I already DO know why HT is abandoned. Full stop.
2. I, however, have no clue why E-cores are still a thing. And not in a "you must stop" way, I want an explanation before I draw final conclusions.

 

[ir_cow]

Okay, let me translate to even more basic English.

1. I already DO know why HT is abandoned. Full stop.
2. I, however, have no clue why E-cores are still a thing. And not in a "you must stop" way, I want an explanation before I draw final conclusions.

Because 4-Cores is better than 1P core. Full stop

 

[Macro Device]

Because 4-Cores is better than 1P core.

In raw performance they might be. In performance per watt, also possible.

However...

 

[anonuser57]

The point of HT over other core sharing strategies of sharing a single core is that inorder to exploit instruction level parallelism you need a lot of extra execution hardware. Since going in that direction leaves you with extra unused execution resources not used by the thread you are running, you may as well add additional hardware to make the unused resources usable by another thread. As CPU designers get better at instruction level parallelism they leave fewer execution resources unused by a single thread and the cost benefit analysis swings towards getting rid of HT.

Algorithm improvements seen in big cores can and do trickle down into lower power cores but with the lower power cores you have different tradeoffs you want to make. For example more aggressive speculative execution leads to faster execution times but also more work getting thrown out. So, how you get to something like Haswell like performance today is going to be different from how you got there 10 years ago. Especially if you're other considerations have also changed your cost-benefit analysis.

In short CPU engineering involves a lot of bean counting even before you consider money.

 

[xSneak]

x86 needs small, cheap, power efficient cores to compete with ARM desgins. A basic google search says almost 2/3s of PCs are mobile form factor where battery life, size, and thermals are more important than overall performance. Not to mention even on desktop PCs itself the amount of people who play games is a fraction of users. I doubt there's any real benefit to pc users in going beyond 8 p-cores right now compared to just adding more e-cores.
According to AMD a zen5c "small" core is 25% smaller so you could get 16 e-cores instead of 12 p-cores for the same die space.

 

[SL2]

Any specific reason we're dicsussing this right now? Desktop Ecores has been at a standstill for exactly 3 years, save for some Hz and cache work.
Arrow lake will have the first substantial Ecore improvement according to Intel, and the reviews will be out in less than 72 hours.


One of my speculations is that it's a compromize. Intel wants more Pcores, but that takes up too much space, or generates too much heat. Or, they got some ideas from looking at Qualcomm.

Remember that how the current desktop CPU's are currently running may not reflect on Intel's original intentions 100 %. Hence it's hard if not impossible to tell exactly why they're there to begin with, especially since we've only seen pretty much one generation on desktop so far.

Dont't forget that Raptor lake wasn't supposed to exist at all. Meteor lake was next after Alder lake originally.

 

[AnotherReader]

In raw performance they might be. In performance per watt, also possible.

However...

Even in performance per watt, E cores are worse at the ridiculous settings chosen by Intel. This might have changed with Skymont, but we won't know until it's tested. The real reason for E cores is higher multithreaded performance within the same die area. Your hope of 16 P cores is just that: a hope. In reality, you would get 10 P cores if you removed the 16 E cores from Arrow Lake.

For the buyers of high thread count CPUs, E cores make a lot of sense and are definitely superior to cramming in fewer P cores.

 

[ShrimpBrime]

Because HT is old tech and a powerhog. It was a way to get more threads when everyone still had single core CPUs

It's a way to "get more threads" without using die space.

Doesn't matter how old the tech is. AMD implemented HT (SMT) like 1.5 decades after Intel. So it's obviously very useful, or AMD wouldn't have done it. And it didn't cost die space. Even got room for 3 dimensional cache!!!

 

[SL2]

Even got room for 3 dimensional cache!!!

3D is over, waiting for 4D cache.

 

[Macro Device]

higher multithreaded performance within the same die area.

It's a way to "get more threads" without using die space.

Finally, an explanation that's totally valid.

you would get 10 P cores if you removed the 16 E cores from Arrow Lake.

As a one who doesn't do productivity on my PC I wouldn't mind that AT ALL. However, I agree, gaming isn't the pinnacle of PC use so it can't be applied to everyone.

But... why not making P-cores a little less complex so they take up less space so we get, like, 5 to 15 % less speed per P-core but 16 P-cores instead of 8+16? What is wrong in this approach? It's a genuine question, I'm not saying it's definitely better.