News Intel Alder Lake-S 16-Core CPU Shows Similar Multi-Core Performance as Ryzen 5 3600X

[Admin]

A new and unreleased engineering sample of Intel's forthcoming Alder Laker-S hybrid processor has appeared on Geekbench 5.

Intel Alder Lake-S 16-Core CPU Shows Similar Multi-Core Performance as Ryzen 5 3600X : Read more

 

[tlmiller76]

I can't imagine Intel would WANT to release a teaser of the CPU's if this is ANYTHING like what they're expecting they'd actually be capable of. 24 threads that perform on par with last generation 12 threaded AMD CPU? Not exactly going to give anyone massive anticipations for the release. Usually you want your teaser to make it look amazing, this makes it look...pathetic.

 

[jpe1701]

I can't imagine Intel would WANT to release a teaser of the CPU's if this is ANYTHING like what they're expecting they'd actually be capable of. 24 threads that perform on par with last generation 12 threaded AMD CPU? Not exactly going to give anyone massive anticipations for the release. Usually you want your teaser to make it look amazing, this makes it look...pathetic.

I would assume that they are more concerned with getting it to work right with windows and such for now. I would expect though, if I grasp the concept, that in a game or something requiring high performance it would act like an 8 core 16 thread, while using the atom cores to run the background stuff. Should be really interesting.

 

[JfromNucleon]

I can't imagine Intel would WANT to release a teaser of the CPU's if this is ANYTHING like what they're expecting they'd actually be capable of. 24 threads that perform on par with last generation 12 threaded AMD CPU? Not exactly going to give anyone massive anticipations for the release. Usually you want your teaser to make it look amazing, this makes it look...pathetic.

Maybe they might be trying to set the bar of expectation really low but at this point anything could happen in a year.

 

[waltc3]

If this is factual, it puts a decisive end to the notion that "Intel's 10nm = AMD's 7nm"... It doesn't appear to even be close, if this is the case.

 

[mrv_co]

I would assume that they are more concerned with getting it to work right with windows and such for now. I would expect though, if I grasp the concept, that in a game or something requiring high performance it would act like an 8 core 16 thread, while using the atom cores to run the background stuff. Should be really interesting.

This is an interesting concept and I'll be interested to see how Intel markets it, but this approach seems like it would be more applicable to mobile than desktop.

 

[everettfsargent]

"Today's outing doesn't look very imposing, although Intel won't likely release Alder Lake until the second half of 2021, so this may just be a teaser of what Alder Lake could offer. "

Replace 'a teaser' with 'lowering expectations' as that is much closer to the current truth of the matter.

Oh and can we continue to expect further rumors of future CPU/GPU benchmarks at the current, greater then one, ratio?

 

[FakeMike]

These results cannot be indicative of real performance because Alder Lake being the first hybrid CPU will require full support for Windows and apps, of which there is none right now. I might be wrong but I doubt little cores will run at the same time as big cores (either through a power constraint limit or to prioritize efficiency) so 24 threads could only be a paper spec and benchmarking software will need to be updated to accurately measure its performance.

 

[watzupken]

These results cannot be indicative of real performance because Alder Lake being the first hybrid CPU will require full support for Windows and apps, of which there is none right now. I might be wrong but I doubt little cores will run at the same time as big cores (either through a power constraint limit or to prioritize efficiency) so 24 threads could only be a paper spec and benchmarking software will need to be updated to accurately measure its performance.

I agree with what you mentioned. However, this also raises the issue that the app/game will require some changes (not sure big /small changes) in order for the processor to work at its full potential.

The big/little config in my opinion is a double edged sword because while it helps keep power consumption lower when using less taxing applications, it also wastes die space when you just want as much performance as possible. If most gamers are so concerned about power consumption of their PC, they would not have purchase a high end Comet Lake CPU or any of the top end GPUs now that is sucking in north of 300W of power. On mobile devices, it makes sense because there needs a good balance between longer battery life without compromising on performance too much.

 

[spongiemaster]

I agree with what you mentioned. However, this also raises the issue that the app/game will require some changes (not sure big /small changes) in order for the processor to work at its full potential.

The big/little config in my opinion is a double edged sword because while it helps keep power consumption lower when using less taxing applications, it also wastes die space when you just want as much performance as possible. If most gamers are so concerned about power consumption of their PC, they would not have purchase a high end Comet Lake CPU or any of the top end GPUs now that is sucking in north of 300W of power. On mobile devices, it makes sense because there needs a good balance between longer battery life without compromising on performance too much.

I think people are looking at this the wrong way. Hybrid CPU's in a desktop system the way Intel is implementing this isn't really about saving power in the same way it's done in mobile products. It's getting much more difficult to improve per core performance now without making massive cores. The overwhelming majority of workloads for desktops now aren't going to benefit from more than 8 cores, so by going hybrid, Intel can make 8 massive cores to maximize performance for the workloads users are most likely to be encountering. What's the point of hurting yields and wasting die space and power with 4 or more massive cores above 8 sitting idle 99% of time? The smaller cores offer a more space/power efficient way to increase highly multithreaded performance during the rare occasions they occur. This design would allow for larger cores for the lower threaded workloads than would be possible if Intel wanted all the cores to be the same. So, yes, this is about saving power, but not in the absolute sense. The saved power is shifted to help improve lower threaded workloads.

Jim Keller said that Sunny Cove dies were about 40% larger than Coffee Lake dies, and that Golden Cove dies (Alder Lake) will be "significantly bigger" than Sunny Cove, and that the roadmap was trying to achieve 50x more transistors within a few years. Node shrinks aren't likely to keep up with that, so that's where the hybrid design comes in.

 

[TerryLaze]

The overwhelming majority of workloads for desktops now aren't going to benefit from more than 8 cores, so by going hybrid, Intel can make 8 massive cores to maximize performance for the workloads users are most likely to be encountering. What's the point of hurting yields and wasting die space and power with 4 or more massive cores above 8 sitting idle 99% of time? The smaller cores offer a more space/power efficient way to increase highly multithreaded performance during the rare occasions they occur.

Very doubtful, if that was the problem they would have invented/bought a three way hyperthreading to increase multithreading in tight spots which would be much more efficient since the threads would still be running on the full clock speed of the big cores, on the wider cores smaller threads would not even run at a IPC deficit because the core would have enough instructions available and it would take up much less space.

 

[PCWarrior]

First of all this article has a clickbait title to attract and appease AMD fbs. I mean you could also say that Alderlake performs worse than the 6-core i5 10600K. But nope, you had to compare it with AMD’s 3600X, didn’t you? Why? Second, are we really going to draw any negative conclusions from such an early engineering sample? At this stage (11 months prior to launch and we don’t know how recent this ES is – it could well be from 1-2 months ago) the fact that there is even working silicon (and for such a novel technology) to submit for benchmarks is actually impressive. It is surely still buggy in many ways, and we don’t even know how many cores were actually in “fully” functional condition in the first place.

Also considering the low frequency, the scores are actually not that low. The final all-core operating frequency will be at the very least 3x higher than the one it was running here (1.4GHz, 1.4GHzx3=4.2GHz)). So, multiply this score by 3 and you have a score of around 21000. So, even without further improvements in IPC from ironing out the bugs, we are talking about performance that is 25% higher than a 5950X. Also single core score was with just 1.8GHz. Multiply this by at least 2.5x and you get a score of around 2500 – that is 40% higher than even Rocketlake.

Cue to someone pedantic to tell me that frequency and performance don’t always scale perfectly linearly due to cache misses, etc. So I will give my answer in advance. A bit of an IPC improvement from fixing the bugs and a bit of higher final operating frequencies (say 4.4GHz all-core instead of 4.2GHz and 4.8GHz single-core instead of 4.5Ghz) will offset any discrepancy, so my approximation is more than valid. If you don’t believe me let’s wait and see the performance in November 2021 as we should do, rather than trash-talking about an unreleased cpu 11-months in advance.

Hyperthreading (2-way smt) achieves around 25-35% more MT performance compared to no hyperthreading. In other words a “hyperthreaded logical core” has 25-35% the performance of a physical core. A logical core from a 3-way SMT would have even less performance. On the contrary Intel’s approach of adding small physical cores with 25-35% the performance of a big physical core (~0.5x IPC and ~0.6x frequency) ensures that they will act just like hyperthreading would do if it were possible to have logical cores from 3-way SMT having the same per core performance as those from 2-way. And is likely you could possibly have such 3-way hyperthreading but likely the cores would be so huge that it would definitely not be the most optimum solution.

 

[TerryLaze]

Hyperthreading (2-way smt) achieves around 25-35% more MT performance compared to no hyperthreading. In other words a “hyperthreaded logical core” has 25-35% the performance of a physical core.

No it doesn't, 25-25% is what it can do if you test with the heaviest thread possible that uses the most IPC possible, even then it gives you 25-35% more.
If you run a smaller thread that only uses less than 25-30% of all available IPC anyway then that thread will run full speed.
If you had three way and the background threads would only use 10-15% each than they would run full speed as well.
As long as they all use different instructions or at least the mix that is available on the core.

Raytracing or video converting etc is not the kind of thread that you would call background tasks.

 

[PCWarrior]

No it doesn't, 25-25% is what it can do if you test with the heaviest thread possible that uses the most IPC possible, even then it gives you 25-35% more.
If you run a smaller thread that only uses less than 25-30% of all available IPC anyway then that thread will run full speed.
If you had three way and the background threads would only use 10-15% each than they would run full speed as well.
As long as they all use different instructions or at least the mix that is available on the core.

Raytracing or video converting etc is not the kind of thread that you would call background tasks.

If you run a light thread then you can practically think of it as running on a physical core without invoking hyperthreading. In any case you are missing the point. Why do you care about light background tasks running on a slice of a big core or even on their own physical cores at full frequency when they can be handled just as well by little efficient physical cores running at lower speeds?

The goal of Intel in desktop is to increase ST and MT performance. The latter is required for heavy tasks which can scale with more threads. You can of course add more big physical cores but that is not an option due to cost and yields. Then you either do 3-way SMT or you add little physical cores (which add another 25-35% of performance like SMT would do in 2-way mode). Intel went with the latter for multiple reasons, one of them besides yields/cost also being that they prefer to have a unified approach in the mobile and desktop cpu space.

Anyway, what Intel is doing is like offering a regular 10-11 core CPU with 2-way SMT. And on top of that, these “10-11 cores” will have an IPC that is 50-60% higher than Skylake. So, this Alderlake cpu should be comparable in MT performance at least with a 16-18core Skylake. And by the looks of it, in some workloads at least, it will probably have performance comparable to an OCed 20core Skylake. Not bad.

 

[TerryLaze]

Why do you care about light background tasks running on a slice of a big core or even on their own physical cores at full frequency

The goal of Intel in desktop is to increase ST and MT performance.

You answered your own question, a light thread running at full speed under HT at full clocks is faster ST than running it on a lower clocked core on its own.