Incoming Ryzen with Efficiency Cores

[Woomack]

So, the AMD fan boys were very critical of Intel's efficiency cores, what will they do when AMD starts selling them?

Only performance cores are one of the main reasons why many gamers/enthusiasts pick AMD nowadays. I won't hide I dislike efficient cores, and I like how new AMD CPUs act in general. Without a hassle can keep high performance, and with easy 2-3 settings can improve performance. It's maybe not overclock well, but the same is Intel.
As long as I enjoy overclocking and tweaking, for my daily PC, I care about stability, and it doesn't matter if it's overclocked or not, as long as it's fast enough. Efficient cores feel like a waste when I already have 6-8 performance cores, and I guess the same feeling has many other users. This is how the current Intel i7/i9 looks. I could disable all the E-cores, and I wouldn't even care, but if I want a top-performing CPU from Intel for gaming, then I have no choice, and I have to pay for these "useless" E-cores too. AMD went both ways, X3D CPUs for gaming, regular for mixed work, or more professional work. Intel has only one path nowadays and not even the HEDT series.

TBF I still don't see the point of efficiency cores at all, Intel or AMD, maybe if they used them to replace HT/SMT? Anyone care to explain?

HT/SMT is faster than efficient cores. It's, in general, 20-30% slower than regular performance cores, but still, performance cores nowadays go up to 6GHz - efficient cores under 4GHz. There is also no problem with scheduling HT/SMT, while E-cores are sometimes causing lower results than expected.

As I said before. E-cores are a way to balance TDP and give more threads for a modern, multithreading environment. Run task manager in Win11, and you will see how many things run in the background. I assume ( I didn't test it directly) that E-cores are helping more in typical daily work like office applications or web browsing but are pretty bad for software that needs fast main cores, like games. In this case, you can find CPUs with many E-cores handy in mobile or business devices. Lower power, lower heat, still good enough performance.
For me, it's also a way to "cheat" potential users. CPUs are advertised as 10, 12, 16, ... cores. In reality, perform like 6 or 8 cores from the last generation. Higher numbers are the marketing tool, and they simply sell better. Most people buy computers like they buy a fridge or microwave.


Off-topic. I just installed Noctua NH-U12A with offset + no IHS mount on the delidded Ryzen 7600. It's too late today, but tomorrow I will test it again on ASRock B650E PG-ITX with the new G.Skill 2x24GB DDR5-7600 kit. KLEVV 8000 memory kit will be published soon; next will be G.Skill.

 

[mackerel]

How many high performance cores does a typical user really need? 8 cores perhaps? If you're using more than 16 threads, chances are the software you're using can scale well with threads. At that point, you're probably not looking at per-thread performance any more, but overall throughput from all those threads. If these multiple smaller cores can do more work than an equivalent sized bigger core, and/or at lower power, then it is a win overall.

Intel are kinda using E cores to replace HT, since they don't come with HT. Again I like this, since SMT does reduce per-thread performance as a trade off to gain more multi-thread performance.

There will be some niche use cases where you want only one type of core, but there's WS CPUs to cover those that genuinely need it, as opposed to just want it as a play thing like me.

 

[Kenrou]

HT/SMT is faster than efficient cores. It's, in general, 20-30% slower than regular performance cores

Erm... sorry, no? Did you mean E cores? HT/SMT was always advertised giving anywhere from 10%-30% multithreading boost, nowhere near "20-30% slower than regular performance cores". Just to double-check, I ran Cinebench r23 and got 13710 SMT ON and 10137 SMT OFF. There are plenty of games that run slower with SMT/HT ON, depending on core count and optimization (4c/8t vs 16c/32t for example), in my personal case, stock settings, I lose max fps to gain 0.1% and 1%. It might have been somewhat improved in more recent generations than my 3800x, ofc, but not to that extent?

If I follow that speed and architecture logic (AMD if all goes well), I would assume that every E core is, on average and scheduler problems aside, 30%-50% of the performance of a P core, still faster than SMT/HT, even though technically it's "free performance"... I don't know, it's a weird thing they're doing, you would think that this would be more expensive overall, unless MAYBE they do exactly what Intel did and use older hardware for E cores, stuff they had "lying around" and not selling?

 

[Woomack]

Erm... sorry, no? Did you mean E cores? HT/SMT was always advertised giving anywhere from 10%-30% multithreading boost, nowhere near "20-30% slower than regular performance cores". Just to double-check, I ran Cinebench r23 and got 13710 SMT ON and 10137 SMT OFF. There are plenty of games that run slower with SMT/HT ON, depending on core count and optimization (4c/8t vs 16c/32t for example), in my personal case, stock settings, I lose max fps to gain 0.1% and 1%. It might have been somewhat improved in more recent generations than my 3800x, ofc, but not to that extent?

If I follow that speed and architecture logic (AMD if all goes well), I would assume that every E core is, on average and scheduler problems aside, 30%-50% of the performance of a P core, still faster than SMT/HT, even though technically it's "free performance"... I don't know, it's a weird thing they're doing, you would think that this would be more expensive overall, unless MAYBE they do exactly what Intel did and use older hardware for E cores, stuff they had "lying around" and not selling?

Yes, you are right. I don't know why but when I read it now then I messed up my comment. Results are typically 20-40% higher, depending on the environment.
As I remember, SMT scheduling was improved a lot a while ago and the last-gen OS had improvements for that too. They still didn't improve E-core scheduling and also AMD CCD/CCX isn't perfect. There was a 3rd party soft to fix it just after the last CPU release, but it should be in the OS. AMD was pointing out how well they cooperate with Microsoft. Somehow Microsoft didn't release patches for Ryzen 7k CPUs on time when 8-9 months before the premiere were comments about it around the web (based on leaked results made on early samples).

 

[mackerel]

With HT/SMT consider what you're comparing.

If you have 1 core, then 2 threads running through it will typically have some increased throughput, at a cost to reduced individual thread speed. For real world workloads I've seen 0-50% throughput increase. Cinebench through recent generations sits around 30%. Some CPU renderers might hit a peak of 50%. These are "good" cases. I don't know if there is an average for a wide variety of workloads, but I'd expect it to be much lower than these. When HT was introduced in P4 era, I think it was said it added about 5% to core size. To be worth it, the performance boost from it should be above that.

Let's take the Cinebench example, with HT/SMT you get 130% throughput vs not using it. But this means, the two threads are each running 65% that of a single. It's a 35% drop in a single thread. This doesn't matter for rendering like Cinebench, but if that thread was a time critical one in a game, it will impact performance. Turning off HT/SMT can help in situations like that, providing you're not limited on total CPU performance. I feel if we get to ever higher core counts, then dropping HT/SMT might become viable, especially if the overhead of managing threads starts to negatively impact performance also. Not a typical end user problem (yet) but I recall seeing reports of weird Windows things once thread count gets so high.

 

[Archer0915]

Fastest standard cores with the ability to clock way down and disable features/components is better than throwing crap cores in to win a Core war.

Post magically merged: Aug 17, 2023

Oh! Better yet! Socket X AMD64E(fficency)C(o)P(processor).

Give me all the power and let me add an efficiency co processor for simple daily crap.

 

[unsafesteagle]

Do they have those already ?
A system available on the network with 24 ram slots or something?