Intel Raptor Lake Processor with 34 P-Cores Spotted

[Frank_100]

No. I hope you remember what Phi was?

I still have one squired away.
It was the surviving part from the Sandy Bridge E5-2687 machine that I burned out designing FIR filters.

Xeon Phi - Wikipedia

en.wikipedia.org

Except for having to write my own code,
what part of throwing lots of vector operations at lots of cores is different?
Or am I being too general?

 

[Valantar]

Are you willing to do a benchmarking experiment on your 5800X? I can't , my i5-6600K has had HT disabled forever by its creator. You'd need to run two instances of a single-threaded benchmarking program such as Super Pi at the same time, one pinned (via affinity settings) to virtual CPU 0, the other to virtual CPU 1. Or any pair that belong to the same core. What results do you get, is one instance slower than the other?

I can't find any technical documents or discussion or benchmarking results that would confirm that there exists a "main thread" with higher priority and a "HT thread" with lower priority, so that the main thread would never be slowed down substantially, but the HT thread would "take whatever remains" of execution units and run very slowly. AnandTech has had some great articles on HT since 2002, with this one being the most recent, and there's no mention of the two threads being unequal.

The OS scheduled clearly knows the consequences of dispatching threads to fewer physical cores (preferring HT) vs. more physical cores (preferring no HT), but this is a different matter.

Unfortunately I'm not at home these days, so I don't have access to my main system, and all I've got where I am right now is an old i5-4670S. Otherwise I would definitely test this. I'm reasonably sure the Windows scheduler actively treats HT threads as "lower performance" somehow and mainly allocates lighter loads to these unless fully saturated with high performance tasks - but unfortunately I'm not able to test it until I can go back to Sweden and get my stuff moved to where I am now, which will take a while.

 

[OkieDan]

What do you mean here? When two threads are running on the same core, they have the same performance. The OS cannot change that because a core has no concept of thread priority. Here is some discussion about that.

As you can see in "slide2" image, after you get past the 8th thread on 12900K where it starts to use hyper threading, the 7950x starts beating the snot out of it. Then the 7950x gets to it's 15th/16th thread and stops leaving it in the dust as it has to start using HT. At the 24th thread, the 12900k runs out of E cores and the 7950x continues its lead.

 

[Prima.Vera]

Games wont ever need more than 12 cores/threads, ever. Many game devs are still struggling with using more than 4 cores.

Latest Unreal, Frostbite and Unity engines can all use up to 32 Cores/Threads. Just FYI.

 

[DeathtoGnomes]

Latest Unreal, Frostbite and Unity engines can all use up to 32 Cores/Threads. Just FYI.

Unity I knew used up to 16 but it still needs to coded for anything more. None of them by default, but that could have changed in this last year.

 

[Wirko]

As you can see in "slide2" image, after you get past the 8th thread on 12900K where it starts to use hyper threading, the 7950x starts beating the snot out of it. Then the 7950x gets to it's 15th/16th thread and stops leaving it in the dust as it has to start using HT. At the 24th thread, the 12900k runs out of E cores and the 7950x continues its lead.

Thanks, it's just what I've been looking for. Although the 12900K starts using E-cores before HT. Intel revealed the formula P > E > HT (performance wise), and it's pretty logical, as an E-core has about 60-65% of performance of a P-core without HT. Difference due to HT is less.

Looking just at 12900K results: if I normalize the 8T performance to 100%, next 8 threads (E, not HT) add 45% more, and then next 8 threads (HT) add another 37%.
Looking just at 7950X results: if I normalize the 8T performance to 100%, next 8 threads (not HT) add 80% more, and then next 8 threads (HT) add another 41%.

+45% seems little for E-cores but even on AMD it doesn't scale to +100%, there must be some contention elsewhere in both systems (cache, RAM, boost limits, I don't know).
+37 is quite close to +41%, indicating a similar efficency of HT. Very good actually, and if there was some contention before, there's more now. From what I've read and understood, the effect of HT is the greatest if the two threads are doing very different things, thus competing less for the same execution units. But in MT benchmarks, all threads are given exactly the same code to crunch, so it's a bad case for HT to show its strength.

But there's a but. MT benchmarks just record the total performance of all running threads. If one runs slower than the other on the same core, it won't show anywhere. So these results don't either confirm or deny what I stated before: that both threads on the same core run equally fast (and each is about 30-35% slower than a thread with no HT).

 

[OkieDan]

Although the 12900K starts using E-cores before HT. Intel revealed the formula P > E > HT (performance wise),

If that's the case then it does indeed look like the HT threads are very close in performance to the primary thread under this workload.

 

[Wirko]

If that's the case then it does indeed look like the HT threads are very close in performance to the primary thread under this workload.

CB would have to keep each of its threads on the same hardware thread throughout the test, then spit out detailed results by thread, then we'd know much more.