Intel 15th-Generation Arrow Lake-S: A Shift Away from Hyper-Threading Technology?
Click here to post a comment for Intel 15th-Generation Arrow Lake-S: A Shift Away from Hyper-Threading Technology? on our message forum
Crazy Joe
It will be interesting what removing HyperThreading from the CPU core will effectively do for the performance. Will the simplified core design allow Intel to keep thermals under better control and thus compensate for the lost performance of the HyperThreads with a higher core clock? Or will the available space on the Core now allow them to massively increase the number of E cores, which would then be used instead of the HyperThreads and boost performance that way. More L2 cache could serve to more easily handover work between the P and E cores (though the HyperThreads access to the L1 cache of the Core do give them an advantage there).
I hadn't heard about these so called "rented cores", but looking at the article on hardware times on the subject it seems somewhat similar to my second idea (but with a intelligent thread splitter)
fantaskarsef
As long as games run mostly one or a few threads at best, this makes no difference, at least that's what I expect.
Ivrogne
interesting, indeed. As long as it performs much better than previous gen, most people don't care,
Kaarme
Yes, a non-SMT P-core plus an E-core will naturally be better than an SMT P-core alone. So, if Intel keeps spamming those E-cores aplenty, it will offer better performance than having less cores with SMT. Of course, like the article stated, when the shift happens, there might actually be a drop in performance, unless the architectural changes can fully compensate for it, or if the number of E-cores makes a jump.
These days gaming performance has been pretty much everything Intel could have been proud of, so this is probably the best time for Intel to make such a change, as it might actually benefit gaming performance.
Crazy Joe
https://www.tomshardware.com/pc-components/cpus/intels-meteor-lake-cpus-are-slower-at-single-core-work-than-previous-gen-models-new-benchmarks-show-ipc-regressions-vs-raptor-lake
I read an article that claimed that the single core performance of Meteor Lake was worse than that of Raptor Lake. So maybe Intel is already in a single core performance downward cycle?
Horus-Anhur
Maybe Intel's OoO got so good with Meteor Lake, that SMT is no longer needed.
Alessio1989
lmao how to lose even more server share (doesn't matter high performance servers are basically no more).
the big.little.fart triple design seems just a joke.
applications already struggle to use multiple cores of same performance/efficiency, having 2 different types of cores performance is still not handled correctly by OS scheduler (of course it cannot do magic) and even more applications. A third stage of performance would be even worse.
SMT was fine since an application that would need to map high intensive thread to a physical core was already doing that, while everything else could run sharing even the cache state without latency issues..
But hell, we are still waiting the 10GHz NetBurst XD
Glottiz
Maybe this will be my next upgrade. I usually upgrade when Intel launches interesting new CPU and this has served me very wall over the years. I went 2500K -> 5820K -> 12700K. Although upgrade to 15700K might be a bit premature looking at my upgrade cycles of the past heh.
H83
I think there are 2 reasons for this (possible) move:
1sr - security, Intel had a lot of security problems with their past CPUs, many derived from HT;
2nd - small cores, it`s possible Intel wants to replace HT with even more e cores, like Kaarme alreadt wrote before.
Undying
I think Intel will get smoked by the Zen5 9950x with 32threads. There wont be even a competition.
Venix
In theory where everything works exactly as intended tlhyperthreading will schedule tasks on a CPU core part if it can fit when that is not doing anything or waiting for other results to continue..... So on the theoretical part it should just cost performance no matter what . Since this is a real world ht does have negative impact on single thread loads we all know that but not always ... maybe the scheduler improved that much that counting in the space that it occupies in the die + what you gain is a net negative? Do they cut it off to release it as a new feature or premium feature on the 16th gen ? We will see I guess !
tsunami231
if adding more e-cores make up to HT lose then it should not matter, but big problem is the scheduling, you more gen before my original plan of going from 6700k to 16xxx intel really need to get thre TDP in check though
Alessio1989
tsunami231
if they can get it to work it wont matter, again IF
schmidtbag
Not surprising. With all the security vulnerabilities, HT is nearly useless. For one thing, due to the nature of how schedulers work, sometimes HT actually hurt performance, even before the Spectre and Meltdown days. But with all the mitigations in place, you're looking at an average of a 10% improvement. When you start comparing thread count to AMD, HT is just making Intel look worse than they really are. Meanwhile, Intel is trying hard to not do chiplets/tiles for their P cores, and as they continue to add instructions, they're making it that much harder to fit in more cores (which they need to compete with AMD). From what I recall, there aren't too many transistors needed to make HT work, but, caches are needed to be enlarged in order to queue the second thread. Caches not only take up a lot of space but the larger they get, the slower they get. By ditching HT, Intel could probably get away with a faster and smaller cache, and therefore squeeze more cores onto the package.
Games will likely run better. This can be for 2 reasons:
1. The caches need to keep getting cleared for alternating threads, which can slow things down. So long as a core is handling the same thread instructions, that should minimize the amount of rewring that has to be done. Without HT, you don't have this problem.
2. Some multi-threaded games are very particular by the order of the threads, since some have a parent process that requires all threads to synchronize in order to proceed to the next frame. To my understanding, HT threads don't necessarily complete at the same time they normally would in their own dedicated core, so depending on the calculation and what the OS scheduler is doing, you might waste extra CPU cycles as parent thread is waiting for the child HT thread to synchronize. Without HT, ideally, the scheduler would allow each thread to not have to wait on any others.
fantaskarsef
user1
Martin2603PL
Arrow Lake is expected to aim for a 30% enhancement in gaming performance through architectural improvements.
Sounds like another Fairy Tale.
schmidtbag
RealNC
I think they will simply re-introduce HT later again. "We don't need it!" at first, then years later "HT is coming back, HYPE!!"
Off-topic:
Half the front page articles look like this when using the newly introduced dark mode: