Skylake CPUs have 'inverse Hyper Threading' to boost single-thread performance | Hacker News Reader

Skylake CPUs have 'inverse Hyper Threading' to boost single-thread performance

myce.com

43 pointskbd21 comments

An interesting idea to ponder, but this isn't a feature present in Skylake or any other processor Intel has announced. The German language source cited in this link retracted their guess a few days later: http://www.heise.de/newsticker/meldung/IDF-2015-Intel-enthue...

"Inverses Hyper-Threading, wie hier aufgrund von merkwürdigen Messergebnissen spekuliert wurde, unterstützt der Skylake indes nicht."

"Inverse Hyper-Threading, as was speculated here due to strange measurement results, is not supported by Skylake."

Here's some English language reporting about the retraction:

"However, Intel has dismissed rumors of Inverse Hyper Threading being present on the Skylake die, something that we talked previously as being a cutting edge methodology of single-thread computing boosting via multi-core help. We expected an answer on this matter at IDF and Intel unfortunately dismisses it as being false. Recent studies made by Heise.de showed incredible single-thread computing performance by the new Core i7-6700K and speculated that Inverse Hyper Threading might be present on the new Skylakes. Apparently, it was not."

http://news.softpedia.com/news/intel-reveals-the-skylake-inn...

Yes I remember the heise reports as well and it was super speculative back then. BTW, your English translation is good. May I ask if that is just with high school German class or have you been to Germany?

I'm American and took a year of German in college, but I cheated here. I scanned the page to find the right sentence, and then copy-and-pasted it into Google Translate. Then I fixed up the English to sound better while trying to stay faithful to the phrasing in German. Without Google's starting point, I'd have been guessing a lot, and probably wouldn't have shared a translation.

AMD did do something like this with their Bulldozer architecture - it had a pair of SIMD dispatch units per pair of processor cores, which could either be split between the cores or used by one of the cores depending on the workload. The public and media perception was that they were lying about the number of cores in the CPUs by doing this, and I believe they're dropping it for the next CPU generation.

I strongly believe that if this feature exists Intel would be making a big deal of it to get people to upgrade - not hide it. Therefor this is nonsense.

The article is entirely speculative. Another entire speculative explanation would be that Skylake simply has better power management allowing single cores to operate at much higher frequencies when other cores are disabled.

Occam's razor (and my background in hardware design and high performance computing) would definitely make me bet my money on the above speculation :)

Or it might just be that when multiple cores are in use, thermal considerations require throttling back from full performance?

That wouldn't explain the jump in performance between generations of processors.

The change from 22nm to 14nm would.

Smaller lithography means less heat dissipation, but the frequency hasn't gone up, the processors executes the same number of instructions per second, doesn't it?

True, but less heat dissipation means less need for thermal throttling. Even when working on a single thread, thermal throttling is likely to kick in - so yep, a die shrink effectively results in higher performance in such a case.

Recent Intel CPUs with Turbo Boost have a frequency range they can operate on, with a spread of over one GHz for some models. Thermal limits dictate how much of this buffer the CPU is able to use, so better thermals do lead to higher frequencies used.

I am a bit outside of my area of confidence but isn't the turbo frequency the max frequency advertised by intel on the CPU? There is not a ratio of 2.5 between these two models so it can only explain a fraction of the delta.

Skylake can also do slightly more instructions per clock, so it is faster even at the same frequency as older generations.

Seven months old. If there were any truth to these claims, more convincing evidence would have surfaced by now.

Had my hopes up for a short while... You'll loose in Dwarf Fortress sooner or later due to "FPS-death" and DF is single-core only.

Is it possible there are shared units between cores? Not a hardware expert but like shared arithmetic units or the like.

Shared units between cores can be done and has been done, by AMD. It's hard; AMD didn't do a great job of it. Intel didn't even try.

And AMD has given up on it, their next-gen Zen CPU will use a more hyperthreading-alike architecture.

Only caches and memory/IO controllers are usually shared between cores.

Not much room for such sharing to exist, every Skylake core can do at least 32 FLOPs per cycle simultaneously.

Besides, it'd be obvious in multi-threaded benchmarks.

My guess is an architecture change that allows more of the 'shared' resources to be used by a single core rather than rigidly enforcing a subdivision of the resources between each core.

Sort of like if you had 2MB of cache per core vs a shared 8MB cache between 4 cores.

Reply on news.ycombinator.com