Wendell, Have you tried disabling SMT?

With the Intel debacle I just thought that SMT might have something to do with the 13/14900K/S frying. With their 15th gen they’re removing SMT out of the blue and MAYBE it’s because they found that with it on you get the problems of 13/14th :man_shrugging:
Intel is stating something about power efficiency as the reason to removing SMT, but I don’t always believe corporations reasonings.
I can’t remember if people were still finding that their CPUs were still frying even with those baseline power profiles that were released for the BIOS, but if they were would you think Disabling SMT might help with that??
I doubt you would be able to to look through those game logs you got to see if they were using SMT, but if you could do you think it might show something?
Also, could it be some type of manufacturing defect like a specific alloy they’re using to cut cost or a chemical in the resist that leaves the pH just a little bit low and over time breaks down an alloy? I know there is no way of finding this stuff out without a REALLY good whistle blower of course, I just wanted to put this idea(and the others) out there. :slight_smile:

1 Like

SMT or multi-threading was a workaround for vendors to pump up their thread count without adding on-die memory for each thread.

There are applications where it is beneficial, but for over a decade best practice in the enterprise space was to disable hyper-threading and watch the benchmarks soar.

This has to do with memory locks during actual workloads, not synthetic benchmarks as servers see erratic data loads.

Not out of the blue, we’ve been instructed to disable it as above for performance and security in the enterprise space since day 0.

There was a white paper in the early 2010’s regarding speculative execution attacks. It has not gotten better and hyper-threading is the same.

2 Likes

Also, no.

Having tested this ourselves, SMT does not affect system stability. Our experiences point to memory related errors with decompression benchmarks being exceptionally problematic.

3 Likes

Would be peculiar, after ALL the years HT existing, having it now be inherently bad [at jump]
HT initially being alternate means, of pushing thread ct [w/o applying more cores]
But the security flaws and net-performance losses, have only been escalating

1 Like

SMT takes an extra 5% of die space, but adds an average of 15% more performance and potentially up to 40% more performance in ideal circumstances.
But branch predictors and cache sizes are getting much better and bigger, so there is less stalling, so less performance gains from SMT. Also, we are getting much higher core counts, especially efficiency cores. So with average performance of SMT getting lower and lower the gains just arent worth 5% extra die space per core anymore, sometimes you even lose performance because of thread contention, and by saving that die space you can add in more efficiency cores instead which are real cores and bring actual performance. That is why Intel decided to “remove SMT out of the blue”.

The problem is voltage related. The i9’s require too much to get their high clocks and request insane levels of voltage, a voltage that overclockers have long said would degrade the CPU over time. Yet Intel decided it was ok, and it turns out the overclockers were actually right. That is why turning the core’s max speed down to a lower level brings stability back somewhat, because it only requests such high voltage above 5.3GHz.

2 Likes

So we have been successfully undervolting the cpu and overclocking, one client is running 2x i9-13900k’s with all cores locked to max turbo on an undervolt, but the memory is capped at 5600 for stability.

3 Likes

This topic was automatically closed 273 days after the last reply. New replies are no longer allowed.