Hyperthreading and o'clking: on or off?

I've worked with some HPC clusters (IBM x3550-M3 and x3550-M4 based) and IBM and Dell (I think) recommend hyperthreading be disabled for HPC applications, because their tests indicated that, for the most part, hyperthreading is bad for performance.

I've read that some people suggest that to maximize your o'clk, hyperthreading should be turned off, which should increase the IPC.. I haven't noticed any games using more than 4 cores and none of them come anywhere near maxing out those cores (at least Bioshock, Bioshock 2, Crysis 3, Crysis 2 and Far Cry 3). None of those games ever even maxed out all 4 cores on my old q9550 either.

Disabling HT would not increase a CPU's IPC. In some rare cases, using HT while gaming can slow things down. That said, if you bought a CPU with that feature, seems silly to disable it.

As far as overclocking further... yes and no. Yes you can likely overclock more with it off as there is less heat and you can usually use less voltage too. That said, see last sentence above as the maybe couple hundred MHz difference you wont notice outside of benchmarking.

EarthDog said:

Disabling HT would not increase a CPU's IPC. In some rare cases, using HT while gaming can slow things down. That said, if you bought a CPU with that feature, seems silly to disable it.

As far as overclocking further... yes and no. Yes you can likely overclock more with it off as there is less heat and you can usually use less voltage too. That said, see last sentence above as the maybe couple hundred MHz difference you wont notice outside of benchmarking.

Click to expand...

The Dell whitepaper article I read had tests that indicated that hyperthreading, more often than not, decreased performance for HPC applications. There's only one "study" I've seen that indicates hyperthreading increases gaming performance and his "study" focused mostly on dual core CPU's.

So why not a higher overclock and consequently higher IPC as opposed to 8 threads that are mostly going to be unused?

Then, why did you purchase an I7 with hyperthreading if you want to disable-it?

I don't know about your processor, but mine doesn't want to overclock any further with ht off than it does with it on, despite the 10℃ or so lower temps. I've pulled 12+ hours of prime stability at 5.0 (about 1.45 cpu-z load voltage)with hyperthreading enabled, and I can only get ~10 minutes at 5.1 with higher (about1.475 cpu-z load) voltage and no ht. It just doesn't like 5.1.

But that's on a custom water loop, it might be a different story on mid range air.

With my 2600k, I hit a wall around 4.7-4.8 with HT on.

Disable HT, and I'm at 5ghz stable with same temps as 4.7-4.8ish with HT on.

Whichever is best for you.

For well-threaded things that don't induce cache thrashing (please note that 2nd/3rd/4th gen I series is far more resistance to cache thrashing than P4 stuff. If the dell paper was on P4 based things it has zero relevance to your i7) HT typically gains 20-40%. 50% if you're really lucky, 5-10 is the low end.
If it uses 4 threads of less there will still be a minor gain due to windows using the HT threads.
If it uses 3 threads or less HT won't help.

Bobnova said:

For well-threaded things that don't induce cache thrashing (please note that 2nd/3rd/4th gen I series is far more resistance to cache thrashing than P4 stuff. If the dell paper was on P4 based things it has zero relevance to your i7) HT typically gains 20-40%. 50% if you're really lucky, 5-10 is the low end.
If it uses 4 threads of less there will still be a minor gain due to windows using the HT threads.
If it uses 3 threads or less HT won't help.

Click to expand...

Some of the following white papers on hyperthreading in HPC applications (for Intel Xeons) don't show any gains like the ones you're claiming (at least for HPC applications on Linux):

Performance of MPI cluster applications with Intel®HyperThreading Technology Enabled
Tim Prince, Intel SSG Developer Relations, Sept. 2009
Contributors: Kathy Carver, George Chaltas, Evgeny Borodin, Arvind Amin, Bob Larson

"With the exception of one application which performed significantly better, enabling Intel®HT Technology produces performance in the range from 3% deficit to 6% gain."

Optimal BIOS Settings for
High Performance Computing
with PowerEdge 11G Servers
A Dell Technical White Paper
is from 2010

"This chart shows that enabling SMT had mixed results. SMT helped performance on two test cases –
Fluent truck_poly and Ansys V12ln-1 – to the order of ~7 to 10%. There were several cases where SMT
hurt performance, including Ansys V12sp-5, whose performance fell by 11%. In other cases the results
were inconclusive. SMT did not help performance on five workloads. Therefore, in order to maximize
performance on HPC workloads, it is recommended that the performance impact of SMT be tested on a
per application basis, and enabled or disabled as necessary."

Intel® Xeon® Processor E5-2600 Family
Cluster Solutions Guide

"Hyperthreading/Simultaneous Multi Processing (HT/SMT)
HT/SMT is Intel’s way of scheduling poorly-threaded applications so that a single core can function like two
virtual cores. Use of HT/SMT helps very few workloads. In a few workloads, having HT/SMT actually hurts
performance. Performance increases in those workloads that HT/SMT helps is generally low.
It is advised to set HT/SMT OFF."

I have also seen an IBM whitepaper that I can't find the source for that
states that as more physical cores are available on a processor the possible gains
realized by hyperthreading drop off substantially while the potential losses in performance remain the same.

magellan said:

The Dell whitepaper article I read had tests that indicated that hyperthreading, more often than not, decreased performance for HPC applications. There's only one "study" I've seen that indicates hyperthreading increases gaming performance and his "study" focused mostly on dual core CPU's.

So why not a higher overclock and consequently higher IPC as opposed to 8 threads that are mostly going to be unused?

Click to expand...

disabling ht does not increase IPC. IPC stays the same no matter what clockspeed or if ht is enabled/disabled.

As far as ht in gaming, correct it usually does not help and so.etimes hurts (as rare as it increases fps).

What is an HPC application?

It all depends on the software as I have said earlier. If its not coded to use as many threads as it can, it won't use them and in some, again, rare cases, hurts performance. What you quoted above is saying exactly what I am (per application basis).

Download Cinebench, run it with and without HT. Marvel at the massively higher score / faster rendering with HT.
Download WPrime, run it, marvel at the massively faster calculation time with HT.

Like I said, it varies per program. That paper is based on two programs that don't benefit as much. If you test your program then you will know whether that program likes it.

THERE ARE NO SWEEPING GENERALIZATIONS FOR HT.