Intel Speedstep - On or Off? What's best?

[seabreeze]

Just in case you were wondering, here's some results from having speedstep on or off. Having not come across much information indicating one way or the other, I did the benchmarks myself and here they are.

The set up is:
i5-750 2.55GHz 8MB L3
Gigabyte GA-P55M-UD2
4 x 2GB Corsair XMS3 1600C8
XFX GTX285 1GB
SuperTalent UltraDrive ME Series 64GB 2.5" MLC SSD
Gigabyte ODIN 800W PSU
Windows XP SP3 32-bit

Speedstep (EIST) enabled in BIOS and Portable/Laptop selected in Power Options for Speedstep On.
Speedstep (EIST) disabled in BIOS and Destop selected in Power Options for Speedstep Off.

If the Power Options are left on Desktop, it makes no difference whether or not Speedstep is Enabled, Disabled or left on Auto with this system. It will function as if Disabled. Unless the correct Power Option (ie Portable/Laptop) is also choosen when Speedstep is Enabled, then Speedstep will not function.

The following two graphs are from 3DMark06 results (default settings). The first is the total 3DMark06 score and the second is the CPU component from the same run.

For those who do not know this benchmarking tool, it is available from http://www.futuremark.com. You can download the free version to try for yourself.

Please note that I prefer leaving speedstep enabled on all my Intel machines, whether they are my own or builds for clients.

e

 

[randomizer]

Interesting that it caused a major drop in performance. I can't say I've witnessed that with C-states, never really checked EIST though.

 

[sminlal]

Interesting that it caused a major drop in performance.

The Y axis of those graphs is not zero, so the differences are exaggerated. It looks to me like the difference is about 2.5% - so I guess it depends on what you mean by "major". I'm not familiar with the benchmarks, but I wonder if it's just the initial ramp-up of the clock speed from idle to full at the start of the test. If not, then I'm guessing it means the CPU clock drops multiple times during idle periods in the test while the GPU is the bottleneck.

 

[seabreeze]

If our power bill is too low, then leave it off.

Still funny...

 

[seabreeze]

The Y axis of those graphs is not zero, so the differences are exaggerated. It looks to me like the difference is about 2.5% - so I guess it depends on what you mean by "major". I'm not familiar with the benchmarks, but I wonder if it's just the initial ramp-up of the clock speed from idle to full at the start of the test. If not, then I'm guessing it means the CPU clock drops multiple times during idle periods in the test while the GPU is the bottleneck.

Yes, the graph was set with a Y-axis of 4000 min and 6000 max for ease of reading. You've estimated correctly the approximate difference, which some people may find hard to do.

Taking that into consideration, here is another image just showing the percentages in gains of having Speedstep Off compared to having it left on:

 

[paranoidmage]

Speedstep (EIST) enabled in BIOS and Portable/Laptop selected in Power Options for Speedstep On.
Speedstep (EIST) disabled in BIOS and Destop selected in Power Options for Speedstep Off.

I think the power options would be the reason for differing performance results, not Speedstep.

 

[paperfox]

^^ ya i was thinking the same, with the power options being different you now have 2 variables to account for not just 1.

 

[seabreeze]

It's not the power options. On this board, with speedstep enabled and power options set to Desktop, it runs just as it would with speedstep disabled.

Without enabling speedstep and selecting power options to Portable/Laptop, the systems will run permanently on the highest multiplier, no changes.

I've already done the benchmarks as follows:

Speedstep On power options to Desktop
Speedstep Off power options to Dektop

There is no difference in results. The both produce outcomes as if speedstep is off.

If I had posted those benchmarks, I would have receive replies of "but the power options are the same".

Apparently, even when sharing some findings, you can't win.

 

[randomizer]

The Y axis of those graphs is not zero, so the differences are exaggerated. It looks to me like the difference is about 2.5% - so I guess it depends on what you mean by "major".

I realised that it was non-zero on the Y-axis. I consider 2.5% to be a pretty major performance drop for simply enabling a power-saving feature. When I tested the effects of C1E on my E6600 it finished a SuperPi 1M run in almost exactly the same time whether on or off. I don't think I tested EIST though.

 

[cjl]

Note that there was a windows power mode change too though. I've tested it on my system before, and I don't think there was a measurable difference. I may get around to trying it again sometime though.

 

[randomizer]

That might be the cause. I was under the impression that EIST and C1E re-evaluate multipliers and voltages quite regularly, meaning that there shouldn't be a measurable difference in performance, let alone a consistent difference.

 

[Mousemonkey]

This topic has been closed by Mousemonkey