RealTemp General Discussion

[somebody]

• 6.92 & 6.93 do not display the IDA multiplier, they only go up to 8.500.
• With IDA off and full load 6.91-6.93 display 8.500.

Your Ratio.zip file downloads as a zero length file so could not try it. Not to worry, I made up my own crude one using 0x30A and 0x30B, see attachment. Also I used the architectural performance MSRs 0xC1 and 0xC2 set to read a range of unhalted core and unhalted bus cycles which enables us read the effective CPU multiplier just by simply dividing one by the other, however with low loads and fixed high multiplier (EIST disabled) the results don't look that flash. I don't know what causes this, maybe it gets messed up with transitioning in and out of sleep states many times, any ideas?

 

[unclewebb]

I renamed the archive RatioCalculator and now it seems to download fine. Not sure what happened before with the 0 byte archive.

http://www.fileden.com/files/2008/3/3/1794507/RatioCalculator.zip

Time to look at your results.

Edit: OK, finally, it all makes sense. It looks like the default multiplier for the P8400 is locked at 8.5 and when Turbo is engaged, the ratio does go beyond 1.0 just like how a Core i7-920 goes up to 1.05 when turbo is enabled. Now I just need to find an MSR which always returns 8.5 regardless if IDA is enabled or not. The value in the high bits of MSR 0x198 flops back and forth between 8.5 and 9.0. If I can't find an MSR that has 8.5 then I guess I'll need to use the trick you suggested by momentarily disabling IDA, reading the maximum multiplier in 0x198 and then if IDA was previously engaged, re-engage it. Time for version 6.94.

At full load you get a very steady multiplier using this method. At idle you don't. I think this is partially do to the fact that these counters are not moving nearly as quickly at idle so the amount of error is greater. Once 6.94 is done you can send me a log file so we can see how things look. At times when there is a slight load on the processor it can be rapidly cycling between two very different multipliers. With a processor that supports SLFM, this probably makes things at idle even worse. The reported multiplier may not be 100% accurate at idle but it's the most accurate multiplier possible from this CPU. The method where you read a single MSR once per second may look nice, but it usually covers up what's really going on inside the CPU.

Edit#2:

Here's version 6.94 that should use 8.5 for the default base multiplier for your P8400.
http://www.fileden.com/files/2008/3/3/1794507/Turbo.zip

I've been having some problems with the FileDen site not letting users download. It gives the 0 byte empty file error. I'm using FireFox and the error might be because of it. Clearing out your cache and clearing out all previous versions of Turbo.zip might work and if that doesn't work, I also uploaded it to SendSpace.com

http://www.sendspace.com/file/ipndt0

Send me a log file or two if you see anything that doesn't look right. With the base multiplier being a fixed value, this method should work just as well on a P8400 as it does on any other Core 2 or Core i7/i5 CPU. A little bit of random fluctuation at idle is normal but if there is more than this then it is likely C1E or the minimum processor state needs adjusting or the amount of background activity is causing the multiplier to be constantly cycling. My E8400 typically only shows a multiplier variation of +/- 0.002 at idle when everything is set up correctly.

If the EIST lock bit in your CPU is not set, then you should be able to toggle the EIST bit if your CPU/motherboard allows this. My motherboard doesn't support this feature so hopefully you can test that out too. If the lock bit is set then the EIST flag will continue to be grayed out. I'm not 100% sure yet that toggling the EIST bit[16] actually toggles the EIST state of the CPU or not. Testing on a Core i7-920 was not conclusive. It didn't seem to change the SpeedStep status so this feature still needs some testing.

 

[somebody]

Congrats, seems like you've cracked it with 6.94. EIST checkbox works also but users may need to adjust the multiplier manually since deselecting it could possibly leave the multiplier set at anywhere from the lowest to highest. Hope it doesn't cause any problems.

I wonder what sort of max ratio the new mobile i7's will have, 1.6-1.77?

 

[unclewebb]

Finally.

I appreciate the help. I had another user check version 6.94 on a T9400 and it also seems to work correctly. When running a single thread of Super PI, rge on XtremeSystems was getting the multiplier to cycle between 9.5 and 10.0. The average on the hardest working core would get up to about 9.9 while the other core was at the default of 9.5. With 2 threads of Super PI running, there was no dynamic acceleration going on so it was locked at a maximum of 9.5 like the Intel docs say it should be. I'm interested in checking this on a P8400 to see if yours works differently than other P8400 mobile chips.

I haven't decided yet whether to allow users to toggle EIST or not. On Core i7 Desktop chips, it doesn't seem to change anything. I hope to play around with this on a friend's P8400 before deciding whether to keep this option or disable it and leave the EIST bit as display only.

Little off topic but the Dell Studio XPS 16 now has an option for the i7-720QM or the i7-820QM. Both are Quad core mobile processors and both support hyper threading.

http://processorfinder.intel.com/List.aspx?ParentRadio=All&ProcFam=3168&SearchKey=

i7-720QM 1600 MHz ~ 12.0X multiplier
Turbo Boost Max = 2800 MHz ~ 21.0X multiplier

i7-820QM 1733 MHz ~ 13.0X multiplier
Turbo Boost Max = 3066 MHz ~ 23.0X multiplier

It's looking like one speedy laptop when running single threaded apps. The way of the future seems to be moderate default MHz to increase battery life and lots and lots of turbo boost available when you are working on something important. Maybe my wife will buy me a new toy so I can make sure that i7 Turbo works with these new CPUs too. I'm pretty sure it does but I'd still like a new toy.

Edit: If anyone has access to one of these you can let me know how i7 Turbo works on it.

http://download.intel.com/design/processor/specupdt/320767.pdf

 

[unclewebb]

RealTemp 3.35

I transferred the i7 Turbo code I've been working on back into RealTemp. If you have a new Core i7/i5 CPU that supports multiple bins of turbo boost when C3/C6 is enabled or Intel Dynamic Acceleration then post your results while running a single threaded task like Super PI 1M. This new version should also support those new Core i7 mobile CPUs I mentioned above.

I'm hoping to see RealTemp report the multiplier on the hardest working core which more accurately reflects what your CPU is doing.

http://www.fileden.com/files/2008/3/3/1794507/RealTempBeta.zip

There have been a lot of changes over the last few months so I need some feedback before uploading this to TPU as the next official version. If you have any problems, let me know. Thanks.

Include your CPU model and OS version as well as 32 or 64 bit. No problems so far on XP and Vista x86 as well as Windows 7 x64. Show some screen shots so I can explain what this new multiplier means.

 

[LocutusX]

Greetings,

I tried out your new RealTemp 3.35 - but really, I was using 3.30 before that, and 3.0 before that although I didn't post in this forum or Xtremesystems. Now I decided to post my result, mainly because I did a Sensor Movement Test and thought I would share the result because I still don't understand what it really means.

My CPU: Core i7-920 D0 stepping
Stock Intel cooler
Mobo: EX58-UD4P
RAM: G.Skill 6GB DDR3 (2GB x 3)
Vista x64 SP2

Generally speaking, stock speed with some minor tweaks:
it runs at 2798mhz (133 x 21)
Vcore is auto
QPI is 4.8GHz (x36)
Uncore is 3.2GHz (x24)
Ram multiplier is x12 (1.6GHz)
BIOS: Turbo mode enabled, EIST disabled, C1E state disabled
All Vista power management features disabled (so CPU is set to 100%-100%)

 

[unclewebb]

LocutusX, welcome to TPU.

Your cool down test looks normal to me. I find that Core 0 on the Core i7 CPUs is always the most accurate from idle to TJMax.

I believe that Intel sets TJMax slightly differently on the other 3 cores to better control thermal throttling but I don't have any Intel documentation to back that up. Intel has publicly admitted that there is a +/- 5C variation at TJMax on some of their processors but they haven't said if this is deliberate or just the nature of these sensors that there is some variation in them.

Core 1, core 2 and core 3 are very similar on your CPU. If you want to do a simple calibration I would set all of them to TJMax = 103C and leave core 0 at TJMax = 100C. This will line things up better so your 4 cores report similar temperatures when equally loaded with Prime 95 Small FFTs. I think your reported temperatures will be more accurate if you do this but we're only talking about a couple of degrees so it's a really difficult thing to prove.

If you are using default MHz then why not enable C3/C6 in your bios so your CPU can access the 22X multiplier when you are running a single threaded app.

Hopefully you can send me a screen shot of the new version of RealTemp along with CPU-Z while running a benchmark like Super PI 1M.

http://www.techpowerup.com/downloads/366/Super_PI_Mod_v1.5.html

I adjusted RealTemp 3.35 so it should more accurately show your peak multiplier on the core that is working the hardest. With C3/C6 enabled in the bios, your CPU should show a multiplier beyond 21X. It will cycle continuously between 21X and 22X as SuperPI is running so RealTemp will report the highest average multiplier on the hardest working core. Background activity that activates a second core will drop it down to 21X but when 3 of your 4 cores are in the inactive C3/C6 state, the core running SuperPI will be able to use the 22X multiplier. CPU-Z doesn't always report this very accurately so I'm hoping to see some screen shots of this and how the new version of RealTemp reports it. Thanks.

 

[LocutusX]

Sure, I will happily try that out if there's a possibility I can get 22x multiplier without overclocking! ... but just to clarify:

Currently I have only TURBO MODE enabled in the BIOS.
I have all 3 of the following settings disabled: EIST, Enhanced Halt (C1), C3/C6/C7 State Support.

Which should I enable for your test?


BTW, if I let Prime95 small FFT 8-thread go for 35 minutes... I tend to get 78-80/74-75/74-75/74-75. Sound normal to you? Also, OCCT (linpack) seems to generate temperatures about 3-4 degrees higher.

 

[unclewebb]

Your temperatures all seem normal. The OEM heatsink is adequate but not much more than that.

To get access to the 22X multiplier, you need to enable C3/C6/C7 in the bios. The inactive cores need to be able to enter these sleep states so the one remaining core can access the 22X multiplier when running a single threaded task.

I'd enable EIST. Some motherboards don't turn this off correctly anyhow. If you don't want your multiplier dropping down at idle then after you enable EIST in the bios, go into your Control Panel -> Power Options and set your Minimum processor state to 100%. Enabling EIST lets turbo mode work the way Intel intended and setting the Minimum processor state will allow you to get the idle multiplier that you would like. You can play around with C1 in RealTemp to see whether you want that enabled or not.

Most users get better idle temps when they enable C3/C6 because when 3 cores have virtually no current flowing through them, they can run a lot cooler.

I think most of your difference in core temperature when running Prime95 is simply a difference in TJMax for each core that Intel set at the factory. Your results are showing me that. Go into the Settings window or RealTemp and try setting TJMax to 100, 103, 103, 103. This doesn't change the temperature of your CPU, it just changes how your temperatures are reported. Hopefully more accurate.

 

[unclewebb]

Here's an example from rge at XtremeSystems.

An i7-950 has a default multiplier of 23. When 2, 3 or 4 cores are in the active state, it can use a +1 turbo boost so the maximum multiplier will be 24. When a single core is active, it can use a +2 turbo boost so then it will be able to use the 25 times multiplier.

What actually happens is that the multiplier will cycle back and forth between 24 and 25 hundreds of times a second. With one core active it will be at 25. As soon as a second core wakes up out of the C3/C6 sleep state to process some background activity, it will become an active core and the maximum multiplier will be limited to 24. When that background task is finished, the second core can go back to sleep (C3/C6) which allows the multiplier to once again jump back up to 25. The less background activity you have running on your computer, the bigger percentage of time you will be able to run while using the highest possible multiplier. This feature is only available when C3/C6 is enabled in the bios.

RealTemp 3.36 now reports the highest multiplier on the hardest working core when this is happening which is what Intel recommends software should report.

I guess you could also disable 3 cores in the bios so only a single core was ever active but that kind of defeats the purpose of owning a multi-core processor. That might be handy though if you're just trying to run a single threaded benchmark at the highest possible multiplier. A Core i7-920 should be locked on a 22X multiplier when 3 cores are disabled in the bios and turbo mode is enabled. Hope that helps you understand how these things work.

For those that like to overclock the BCLK, most Core i7 X58 motherboards disable this feature as soon as you overclock by 1 MHz. The Intel X58 board is one of the few that correctly supports multiple bins of turbo boost, even when overclocking. Most of the new P55 boards support this Intel technology correctly, even when overclocking.

I think RealTemp or i7 Turbo will give you a much clearer look at your multiplier and CPU speed compared to CPU-Z which rounds things off and doesn't tell the whole story when a CPU is lightly loaded.

 

[LocutusX]

Hi unclewebb,

OK, so I first enabled all 3 of the BIOS features I mentioned before: EIST, C1E, C3/C6/C7.

Strange Behaviour #1: Upon booting, I checked cpu speeds in RealTemp and my multipliers were really bouncing up and down, even while idling! It would hit 12.0 briefly and then bounce up and then down again for a second and then up again! Yes, I was truly idling (I waited 10 minutes after Vista boot; killed un-essential processes like AV; HDD light was off so no superfetch thrashing).

Then I went into Control Panel | Power and set my minimum processor speed to 5%. Now, my multiplier wouldn't bounce so much while idling - it would just sit at 12.0. That sounds to me like more reasonable behaviour then going up and down. Thoughts?

Now, on to the test I did.
I got your new version of RealTemp and let that run alongside Super Pi Mod which did a long pi crunching routine. As a sidenote I tried to disable many extraneous processes as mentioned above, including AV scanner, Sidebar, etc.
Unfortunately the 21.5 mult you see in the screenshot below is the highest it ever went up to. The *average* was more like 21.3 or 21.4.

Further thoughts:

I likely won't keep all 3 of my power-saving BIOS features enabled. 2 main reasons at this point:

(1) Reduced performance. I ran Super Pi's 1M test a number of times and it was always 5-7% slower than with those same BIOS features turned off. Which is interesting because my multiplier only goes to 21.0 with them off! So even with access to 22.0 multiplier it can't match my old times.
I swear that my computer also feels a bit slower in terms of general usage, but that's a bit hard to prove.

(2) a strange high-pitched tweaking noise coming from somewhere in my case - no idea where, not sure i have time to troubleshoot!


Do you think that if I maybe selectively enable those bios features, I can eliminate #2 and get an improvement in #1?

 

[LocutusX]

off topic but isn't it annoying how much thrashing vista does after every reboot? obviously with all the tweaking/testing i'm doing I have to reboot a lot and it's annoying seeing all the load on my system for almost 10 minutes after boot!

it seems like SuperFetch wants to fill all of my 6GB with random crap from my HD whenever i boot!

 

[unclewebb]

All I can say for number (2) is, it wasn't me!

Makes you wonder about Intel Turbo Boost Technology. Being able to go beyond 21X looks more like marketing hype. The amount of extra performance you get with the occasional 22X multiplier is being counter acted by the amount of multiplier switching going on within the CPU.

The multiplier bouncing up and down that RealTemp reports is actually going on inside your CPU when your software and hardware power settings are not in agreement. As you found out, when they are in agreement, then Speedstep can work correctly and lower your multiplier down to 12.0.

What you might want to try is to enable EIST and C3/C6/C7 and disable C1 and then set the Minimum processor state to 100%. This might help you get a steady multiplier at idle that doesn't drop down to 12.0. You might get better Super PI times when doing things this way.

For consistent testing, turn off CPU-Z and go into the Task Manager and right click on the Super PI task and select Set Priority and set this to Realtime. That usually helps with consistency when testing.

Edit: That's why I neutered SuperFetch and some other wonderful Microsoft ideas. Without doing this, it can be very difficult to get repeatable results when testing with Vista. Try a couple of XS Bench tests as well. It's usually pretty consistent.

 

[unclewebb]

I had my friend rge do some Super PI testing. He has a Core i7-950 which has a default 23 multiplier and a 24 multiplier when turbo mode is enabled. When C3/C6 is also enabled, he gets the multiplier cycling between 24 and 25 as shown in his screen shot above. Here are his results:

Stock FSB with EIST/C1E on... C3,5,6 off
1M
12.724, 12.656, 12.674, 12.724, 12.677, 12.795

Stock FSB with EIST/C1E on and C3,5,6 on
1M
12.367, 12.468, 12.511, 12.370, 12.380, 12.439, 12.464

Not a huge boost in performance but definitely an improvement during his testing.

That tiny increase in performance explains why Intel has +4 bins of Turbo Boost available on their new i5-750 and their i7-860 and i7-870 both have 5 additional bins of Turbo Boost available when C3/C6 is enabled. A user is more likely to notice a performance increase with that.

Multiplier .. Default --- Maximum
Core i7 920 .... 20 --- 22
Core i5 750 .... 20 --- 24
Core i7 860 .... 21 --- 26
Core i7 870 .... 22 --- 27

 

[LocutusX]

another plus is that OCCT linpack is x64 - it tests all of the CPU, and all of the ram without leaving parts untested like the x86 ones do

but Prime95 has an x64 version too... thats what I use.

 

[Mussels]

but Prime95 has an x64 version too... thats what I use.

it does NOW, it never used to.

 

[wakko709]

do you know if theres

an amd version

 

[Mussels]

an amd version

there is not, and never will be.

 

[andrewsmc]

My i7 920 is only showing 2 cores.... :\ Halp?

 

[unclewebb]

Have you tried the latest version of RealTemp?

http://www.fileden.com/files/2008/3/3/1794507/RealTempBeta.zip

Does the Task Manager show all of your cores?

Post some screen shots so I can have a look.

 

[andrewsmc]

i will just a bit later today.

 

[andrewsmc]

Here.

How on earth do i see my temp on the new realtemp? Am i just crazy and missing it?

 

[unclewebb]

This isn't a RealTemp issue as CPU-Z is also only detecting 2 cores. It shows 4 threads because you have hyper threading enabled but only 2 cores. CPU-Z should be showing 4 cores and 8 threads.

Either your bios needs adjustment so that you can use all of your cores or you need to make an adjustment in Windows. I think in Vista / Windows 7 you have to run something like msconfig and uncheck a box so it will find all of your cores. If you have this set to a specific number it might not be able to use them all. I'm on XP at the moment but time to go fire up my Vista machine. I'll post a screen shot shortly of what I'm talking about.

Download the RealTemp version that I posted and use that. It has a few extra features for the Core i7 CPUs.

 

[andrewsmc]

I think i know the problem... Give me 5 brb. Also.. I did d/l look at my previous post.. I added the pic.

 

[unclewebb]

Here's an urban myth that screws up people's computers. Someone said setting this will speed up your computer but all it does is limit the number of cores Windows can use. Un-check the box and hopefully all your cores and threads will be available after a re-boot.