Unbelieveable 50% speed boost in Mobility Radeon 9600/9700

[Gold]

I just witnessed something incredible.

You know how your computer runs right, you know yours better than anyone else. You download the latest Catalysts or other tweaks and just squeeze the maximum performance out of it.

You know how your comp runs, everyone does, you know exactly what fps you get in rthdrbl or in the dawn demo. Or in the ATI demo's Debevec natural light, or in Farcry, etc. You know it is what everyone else with a machine of your spec gets too, because you check. You know your 3d marks, etc.



5 Minutes ago I did something that made all my FPS values shoot up 15%-50%.


I downclocked this computer.



Yeah you read it right.



This is a Sager 5680, a Pentium 4 3.2 Ghz HT laptop, 800 Mhz FSB with 512 dual channel Ram. I had always ran it normally, but then I started looking around for something that would make it run quieter, cooler, longer and extend the life.

Why have it running at 3.2 Ghz all the time when I'm surfing the net? It just puts wear on the components.

So I found a tool to clock it down to say 2.6 or 2.4 ghz,.. more than enough for everything.



I clocked it down (first by hitting get values, then sliding the fsb, then clicking apply) to 2400.68. Any lower and the screen would freeze and it would need a reboot.


Imagine my surprise when, upon loading all my favourite demos, games, to see how much slower they had gotten, my surprise when I actually saw an increase in FPS in all of them! From 15% to 50%.


Before/After values:

Dawn Ultra Demo (1152x964, 2xAA): 17 fps went to 23 fps
Rthdrbl (ultra cool DX9b Highpass demo, default sett.): 29 fps went to 38 fps
Debevec Natural light (1024 0xAA) (the peak after 10 secs): 42 fps went to - 55fps


Even when I had overclocked my Mobility Radeon 9600 Pro insanely, I never got these sky high fps values, plus what I did get I had massive artifacting. Now it's like this is impossible, it just can't be, I checked radlinker, my clocks are still the same (defaults no less).

It just feels like this little trick opened up a whole other 4 pipelines in the GPU.


I can't explain it I got 50% more fps for free by down clocking???


Can people please check this out and tell me I'm not dreaming.



This is the tool

http://www.cpuid.com/download/CG-ICS952607.zip


If people with other laptops wanna try,.. my motherboard is the I865PE, but the file (CG-ICS952607) is listed as for another type of board but it works!, just goto the homepage for a choice of the different versions.




Downclock the 3.2 to exactly 2.4 (no less) to do exactly what I did, and check with fraps.



I still can't believe it.





P4 3.2 HT (running at 2.4)
512 Ram
DX9b
Win XP SP1
Cat 4.7
Mob 9600




A picture is worth a thousand words: Notice the fps and overall low temperature.



This is before:

This is after:

This is entirely reproducible on any mobile P4 system.

This was posted at notebookforums.com an hour ago and is running into it's 5th page.

 

[Bouncing Zabaglione Bros.]

Are you sure it isn't just fiddling with your framecounters? This is how the UT2K4 speedhack worked. Artificially force a downclock of the CPU, and the game gets its timing wrong relative to the player and you only see every nth frame, allowing you to speed everywhere at ten-times normal speed. Maybe downclocking is doing something similiar?

 

[kyleb]

it bet it has more to do with less heat from the cpu alowing the radeon faster.

 

[AlNom]

I bet it's the babe in the wallpaper wanting to take it slow.

(WHO IS SHE?!!?!?!!!!)

 

[Bouncing Zabaglione Bros.]

it bet it has more to do with less heat from the cpu alowing the radeon faster.

That's a very good point. The mobility chips are full of all kinds of downclocking power/heat conservation stuff arn't they?

 

[Humus]

Are you sure it isn't just fiddling with your framecounters? This is how the UT2K4 speedhack worked. Artificially force a downclock of the CPU, and the game gets its timing wrong relative to the player and you only see every nth frame, allowing you to speed everywhere at ten-times normal speed. Maybe downclocking is doing something similiar?

I would say this is the most likely explanation. I don't know how the apps does their timing, but say the app detects the CPU as an 3.2 GHz, but it's really running at 2.4, then for each 3.2 billion cycles the app counts one second, while in reality 1.333 seconds have passed, making it appear as the framerate is 33% higher than it really is.

 

[Gold]

Yes it has to do with the realtime clock bu that is not the strange issue.

The fps is real, not a corrupt counter, I can feel it. (12 year pc gamer)


RTHDRIBL:

But the interesting thing is that mouse movements are faster! Just try the skull (rthdribl) , hold both mouse buttons and rotate,.. you'll see that mouse movements are unaffected, it rotates silky smooth. While it didn't at 3.2 ghz.


Same as what I said for counterstrike, you walk slow, yet head turning is the same speed. This is the crucial dilemma, how can objects turned by a mouse maintain this high fps, (the skull in rthdribl) while scripted movement (rotating balls, walking/running in CS, slow down 10%) due to the reduction affecting the system time running slower than realtime.?


This is the question.

Just try the skull thing in rthdrbl. (o to change objects), please look carefully, you'll notice that the fps in the counter is actually the fps that your getting, it is not a false reading.



It is weird, at 3.2 ghz, when I click and rotate the skull I can see the effect of a low framerate (21fps), I.e. it rotates and swivels jaggedy.

When I run at 2.4 ghz yes the skull auto rotates slower but at 45 fps, but when I click and manually drag/rotate it now I can obviously see it swivelling left right very rapidly (at 40fps). When I click and drag it doesn't swivel at the jaggedy 25 fps but at the smooth 45 fps.


Just download rthdribl and you'll see what i mean.



Your method of calculation would not work if it is a script that depends on the realtime clock, eg a racing game with a laptime counter that is synched to the real world. Or any scripted motion.


Again just imagine a 3d model that auto rotates, but that you can also manually rotate (ala 3Dmax0, or menu screen of f1 Challenge '99 -'02)).


Downclocking slows down the speed of auto rotation (it rotates it 10% more in slomotion) but increases the fps at which it does run from 25 to 40 fps.


But when you cancel auto rotation by clicking on the object and drag it around yourself, it gets dragged at 40 fps! (you can feel this)

This is what seems impossible, for logically as you begin dragging,.. the fps should fall. It doesn't.

So in practice the Mobility 9600/9700 is capable of moving around a DX9 rthdribl model dragged around the screen by you at 45 fps,

while in normal operating condition it does so at 25.


Somebody suggest any way to officially bechmark this and I will do it.
PCmark and Pi are no good, neither is 3dmark03 as 3dmark results stay the same (as per shapeshifters findings) and the cpu tests obviously score less.