R600 overclocked - Sven Olsen

[otispunkmeyer]

http://www.dailytech.com/Overclocking+the+R600/article7044.htm
its from DT

Sven Olsen's blog on DT - details some OC'ing and he states that hes currently benching an HD2900XTX

not sure if you lot have seen it and im not sure if this is a pea roast but here goes

he wasnt quite able to push 2.0ghz from the memory, but then again this isnt the top card (i hope)

however it seems to clock nicely.... going a full 100mhz to 845mhz.

with that core clock and a memory clock just shy of 2.0ghz its pretty much necking it with the 8800GTX.

this to me, kinda highlights the efficiency of the nvidia part, i know this is 2nd rung down card vs a top rung card, but 200Mhz extra core speed, and a healthy amount of extra memory bandwidth (albiet hooked to 256mb less memory) and as it stands now... its just as fast as a possibly more expensive 8800gtx

now the stock clocks for the XT are 745Mhz and 1600Mhz.

what were the XTX clocks? im guessing 2.0Ghz or more on the memory, and 1gb of it, but will the core come a full 200mhz higher?

 

[tuteja1986]

Coolies

 

[NoStateofMind]

Wow, those are some insane clocks....

 

[CrystalBay]

groovy...

 

[TecHNooB]

If the XTX version doesn't absolutely destroy the 8800GTX.. there better be all the extra goodies mentioned/rumored (built in HD-sound, physics processor, etc). The R600XT should have killed the 8800GTX by itself :| That would have got everyone drooling. Oh well..

 

[Keysplayr]

Just be mad glad that it at least exists and will be here soon. (Hopefully).

 

[Cabages]

I really cant wait for the benchies.

I think a gb of gddr4 is just insane, lets see if it really makes a difference.

 

[yacoub]

hahah wow, a couple retards who can't read responded to my comment in that thread.

I said I prefer a card that is clocked lower and runs cooler and quieter and also happens to outperform the higher-clocked card even when that card is overclocked.

Wrong thing to say, apparently. :roll:

 

[chizow]

Not sure what to make of those results....its good to see the XT has some OC'ing headroom but no one else noticed it looks like its just peddling faster and not going anywhere? I'd like to see if those are actual clock speeds, since the G80s have speed straps/steps before they hit actual increases in clockspeeds. For instance you can set your OC to 660 and your actual core would be 648 etc.

Just seems really odd there's only a 450/200 point 3dmark difference at 1280/1600 respectively for a 100MHz core and nearly 200MHz memory overclock. Either the actual clockspeeds aren't changing much or the R600 doesn't scale particularly well at higher speeds.

 

[AnotherGuy]

Originally posted by: yacoub
hahah wow, a couple retards who can't read responded to my comment in that thread.

I said I prefer a card that is clocked lower and runs cooler and quieter and also happens to outperform the higher-clocked card even when that card is overclocked.

Wrong thing to say, apparently. :roll:

So u prefer spending $550 instead of $450 for 2 almost identical performance cards....

Godspeed!

 

[Matt2]

Originally posted by: chizow
Not sure what to make of those results....its good to see the XT has some OC'ing headroom but no one else noticed it looks like its just peddling faster and not going anywhere? I'd like to see if those are actual clock speeds, since the G80s have speed straps/steps before they hit actual increases in clockspeeds. For instance you can set your OC to 660 and your actual core would be 648 etc.

Just seems really odd there's only a 450/200 point 3dmark difference at 1280/1600 respectively for a 100MHz core and nearly 200MHz memory overclock. Either the actual clockspeeds aren't changing much or the R600 doesn't scale particularly well at higher speeds.

Agreed.

Something is not right. With a 100mhz core increase and a 200mhz memory increase, the scores should have been through the roof.

 

[swtethan]

i had a jump from 500mhz to 610mhz and my 3dmark scores jumped something like 1000+ pts

 

[dreddfunk]

I'm not an expert but that does seem a bit odd, chizow.

I entered all of the numbers into an Excel spreadsheet to do some calculations and it seems that the HD2900XT 512MB may need a balance between memory overclocks and core overclocks. Below is a chart that shows the respective % overclocks on core and memory with the % increase in 3dMark performance.

---------------------------------------------------------------------------
------------------Core--------Memory--------12x10--------16x12----
---------------------------------------------------------------------------
800/900---------7.38%-------12.50%--------2.27%-------1.69%----
---------------------------------------------------------------------------
845/950--------13.42%-------18.75%-------2.66%-------2.00%----
---------------------------------------------------------------------------
845/995--------13.42%-------24.38%-------3.38%-------2.62%----
---------------------------------------------------------------------------


In looking at how increases scale with overclocking, the matter is a bit difficult, as the data haven't given us much information with one of the two variables (core, memory) controlled.

I could give a lot more charts but the best results here are when the core is overclocked at slightly more than half the memory overclock.

The first jump (800/900, OC1), an average OC of 9.94%, returns about 2.27% improved performance on 3dMark at 12x10. That's a ratio of about 0.23--or getting 0.23% performance improvement for each 1% in averaged core/memory overclock.

When OC1 is taken as a baseline, however, the additional 6% increase to both core and memory in OC2 (845/950) only yields an additional 0.39% performance, for a dismal ratio of 0.06% performance improvement for each 1% of averaged OC (after the initial OC, of course).

Yet OC3 fares better even though it is a smaller jump from OC2 to OC3 than from OC1 to OC2. For just an additional 6% to memory alone (the increase from OC2, OC3) an additional 0.72% of performance is gained.

Unfortunately, I had to average the overclocks (core & memory) to establish any sort of single ratio relationship between the overclock and the corresponding improvement in 3dMark, so these aren't mathematically conclusive. To be conclusive, we'd need numbers that hold one of the two variables (core and memory) stationary while varying the other.

But these results make me think two things about R600.

1) I'm not sure that overclocking is really all that successful. Even an--at best-- ratio of 0.23% of improved performance for each 1% of overclocking doesn't seem particularly high, although I'll defer to those with a lot more knowledge of the GPU industry. This ratio also decreases pretty rapidly. While the first 9.95% of overclock (stock to OC1) yields 0.23% of improvement per 1% of overclock, the next 8.96% of overclock (OC1 to OC3) only yields 0.12% of improvement per 1% of overclock.

2) There seems to be a necessary balance between the core and memory clocks at which R600 works most efficiently. The jump from OC1 to OC2, which included both a core and memory jump, was less effective (as a ratio of oveclock to performance increase) than the smaller jump from OC2 to OC3.

I'm sure that better numbers will eventually be available. So far it seems a decent part but far from overwhelming.

 

[aka1nas]

Could it be slightly CPU-bottlenecked in 3dmark at those resolutions? Also, could be that there is something wrong with OverDrive in the Beta Cats he is using and he is not getting the clocks that it says he is. He didn't mention using any third-party apps to detect actual clocks AFAIK. Such apps may not even work with the X2900XT yet for that matter.

 

[yacoub]

Originally posted by: AnotherGuy

Originally posted by: yacoub
hahah wow, a couple retards who can't read responded to my comment in that thread.

I said I prefer a card that is clocked lower and runs cooler and quieter and also happens to outperform the higher-clocked card even when that card is overclocked.

Wrong thing to say, apparently. :roll:

So u prefer

no i was making a general statement.

 

[dreddfunk]

aka - I thought about CPU bottlenecking but they seem to be using a quad-core Core 2, so that seems unlikely.

As someone else suggested, it could be that the actual clocks and the reported clocks are different, although it would seem that they would have to be substantially different. That, any event, seems more likely of these two options.

I'm sure we'll see more benches before too long.

 

[otispunkmeyer]

Originally posted by: aka1nas
Could it be slightly CPU-bottlenecked in 3dmark at those resolutions? Also, could be that there is something wrong with OverDrive in the Beta Cats he is using and he is not getting the clocks that it says he is. He didn't mention using any third-party apps to detect actual clocks AFAIK. Such apps may not even work with the X2900XT yet for that matter.

unlikely, this is a core 2 quad, it should rip thru the cpu tests in 06 which do have an effect on the overall score but the actual graphics tests, SM2 and SM3 arent cpu bound and they affect the score more.

yeah i have no idea what he's using, i dont know who sven olsen is, but he sounds norweigen or scandanavian or whatever and they like their hard core overclocking so im sure he knows what hes doing.

 

[deadseasquirrel]

Originally posted by: dreddfunk
aka - I thought about CPU bottlenecking but they seem to be using a quad-core Core 2, so that seems unlikely.

I don't think he meant "cpu bottlenecking" as in the cpu holding back the GPU... but rather the test being run may be more stressful to the CPU vs the GPU. e.g. If you run some games at 1024x768 with the same CPU but compare a GTX to a GTS, you'll likely show them to be equal in a benchmark since more stress is being put on the CPU in that test.

Dunno if that's the issue or what, but I agree with everyone else-- we need more data.

 

[Matt2]

dreddfunk-

Good analysis of the overclocking numbers. I'm inclined to believe that the card was actually overclocked or else we wouldnt see any improvement on performance. I'm not sure I believe that the card was CPU bottlenecked.

Could be driver issues, could be speed straps, I dunno. Historically ATI has not used speed straps or clock domains, but since this is a brand new architecture, we'll have to wait for full reviews to find out.

 

[imported_thefonz]

This could also be a hand picked card, which can achieve these high clocks, I'd wait till we see some forums user reviews before you start assuming that 2900's will clock like this all the time.

 

[Matt2]

Originally posted by: thefonz
This could also be a hand picked card, which can achieve these high clocks, I'd wait till we see some forums user reviews before you start assuming that 2900's will clock like this all the time.

Good catch. Didnt even think of that.

As always YMMV, but this wouldnt be the first time cherry picked cards were sent to reviewers.

 

[dreddfunk]

Matt2 - what are 'speed straps' and 'clock domains'? I'm very unenlightened when it comes to the inner workings of GPUs. I tend to agree that the cards were actually probably overclocked. The numbers are just a little strange, though. Still, I've no idea what is average when it comes to ROI, so to speak, with overclocking.

Fonz - you bet, and good call.

Dead - that makes more sense--as a couple of folks have mentioned, benching these cards at 12x10 and 16x12 isn't really going to that revealing.

 

[CrystalBay]

Yeah Chizow and DreadF that card doesn't look to scale well, WTF . Maybe it needs more juice , I dunno... or more driver work.

The thing is we don't really know the the full story on the chip architecture...

 

[BladeVenom]

Originally posted by: yacoub
I said I prefer a card that is clocked lower and runs cooler and quieter and also happens to outperform the higher-clocked card even when that card is overclocked.

Point to one review that says the 2900XT is louder and hotter. You keep inferring that it is, but have shown no evidence.

 

[aka1nas]

At this point, I don't think we actually have power consumption data on the XT. The cooler looks to be a rip-off of the 8800 series cooler, so I don't know why you are complaining yacoub.