5870 Bottleneck Investigation (CPU and/or Memory Bandwidth)

[Astennu]

Furmark has a strange way of stressing the cards to there max. You can overheat a HD4870 or HD4890 at ease even without renaming the .exe file A twin turbo could not cool a HD4890 VRM's got to hot. Stock cooler does a better job there bot you do have a lot more noise. The power draw is also way higher. So its deffinatly doing something. I also think furmark is able to keep all the 800 (RV770 and RV790) and 1600 shaders (RV870) fully loaded where games can not.

[saaya]

i dont think its that all the shader cores are loaded, games can def do that, what furmark does is load the cores completely, all 5 parts of it, and occt even feeds the rest of the logic with work and keeps texture and geometry setup busy with bogus work i think... its kinda like linx that bombards cpus with random instructions without any dependencies so every single unit is fully loaded. i think...

[Astennu]

i dont think its that all the shader cores are loaded, games can def do that, what furmark does is load the cores completely, all 5 parts of it, and occt even feeds the rest of the logic with work and keeps texture and geometry setup busy with bogus work i think... its kinda like linx that bombards cpus with random instructions without any dependencies so every single unit is fully loaded. i think...

Indeed thats what i tried to say :P Games mostly load 2-4 out of 5 alu's. And sometimes they get up to 5.

[largon]

Furmark 1.7
1280x1024 8x AA 60000MS
HD 5870 1GB

Mem - Score
1300 - 2733
1200 - 2537
1100 - 2360
1000 - 2187
900 - 1963

http://img194.imageshack.us/img194/9013/graphtv.jpg

My HD4890 at SXGA, 8xMSAA scored 1694 at core 850MHz, memory 900MHz and 1859 at memory 1000MHz. So HD5870 at same clocks is whopping 16-17% faster. And that, ladies and gentlemen, is nothing short of messed up considering Cypress' overwhelming HW level superiority and the fact Furmark is very, very shader dependant...

[Miwo]

Any conspiracy theories yet?

Witholding performance via driver for GT300 launch? lol...

[gamervivek]

My HD4890 at SXGA, 8xMSAA scored 1694 at core 850MHz, memory 900MHz and 1859 at memory 1000MHz. So HD5870 at same clocks is whopping 16-17% faster. And that, ladies and gentlemen, is nothing short of messed up considering Cypress' overwhelming HW level superiority and the fact Furmark is very, very shader dependant...

doesn't furmark have to be updated for better loading of newer cards?

[wez]

Any conspiracy theories yet?

Witholding performance via driver for GT300 launch? lol...

Well we haven't seen what the x2 can do yet. It's always possible that they've figured out an more efficient way for the GPU's to work over the traditional cfx..

wild speculations

[largon]

gamervivek,
Of course not. It's just an OpenGL benchmark app. It's not like games and all 3D aps need to be patched to "support" new HW.

[Chickenfeed]

My HD4890 at SXGA, 8xMSAA scored 1694 at core 850MHz, memory 900MHz and 1859 at memory 1000MHz. So HD5870 at same clocks is whopping 16-17% faster. And that, ladies and gentlemen, is nothing short of messed up considering Cypress' overwhelming HW level superiority and the fact Furmark is very, very shader dependant...

I will be doing this test at more "real" world resolutions soon ( real world as far as these cards go - I know 1280x1024 is still widely used though ) I do agree that something is pretty messed up given the small difference.

I'll mention I've found that I get a display driver has stopped running error at 900/1300 when running OCCTs furmark type test after 30-45min ( this got core to like 88C and fan to over 50% which gets quite loud ) With actual games though it usually is in the 70s and only Crysis has managed to get any higher. I'm not sure if this is an indication of the card not being able to handle the load at these clocks in something like Furmark or merely drivers in their infancy. All I know is no games or other programs have shown signs of issue yet so far.

[Kuntz]

rv600 and rv670 and rv770 all adjusted the timings when you overclocked the memory... id be surprised if rv870 doesnt do it anymore...

a drop from 8.3 to 7.7ns is a latency reduction of 7% though, then how come we still only get a 1.5% boost? and memory timings tend to not really matter on vgas, they need bw bw bw... they usually have cas15 or so...

i dont think the boost we see is only from reduced memory latency, as a matter of fact, i dont think memory latency decreases a lot, since r600 all gpus use a formula to calculate memory timings based on memory clock. the latency still decreases, and bw still increases, otherwise we wouldnt see any gains at all, but its less than we would see from a static timing config with increased memory speed.

you are probably thinking of system memory performing better with lower timings, for vgas that doesnt apply... they do score better with lower timings by the boost is tiny compared to cpu system memory...

nice tests chumbucket and chickenfield (sp?)
what it shows is that in furmark the 5870 def is bw limited, but that doesnt seem to be the case in actual games...
furmark is 99% shader heavy load only isnt it?

Ah I never knew that about the memory timings on ATI's cards.

I still don't think we have a memory bandwidth issue in games.

If a core increase of 10% nets 5% on average, and a memory bandwidth increase only nets 0.5% to 3.5%, I'd say we have a core bottleneck more than anything. :p

Though this benchmark stands out for oddity:



-A 10% increase in memory nets ... 0% increase in frame rate.
-A 10% increase in core nets ... 0.6% increase in frame rate.
-Both increases combined and now we have a 7.4% increase in frame rate.

[Chumbucket843]

i dont think its that all the shader cores are loaded, games can def do that, what furmark does is load the cores completely, all 5 parts of it, and occt even feeds the rest of the logic with work and keeps texture and geometry setup busy with bogus work i think... its kinda like linx that bombards cpus with random instructions without any dependencies so every single unit is fully loaded. i think...

its more complicated than that. if a shader program is flow heavy it will do a lot better on nvidia cards. loading all of the alu's is harder than it sounds. its vliw too so not all instructions can take full advantage of all of the shaders. they encourage programmers to use certain floats and vecs to get the fastest performance. memory size is probably the issue. think about it. if you double the amount of threads you double the required memory. its like having a 4870 512mb. it makes sense that seeing that they doubled everything but memory and bandwidth.

[Crankybugga]

Any conspiracy theories yet?

Witholding performance via driver for GT300 launch? lol...

Im still waiting for this "investigation" to be finalised so we can see how ATi have completely stuffed the HD5870

[Soultaker52]

Any conspiracy theories yet?

Witholding performance via driver for GT300 launch? lol...

I know that's what's on my mind. I tried asking on Terry Makedon's (CatalystMaker) twitter, but he only evaded the question, then stopped answering when I pushed harder. I'm really beginning to think they're holding the card back.

[JimmyH]

One theory why 5870 is slower than its spec suggest: Its 1600 shaders are nowhere near working at their max capacity. This would explain why

1) increasing memory clocks does not improve results as much as expected.
2) load power is lower than other cards in games but much higher in occt.
3) raising memory has a greater effect in furmark

[flopper]

"if" they can hold it back a little, it make sense as they wait for the late and very late g300 performances.

[iandh]

You know, ATI holding the card back really goes pretty far into tinfoil hat territory, but it almost makes sense...

[Mr. K6]

I think it might be accidentally held back due to drivers. If I understand it correctly, a lot of the performance from ATI's design comes from the GPU being fed information optimally. I'd wait for a few driver releases to see final performance numbers.

[zalbard]

Yep, need better drivers, hard to expect some crappy release candidate version to offer 100% performance.

[Astennu]

I think its driver related but i dont think AMD does this in purpose. Because the card now looks bad vs the HD4870 X2. The X2 is out gunning it. They did change a lot in the core. So i expect they will be able to boost the performance by 10-20% in the future. We have also seen massive gains with newer drivers for the HD48xx cards.

And about furmark. I know AMD does something in there drivers to cut down the load on the GFX card. Furmark puts a rare load on the card causing some to overheat. Because of that AMD has put a limiter in the driver. You can disable it by renaming the .exe in something else. You might wanna try that if you want a fair compare. But keep an eye out for you temperatures. They can get higher then before !

[JimmyH]

So despite what they claimed the new AF does take more serious hit than before:

http://pclab.pl/art38674-9.html

Might be a better idea to run these cards at 8xAF instead of 16x

[fornowagain]

rv600 and rv670 and rv770 all adjusted the timings when you overclocked the memory... id be surprised if rv870 doesnt do it anymore...

That's very interesting, I remember reading that before somewhere. Don't suppose you have a link?

[PaganII]

Maybe a bottleneck in crossfire.

http://www.hardwarezone.com/articles...?cid=3&id=3032

[Hornet331]

... its kinda like linx that bombards cpus with random instructions without any dependencies so every single unit is fully loaded. i think...

Sry but linx is nothing more then a gui for linpack...

[demonkevy666]

http://www.xtremesystems.org/forums/...=235831&page=2

look at the score on the pages before oc on card and after.
seems more like a bios problem, note this was an AMD system though

[Metatron]

One theory why 5870 is slower than its spec suggest: Its 1600 shaders are nowhere near working at their max capacity. This would explain why

1) increasing memory clocks does not improve results as much as expected.
2) load power is lower than other cards in games but much higher in occt.
3) raising memory has a greater effect in furmark

If you want to verify the capacities of the ALUs you should consider testing the same stuff on the "Froblin"-demo with a 4870 and a 5870, because you got adaptive tessalation even without DX11.
That somewhat disconnects the internal from the external bandwidth requirements, as no (proportionally) more geometry is streamed into the processor.
I would also make the very vague guess that crunching on up-LODed geometry creates a somewhat shaderinstruction-grouping favoring situation. But that's very hard to say. That can really only be verified if you can play around with the LOD-strength in the demo.

You'd check something like this combinations:

LOD: 1, 2, 4
Chips: 4870, 5870
Core-Clock: x, y, z

BTW the architecture is well described (ATI published ta 392 page doc to the linux guys), so someone eager may find the reason in there.

Good luck