HD 5870 Discussion thread.

[Bo_Fox]

Yeah, that's usually the percentage gains you get with overclocking the memory. It's not such a linear bottleneck in that you get the same percentage gain with the MHz gains, but sometimes when a bottleneck threshold level is overcome in a certain scenario that is so bandwidth-hungry, there will be a huge boost in performance after overclocking it past a certain level. It only happens in critical scenarios, however.

The bottom line is that increasing the bandwidth still does give performance gains. If increasing the clock by 18% gives a 3% boost, then increasing it by 100% could as well give an 18% boost, but that figure is rather low according to the difference between a 5770 and a 4890 despite slight architecture optimizations. A 4890 has 62% greater bandwidth, but performs "only" 20% better.

 

[a_ump]

true but what else is different about RV870? i personally do not know all the differences and i agree that in a bandwidth hungry app the overclock would attribute a greater increase in performance, but i believe that the bottleneck lies elseware, be it something in the architecture or drivers, or dx10 games simple not being able to make full use of the 1600 shaders but a dx11 game would. idk lol

 

[mav2000]

Guys I need some help here. Got my 5870 last evening and even though the 3dm06 scores seem to be in line with others at 1 Ghz core and 1.25 mem, the 3dmV score is around 12,000 for the GPU. This is not correct as it should be something close to 17,000 by my calculation. My system specs are in my sig and Im running it at 4 Ghz and getting an absolute crap 3dmv score...any help guys??? I reinstalled 3dm Vantage again, as there was a gtx 295 before this, but that didnt seem to help it.

Edit: even Heaven benchmark is at half the score....wonder whats happening, only 3dmark 06 seems to be working good.

 

[Benetanegia]

Guys I need some help here. Got my 5870 last evening and even though the 3dm06 scores seem to be in line with others at 1 Ghz core and 1.25 mem, the 3dmV score is around 12,000 for the GPU. This is not correct as it should be something close to 17,000 by my calculation. My system specs are in my sig and Im running it at 4 Ghz and getting an absolute crap 3dmv score...any help guys??? I reinstalled 3dm Vantage again, as there was a gtx 295 before this, but that didnt seem to help it.

Edit: even Heaven benchmark is at half the score....wonder whats happening, only 3dmark 06 seems to be working good.

Have you benchmarked in every step up to 1000 Mhz? Remember that HD5xxx series have error detection, so you are not going to see artifacts even if the card is failing due to overheating/instability. Only way of knowing that is when performance drops instead of increasing as you go higher in clocks. 1 Ghz does seem high according to what most people are reaching AFAIK (on stock cooling and Vcore that is). 3Dmark 06 is far from maxing out current generation cards (not even last generation cards) so it is possible that your card is fine when running 06, but is failing in more taxing benchmarks.

 

[mankind]

The error correction is ONLY on the memory, the core will still artifact or lock the machine if to high.

 

[Benetanegia]

The error correction is ONLY on the memory, the core will still artifact or lock the machine if to high.

And cache and registers AFAIK.

EDIT: Ooops my bad, it's Fermi that has that ability on cache and registers.

 

[mav2000]

Ok...so heres a 3dmark vantge shot with the gpu running 850/1200 and the cpu at 3.6....crappy scores..HELP..........I tried with the cpu at 4 ghz and it was the same. I made sure that it was running 850/1200 by looking at the sensrs screen on gpuz...so it is running as per spec only the scores are mighty low.....please HELP....

 

[wolf]

mav2000, is anything forced in CCC at all? we have a member who forgot he had 16xAF forced in CCC, make sure everything is set to application choice of settings.

 

[HalfAHertz]

Guys I need some help here. Got my 5870 last evening and even though the 3dm06 scores seem to be in line with others at 1 Ghz core and 1.25 mem, the 3dmV score is around 12,000 for the GPU. This is not correct as it should be something close to 17,000 by my calculation. My system specs are in my sig and Im running it at 4 Ghz and getting an absolute crap 3dmv score...any help guys??? I reinstalled 3dm Vantage again, as there was a gtx 295 before this, but that didnt seem to help it.

Edit: even Heaven benchmark is at half the score....wonder whats happening, only 3dmark 06 seems to be working good.

You said you had a GTX295 in the system before that? Are you sure you re-installed the drivers properly?

 

[mav2000]

Well I did a reinstall of windows yesterday and the results have not changed. So it can't be a driver issue. Wonder why its acting up

 

[phanbuey]

Well I did a reinstall of windows yesterday and the results have not changed. So it can't be a driver issue. Wonder why its acting up

hrrmrmrm... could be a bios bug... have you tried testing your rig at stock and just the card OC'd?

 

[Jstn7477]

Ok...so heres a 3dmark vantge shot with the gpu running 850/1200 and the cpu at 3.6....crappy scores..HELP..........I tried with the cpu at 4 ghz and it was the same. I made sure that it was running 850/1200 by looking at the sensrs screen on gpuz...so it is running as per spec only the scores are mighty low.....please HELP....

http://img682.imageshack.us/img682/1319/62473864.jpg

It doesn't look like all the tests were selected, only the first two. You only have the results for the Game tests, there aren't any results for the CPU, hence a score of 0.

 

[mav2000]

Well my friend the idea was to test only the gpu and see the score. The gpu score will be from the first two tests only and not from the other two. I am talking about the graphics score which should e around 15-16k at stock for this card.

 

[HalfAHertz]

Could we get a GPU-Z shot?

 

[lemonadesoda]

i do believe the ati engineers dont hang out on this forum, or they would have read THIS THREAD and we wouldnt be having these performance questions again.

 

[Zubasa]

i do believe the ati engineers dont hang out on this forum, or they would have read THIS THREAD and we wouldnt be having these performance questions again.

The only "problem" with those tests, is the fact that it is done using the 4xAA "Box" filter.
The problem with the R(V)6XX is they don' have enough shaders, ad as a result that generation performed quite poorly in FSAA.
On the other hand the RV7XX have much more shaders per ROP which makes them performs very well in FSAA.

What we need to know is, what happens if we use the "Edge Detect" filter with 4xAA+?
Since the "Edge Detect" filter mainly uses the cards shaders for the effect.
Although, the 4830 will have a hard time maintaining high FPS under 4xAA on new games, so it probably doesn't matter.

 

[HalfAHertz]

The only "problem" with those tests, is the fact that it is done using the 4xAA "Box" filter.
The problem with the R(V)6XX is they don' have enough shaders, ad as a result that generation performed quite poorly in FSAA.
On the other hand the RV7XX have much more shaders per ROP which makes them performs very well in FSAA.

What we need to know is, what happens if we use the "Edge Detect" filter with 4xAA+?
Since the "Edge Detect" filter mainly uses the cards shaders for the effect.
Although, the 4830 will have a hard time maintaining high FPS under 4xAA on new games, so it probably doesn't matter.

It also shows that speeding up the shaders (compare the 4850 to the 4870 - 625mHz to 750mHz) the fps per shader value increases. I'd like to see a 4890 OCed to 1GHz in that comparison... or it could be a memory bandwith limitation I hope Bo Fox doesn't read this

 

[grimeleven]

i do believe the ati engineers dont hang out on this forum, or they would have read THIS THREAD and we wouldnt be having these performance questions again.

Yeah ATI has been way behind in that area, the 5800s still has same ROPs as Nvidia's previous gen -1 year (GT200). They started 16 ROPs way back with the Radeon X800 series in 2004.... 32 ROPs 5 years later lol... says it all.

X800 XT PE= 8320 MPixels/s
+4 years
4870 = 12000 MPixels/s

Go figure

WTB 64 ROPs for a nice 54400 MPixels/s
*Edit* Maybe its just their scalar design preference, a 5970 will have 46400 MPixels/s, fair enough

 

[Bo_Fox]

i do believe the ati engineers dont hang out on this forum, or they would have read THIS THREAD and we wouldnt be having these performance questions again.

Ahh, never seen this thread about the ROP's.

Wow, I'm surprised how many people went along with this theory about ROP's just nicely, but when I try to say that it's bottlenecked by the memory bandwidth, everybody wants to go against me.

Must be bad timing or some kind of a curse? Well, I'm just in a state of self-pity right now. *sighs*

At least I didnt start this thread about the 5870. And I'm not the only one who believes that a 5870 needs 512-bit bandwidth first before anything else. Over at Anand forums, etc.. it's what a 5770 needed before anything else (2x the bandwidth), since slight architecture optimizations still did not let come within 20% of a 4890.

 

[Binge]

Now's not the time to hate the community, just feel sour that you didn't include ROPs in your list of problems with the card

 

[Bo_Fox]

LOL, yeah, I was one of the people who voiced out against the low number of ROP's when the 4870 came out.

But now that it's been doubled to 32, it was nice for a change (the first time ATI increased it after 4 generations, since the X800XT). So I was happy.

I was also one of those who complained against a 4870 having only 512MB of RAM (at least for a few months until there was a 1GB version). It was unbelievable how many people said that 1GB was a waste, but they were all silenced after the 1GB version was finally released.

I guess we'll just have to wait and see if ATI releases a new 512-bit card using similar specifications (same # of shaders, TMU's, ROP's, etc..). The engineers know best anyways, so let's just see..

Thanks Binge for theheads-up about not giving up on the community! I know there are a few here who do support the theory of a 5870 being in dire need of 512-bit bus. Those who wish to argue otherwise, I do not know what they are getting at, but I'll decline to respond to arguments against this theory for the rest of the thread. I'm the only one who provided more than 20 benchmark charts in this thread and it's being refuted by 1-2 games tested with a 5870's memory being overclocked (probably with "undue" changes in latency that has not yet been tested for, or with the error-correcting algorithm getting in the way with overclocked memory).

That's all for now.. at least I feel better not being in a constant argument with somebody.

 

[SNiiPE_DoGG]

I think we will see what drivers can do and DX11 games.... Honestly we do not know the real performance of the 58XX series until we see the comparison between the GTX295 and a 5870 in a true DX11 game; utilizing all dx11 features.

right now we are at a crossroads where the old generation of graphics engine has not me the capabilities of the new generation hardware and that is sorely skewing the ruslts of benchmark comparisons.

 

[mav2000]

Hi ...solved the problem with beta 9.11 drivers. I wonder what the issue was with 9.10, 9.9 etc.....and i did a whole reinstall...anyway that was pending for a while.

 

[grimeleven]

Ahh, never seen this thread about the ROP's.

Wow, I'm surprised how many people went along with this theory about ROP's just nicely, but when I try to say that it's bottlenecked by the memory bandwidth, everybody wants to go against me.

Must be bad timing or some kind of a curse? Well, I'm just in a state of self-pity right now. *sighs*

At least I didnt start this thread about the 5870. And I'm not the only one who believes that a 5870 needs 512-bit bandwidth first before anything else. Over at Anand forums, etc.. it's what a 5770 needed before anything else (2x the bandwidth), since slight architecture optimizations still did not let come within 20% of a 4890.

It's not a "theory" the numbers speak for themselves that it would increase performance a lot (GF100 is going for 44800 MPixels/s), among other things those 5800s are just castrated for costs reason. Everyone knew they were low with this, it's just pointing out other bottlenecks that were chosen for costs reasons again.

 

[grimeleven]

Hey guys i found this link to a really nice document explaining stuff related to the thread and Hydra and then some, good read http://jonpeddie.com/publications/whitepapers/multi-gpu/ download the zip, PDF inside.

Page 13 on possible bottlenecks

The pipeline doesn’t tell the whole story. The load on the CPU or GPU can vary widely from
game to game, and even from one scene to the next in the same game. For example, F.E.A.R.
spends a huge amount of time in rasterization, depth tests, and alpha blending. By comparison,
each frame of Gears of War is extraordinarily heavy on pixel shaders, but comparatively low on
depth tests and rasterization.

That quote explains pretty much how a game such as Crysis can run better on Nvidia than competition, it has heavy use of pixel shaders etc.. not so much about the TWIWMTBP crap we hear all over.