Firingsquad 5870 Overclocking article

[AzN]

http://firingsquad.com/hardwar...verclocking/page12.asp

Based on the benchmarks we just saw, it?s pretty safe to say that when OC?ing the Radeon 5870 you?ll get the best gains from GPU rather than memory OC?ing. We setup both components to be OC?ed by the same 9% ratio, and in most cases the GPU OC scenario came out on top in performance, generally by about 5%. Only in Batman: Arkham Asylum did we finally see memory overclocking deliver greater gains than GPU OC?ing.

This can probably be attributed to the fact that we?re enabling 4xAA via control panel in the case of this game, which forces 4xAA on everything in the scene. This brute force method pushes the memory subsystem harder than using in-game AA, which selectively applies AA in the scene. This is reflected in the Batman OC?ing benchmarks. Here OC?ing the memory 9% bought us an additional 3-4% in performance, which is better than the 2-3% we saw in the other games we used for testing.

These results should finally put any lingering criticism of the Radeon 5870?s 256-bit memory interface to bed. Clearly this card isn?t being held back by its 256-bit memory subsystem with today?s latest games.

We would like to see ATI open up more clock speeds inside Catalyst Control Center though. Early indications today suggest this card can scale much further than ATI currently allows. It doesn?t appear to scale like the Radeon 4890 did, but the 5870 certainly looks like it?s a better OC?er than the Radeon 4870 was.

When determining Overdrive speeds ATI obviously has to weigh the wants of the enthusiast crowd with the needs of the mainstream user who may not know a lot about OC?ing. We realize this. That?s why we think a good compromise would be to provide higher speeds once a config file or registry key is modified by hand by the end user. That keeps the enthusiast and experienced OC?ers happy, while the average joe can?t accidentally OC his card to the point where stability becomes compromised, which could lead to a pesky tech support call.

We?ve got more RV870-related topics to explore, including Radeon 5850 performance and supersample AA benchmarks. Be on the lookout for those articles in the days and weeks ahead?

The best gains would be from GPU and memory overclocking. Not just GPU overclocking alone.

The fact that you are also raising shader clocks along with core clocks when you raise core on 5870 nets you bigger gains over memory but the memory clocks aren't far behind at all. Kind of wacky conclusion considering the writer did not point this out.

It definitely is being held back but not enough that it's being hampered like G92 did. 4870 in other hand had all the bandwidth it needed and then some. If it had more bandwidth it would be more faster. The fact that 16ROP 40TMU 800SP 4890 is only 25% slower than 5850 with 32ROP and 72TMU 1440SP tells you it's definitely being limited to a degree.

 

[cusideabelincoln]

The 4890 has higher clock speeds on those 800SPs, though, which probably accounts for a lot of the difference (see the 4770 vs. 4850).

But yeah, honestly, I expected some bigger gains from only overclocking the memory. Glad to see in this early testing it isn't severely chocked, but I think this issue could use more attention especially as newer games and drivers come out.

edit: Yep, just remembered AT's 5870 launch article and I do think memory overclocking needs more attention. http://www.anandtech.com/video/showdoc.aspx?i=3643&p=12

 

[MarcVenice]

Originally posted by: Azn
http://firingsquad.com/hardwar...verclocking/page12.asp

Based on the benchmarks we just saw, it?s pretty safe to say that when OC?ing the Radeon 5870 you?ll get the best gains from GPU rather than memory OC?ing. We setup both components to be OC?ed by the same 9% ratio, and in most cases the GPU OC scenario came out on top in performance, generally by about 5%. Only in Batman: Arkham Asylum did we finally see memory overclocking deliver greater gains than GPU OC?ing.

This can probably be attributed to the fact that we?re enabling 4xAA via control panel in the case of this game, which forces 4xAA on everything in the scene. This brute force method pushes the memory subsystem harder than using in-game AA, which selectively applies AA in the scene. This is reflected in the Batman OC?ing benchmarks. Here OC?ing the memory 9% bought us an additional 3-4% in performance, which is better than the 2-3% we saw in the other games we used for testing.

These results should finally put any lingering criticism of the Radeon 5870?s 256-bit memory interface to bed. Clearly this card isn?t being held back by its 256-bit memory subsystem with today?s latest games.

We would like to see ATI open up more clock speeds inside Catalyst Control Center though. Early indications today suggest this card can scale much further than ATI currently allows. It doesn?t appear to scale like the Radeon 4890 did, but the 5870 certainly looks like it?s a better OC?er than the Radeon 4870 was.

When determining Overdrive speeds ATI obviously has to weigh the wants of the enthusiast crowd with the needs of the mainstream user who may not know a lot about OC?ing. We realize this. That?s why we think a good compromise would be to provide higher speeds once a config file or registry key is modified by hand by the end user. That keeps the enthusiast and experienced OC?ers happy, while the average joe can?t accidentally OC his card to the point where stability becomes compromised, which could lead to a pesky tech support call.

We?ve got more RV870-related topics to explore, including Radeon 5850 performance and supersample AA benchmarks. Be on the lookout for those articles in the days and weeks ahead?

The best gains would be from GPU and memory overclocking. Not just GPU overclocking alone.

The fact that you are also raising shader clocks along with core clocks when you raise core on 5870 nets you bigger gains over memory but the memory clocks aren't far behind at all. Kind of wacky conclusion considering the writer did not point this out.

It definitely is being held back but not enough that it's being hampered like G92 did. 4870 in other hand had all the bandwidth it needed and then some. If it had more bandwidth it would be more faster. The fact that 16ROP 40TMU 800SP 4890 is only 25% slower than 5850 with 32ROP and 72TMU 1440SP tells you it's definitely being limited to a degree.

Wait, the article concludes that oc-ing the ram 9% nets 2-3% performance gains in most games. Yet you remain convinced the HD 5870 is memory bandwith starved? I've had the same feeling, but it's clearly not the case, otherwise you'd see a near linear increase when oc-ing the ram.

 

[BFG10K]

Originally posted by: Azn

The best gains would be from GPU and memory overclocking. Not just GPU overclocking alone.

Uh, that applies to pretty much every video card ever made.

The fact that you are also raising shader clocks along with core clocks when you raise core on 5870 nets you bigger gains over memory but the memory clocks aren't far behind at all. Kind of wacky conclusion considering the writer did not point this out.

The conclusion is fine given it?s backed by their results. I?m not sure how you derived your conclusion of ?memory clocks not being far behind? given a 5% performance improvement to the core compared to 2% memory means that on average the core has 2.5 times the performance impact as memory.

Also had they gone clocked over 9% the absolute performance difference would?ve grown even more.

4870 in other hand had all the bandwidth it needed and then some.

If you overclocked the 4870 in the same way you would likely see similar results.

If it had more bandwidth it would be more faster. The fact that 16ROP 40TMU 800SP 4890 is only 25% slower than 5850 with 32ROP and 72TMU 1440SP tells you it's definitely being limited to a degree.

Again, almost every video card is bandwidth limited to some degree and would be faster with more bandwidth. This is nothing new or surprising.

 

[Forumpanda]

It doesn't have to prove that it isn't bandwidth starved, since the card has the new error checking feature it could simple be that the memory is not performing 9% better.

Also memory would be more important as the GPU clock speed is increased, either wit ha HD5980 varient or just from OCing.

And as BFG pointed out, OCing *both* GPU and memory by 9% would (should) give a 9% increase in performance, so when OCing one or the other they are sharing that 9% increase.

So some bad math would tell us that its memory bottlenecked at least 25% of the time.

 

[AzN]

Originally posted by: BFG10K

Originally posted by: Azn

The best gains would be from GPU and memory overclocking. Not just GPU overclocking alone.

Uh, that applies to pretty much every video card ever made.

True. But the best performance derived from gpu and memory clocking combined. Perhaps I took it as literally.

The conclusion is fine given it?s backed by their results. I?m not sure how you derived your conclusion of ?memory clocks not being far behind? given a 5% performance improvement to the core compared to 2% memory means that on average the core has 2.5 times the performance impact as memory.

Also had they gone clocked over 9% the absolute performance difference would?ve grown even more.

I didn't particularly liked how the writer didn't point SP clocks were being raised when core clocks were being raised since Nvidia cards are capable of this. Now if you could raise SP and core clocks separately what do you think will happen? That gap would quickly shrink and memory would be equal footing as core clocks considering bandwidth usually only hinders fillrate.

If you overclocked the 4870 in the same way you would likely see similar results.

Not quite. You would get absolute minimal gains from memory when you overclock 4870 the same way showing that 4870 is much more bandwidth saturated compared to 5870. But I see you were just generalizing.

Again, almost every video card is bandwidth limited to some degree and would be faster with more bandwidth. This is nothing new or surprising.

Degree sure but the fact that 5850 with 70% more pixel fill, 53% texture fill, 53% more processing power only nets you 25% over 4890 tells you that it's more bandwidth limited.

 

[AzN]

Originally posted by: MarcVenice
Wait, the article concludes that oc-ing the ram 9% nets 2-3% performance gains in most games. Yet you remain convinced the HD 5870 is memory bandwith starved? I've had the same feeling, but it's clearly not the case, otherwise you'd see a near linear increase when oc-ing the ram.

Memory bandwidth raises minimal frame rates and helps AA performance while fillrate helps peak frame rates. There's an equilibrium between fillrate and bandwidth. Bandwidth doesn't quite help much on average frame rates than minimum frame rate however core clock help peak more max frame rates to get much better average frame rate. To say 5850 or 5870 isn't being bottlenecked is like saying g92 wasn't bottlenecked by bandwidth. All cards are limited to a degree except maybe 2900xt that was fillrate hungry but how much this limitation is the real question.

The added Shader overclocking along with core clocks added performance and gained you extra 4-5% performance compared to 2-3% memory overclocking. Now take out that shader clocking overhead and memory clocking would get equal footing as core clocks. All while GTX285 would show more of a saturated result. Raising only core clocks and not shader would net you much bigger gains over memory or shader clocking in average frame rates. Perhaps 2x fold as your result of 5870.

 

[BFG10K]

Originally posted by: Azn

True. But the best performance derived from gpu and memory clocking combined.

Yes I agree; my point is that this isn?t new or anything surprising. Even a 2900XT had situations where this applied.

Now if you could raise SP and core clocks separately what do you think will happen? That gap would quickly shrink and memory would be equal footing as core clocks considering bandwidth usually only hinders fillrate.

The gap would shrink but I don?t think we could infer which clock would have the highest impact, except to say it isn?t the memory.

With nVidia?s recent parts it?s generally been the core but I?m convinced it?s because their shaders run at twice the core. With ATi the shader clocks are the same as the core, and I?m convinced they would show a bigger difference than with nVidia?s parts.

Anyway, it doesn?t change the outcome since a memory bottleneck would show the memory being the primary limitation, but that isn?t happening here. Raising the core increases performance by a ratio of 2.5 to 1 compared to raising the memory.

Not quite. You would get absolute minimal gains from memory when you overclock 4870 the same way showing that 4870 is much more bandwidth saturated compared to 5870. But I see you were just generalizing.

They were minimal on the 5870 too: 2% for every 9%. I?d expect something similar on the 4870.

Degree sure but the fact that 5850 with 70% more pixel fill, 53% texture fill, 53% more processing power only nets you 25% over 4890 tells you that it's more bandwidth limited.

Yes, or the drivers need work, or possibly some of the tests have elements of CPU/platform limitation.

It?s worth noting that the 5870 is about 25%-30% faster than the 5850 and that?s almost exactly in line with its 30% shader performance advantage:

http://www.computerbase.de/art...rmancerating_qualitaet