5850 just as fast as a 5870?

[AzN]

I'm pretty sure he means that the driver improvements weren't limited to a 4870, but applied across the 4xxx series line. In such a case, this is hardly an argument against the assertion made in the OP, as the 5850's performance will benefit just as well as the 5870's with further driver improvements.

Which has nothing to do with the topic why the 5850 is only 2% slower than 5870 when it's clocked the same.

 

[dguy6789]

With all video cards bandwidth plays the equalizer to the core. Of course bandwidth is the limitation of the 5870 considering all the core doesn't have the breathing room it used to have with rv770. If it didn't it would be faster than 4870x2 by 15% or more in all cases but it's not.

This is exactly what everyone thought when the cards came out. Some sites did testing and proved this theory wrong. There is no use trying to defend it. If your card is memory bandwidth limited, it should improve in performance almost linearly with clock speed improvements. In the case of the 5870, overclocking the memory by a ton doesn't give you much performance improvement. The improvement is nothing of the sort that a memory bandwidth starved card should be receiving.

The link above by OP only strengthen that concept as 5850 with 5870 clocks is only 2% slower than 5870.

What you are proposing is that 5850 have better drivers than 5870. We all know this isn't true.

We all know this isn't true because why? For whatever reason, the 5870 is 2% faster than the 5850 clock for clock. The one thing we DO know for a fact is that it is not because of a memory bandwidth limitation, yet you keep sticking to that. You insist it can't possibly be because of drivers, something we have yet to prove is impossible. History has shown that drivers have improved performance plenty in the past.

I suggest you try and be a little more open minded rather than picking an idea and sticking to it even when all testing done shows it to be an incorrect hypothesis.

 

[AzN]

This is exactly what everyone thought when the cards came out. Some sites did testing and proved this theory wrong. There is no use trying to defend it. If your card is memory bandwidth limited, it should improve in performance almost linearly with clock speed improvements. In the case of the 5870, overclocking the memory by a ton doesn't give you much performance improvement. The improvement is nothing of the sort that a memory bandwidth starved card should be receiving.

No use in defending? lol.. really what sites? I question these sites and the authors themselves.

You can overclock the memory so much on 5870. Then again we have 4870x2 with same amount of cores and lower core frequency that is getting much better results in many games than 5870 with more bandwidth.

We all know this isn't true because why? For whatever reason, the 5870 is 2% faster than the 5850 clock for clock. The one thing we DO know for a fact is that it is not because of a memory bandwidth limitation, yet you keep sticking to that. You insist it can't possibly be because of drivers, something we have yet to prove is impossible. History has shown that drivers have improved performance plenty in the past.

So 5850 have better drivers than 5870. That's kind of amusing to say the least.

Driver improvements come within the same architecture not just 5870 over 5850.

I suggest you try and be a little more open minded rather than picking an idea and sticking to it even when all testing done shows it to be an incorrect hypothesis.

Was this even necessary? I can say the same thing about yourself. All the hypothesis is shown by 4870x2>5870.

 

[edplayer]

No use in defending? lol.. really what sites? I question these sites and the authors themselves.

Have you tested it yourself?

Have you seen any website test memory bandwidth improvements on these cards?

 

[AzN]

Have you tested it yourself?

Have you seen any website test memory bandwidth improvements on these cards?

You obviously know the answer to your 2nd question as I already answered to what you are quoting to. I gave you an example on your first question.

Seems to me you just want to pick a bone. Another question perhaps?

 

[T2k]

Which has nothing to do with the topic why the 5850 is only 2% slower than 5870 when it's clocked the same.

WTF are you talking about, seriously? Do you even read your own posts?

You said "I don't know why BFG thinks it's drivers when Radeon drivers are mature. There's no evidence of magic drivers out of hat."

Then I showed that it is quite the contrary: massive performance gains in new Catalysts happen pretty frequently, even a year later (from launch.)

Why would any sane person claim "Radeon drivers are mature" regarding a brand new, month-old architecture is beyond me... you either don't even read what are you posting or you have no clue whatsoever how drivers work...

 

[T2k]

I'm pretty sure he means that the driver improvements weren't limited to a 4870, but applied across the 4xxx series line. In such a case, this is hardly an argument against the assertion made in the OP, as the 5850's performance will benefit just as well as the 5870's with further driver improvements.

See above. I simply showed that his "Radeon drivers are mature" is simply ridiculous when it co9mes to 58xx-series.

 

[GodisanAtheist]

Which has nothing to do with the topic why the 5850 is only 2% slower than 5870 when it's clocked the same.

-Yeah... Ok... but it has everything to do with the subthread within the topic that suggested driver improvements would just make the 5870 suddenly unlock hidden performance above and beyond the 5850.

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.

-http://www.firingsquad.com/hardware/ati_radeon_5870_overclocking/

Firing Squad did a nice little OC review of the 5870 and had the decency to break down their graphs into Stock/Mem OC/Core OC/Total OC, making it very easy to demonstrate that the 5870 is NOT bandwidth limited.

 

[AzN]

WTF are you talking about, seriously? Do you even read your own posts?

You said "I don't know why BFG thinks it's drivers when Radeon drivers are mature. There's no evidence of magic drivers out of hat."

Then I showed that it is quite the contrary: massive performance gains in new Catalysts happen pretty frequently, even a year later (from launch.)

Why would any sane person claim "Radeon drivers are mature" regarding a brand new, month-old architecture is beyond me... you either don't even read what are you posting or you have no clue whatsoever how drivers work...

Ah hah. I didn't understand why you posted that now I do. Of course driver optimization come over time never said it didn't. not just magic drivers out of hat. I went on to say this though.

If driver optimization happen it will also be applied to RV770.

 

[AzN]

-Yeah... Ok... but it has everything to do with the subthread within the topic that suggested driver improvements would just make the 5870 suddenly unlock hidden performance above and beyond the 5850.

That's absurd when it's based on the same architecture.

-http://www.firingsquad.com/hardware/ati_radeon_5870_overclocking/

Firing Squad did a nice little OC review of the 5870 and had the decency to break down their graphs into Stock/Mem OC/Core OC/Total OC, making it very easy to demonstrate that the 5870 is NOT bandwidth limited.

Hah. I already left a comment there when they did this benchmark between the 5850/5870 while back...

Like Chizow mentioned in the other overclocking article I don't quite agree with your conclusion either about 58x0 being not bandwidth bottlnecked just by looking at average frame rates.

Shader usually has nothing to do with bandwidth. Games today just aren't shader limited except for maybe Crysis very high settings with these new cards. Sure we get some performance improvements but it's quite small compared to fillrate and bandwidth. 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 fact that you don't mention you are also overclocking the shader portion in your first article when you raised core clocks in your 5870 I had a problem with. 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 or more as your result of 5870 overclocking and now 5850.

The fact that 5850 with 70% more pixel fill, 53% texture fill, 53% more processing power only nets you 25% on average over 4890 tells you that it's more bandwidth limited than the cards before it. Not to mention in your own overclocking result of the 5850 beating 5870 by clocking it to 878/1400 that has 72TMU and 1440SP vs 850/1200 with 80TMU and 1600SP.

Then the author replied with this..

Unfortunately there's no way for me to unlink the stream processors from the rest of the graphics core, so I can't test that, but you're right in the sense that it does help the graphics core OC'ing results that were shown. I didn't mention that, but at the same time I still think you'd see the memory scale better when OC'ed if it truly were bottlenecked.

Perhaps "bottleneck" is the wrong word here. When I hear the word bottleneck, I think of something that's blatantly holding the architecture back from performing to its max potential. Right now, the evidence suggests that the graphics driver is bottlenecking RV870's performance potential more so than its 256-bit memory interface in my opinion. Sure, integrating a 512-bit memory interface may net RV870 a little more performance, but I still think they made the right call here.

Believe me, I was questioning ATI about this quite a bit at first, but now after seeing all the benchmarks I'm a believer. ATI's Dave Nalasco and Baumann made it quite clear that it was an either/or scenario. We weren't going to get 1600 shaders+512-bit memory interface with RV870 and still have a semi-affordable GPU. Something had to give and they decided it was going to be the mem interface. I discussed this in the original 5870 article last week.

Then me

1600SP and 512 bit is probably out of the question while pertaining smaller die and power consumption. ATI did the right thing by going 1600SP with 80TMU instead of going 512bit with lower SP and TMU. There's just more performance to be had with the latter if you want to impress all these hardware/gaming sites who only test average frame rates.

 

[GodisanAtheist]

That's absurd when it's based on the same architecture.

-Are we even talking about the same damn thing? I'm simply saying that expecting a future driver release to increase the performance of the 5870 but not proportionately increase the performance of a 5850 is nearly as absurd because... they're based on the same architecture... weird, its like the twilight zone!

If you're saying something different please elaborate!

 

[AzN]

-Are we even talking about the same damn thing? I'm simply saying that expecting a future driver release to increase the performance of the 5870 but not proportionately increase the performance of a 5850 is nearly as absurd because... they're based on the same architecture... weird, its like the twilight zone!

If you're saying something different please elaborate!

That's what I've been saying the whole time. That driver improvements would show up in both 5850 or 5870.

My logic was this.

5870 is being held back by bandwidth over 5850.

Then likes of dguy6789 is saying drivers is the reason why the 5850 is only 2% slower with 160SP and 8 more TMU at the same clocks.

 

[T2k]

Ah hah. I didn't understand why you posted that now I do. Of course driver optimization come over time never said it didn't. not just magic drivers out of hat. I went on to say this though.

If driver optimization happen it will also be applied to RV770.

But for God sake it: why would be?

It might will or might not - more likely that it won't as it's a new architecture.
FYI unified driver does not mean exactly same codepath - it means a unified package (which could go down to even same driver files, yes) but it's typically different architecture -> different code. Obviously when they design the chip from an existing architecture it will have a lot of commonality but it does not mean anything and certainly not anything you suggest.

 

[T2k]

That's what I've been saying the whole time. That driver improvements would show up in both 5850 or 5870.

Nope, you said RV770, not 5850... :twisted: ()

 

[AzN]

But for God sake it: why would be?

It might will or might not - more likely that it won't as it's a new architecture.
FYI unified driver does not mean exactly same codepath - it means a unified package (which could go down to even same driver files, yes) but it's typically different architecture -> different code. Obviously when they design the chip from an existing architecture it will have a lot of commonality but it does not mean anything and certainly not anything you suggest.

5870 is not new architecture. It is minor improvements over RV770 if anything at all.

 

[AzN]

Nope, you said RV770, not 5850... :twisted: ()

That was responding to your post and I still think driver improvements will also come on RV770. Just differently on the RV770 opposed to 5850/5870.

And your link with driver improvements is mostly cross fire over head that needs constant optimizing. Battleforge being a new game that hasn't had it's optimization on the RV770.

 

[ugaboga232]

Ok I am posting just to set the record straight. Azn you are blatantly wrong and I will give you the links you want.

http://www.xtremesystems.org/forums/showthread.php?t=235181
Go to the last page and see the nice graph which shows a 12% benefit from a near 50% increase in memory bandwidth.

Also you could say the same thing about the Gt200 architecture and g80.

Another reason drivers are more likely is because of the 5 way design of Ati's shaders. That means it is far harder to fully utilize the card and so some apps might only use 1200 shaders or what not (I know its only really 320).

 

[AzN]

Ha 2 posts and you are already making accusations. That graph is silly. Only testing average frame rates to show the difference when bandwidth specifically raise minimum fps.

You do know G80 was faster than g92 when certain conditions were applied although g92 was much stronger core. Specifically minimum fps, AA performance, uber high resolution performance.

You mention 5 way design of ATI's shader. You do know that RV770 has this as well which I've been saying from the beginning that whatever driver performance improvements on the 5870 will show up in RV770.

 

[BFG10K]

I don't know why BFG thinks it's drivers when Radeon drivers are mature.

Uh, because I actually tested a 5770 extensively and found it was slower than a 4850 sometimes? I’m not sure how anyone can claim the drivers are mature on that basis.

That and I found it was faster than a GTX285 at 2560x1600 with 8xAA sometimes, so I’m not sure how anyone can claim it’s starved for bandwidth.

And there’s the fact that I actually did a bottleneck test on the 5770, and I found the core had the biggest impact, not the memory.

We’ve also seen Firingsquad showing the core affects performance more than memory on a 5870 by 2.5 to 1, and the 5770 is literally it cut in half.

Tell me, what tests have you run on a 5770? What evidence have you shown us that the core affects only affects maximum while the memory only affects minimums like you’re claiming?

The problem here is that you get an idea in your head and your refuse to accept anything that disproves it, even the most basic common sense. I’ve produced benchmarks that disprove your theories. What benchmarks have you produced on the 5770 to disprove mine?

If driver optimization happen it will also be applied to RV770.

Nope, because the previous architecture doesn’t have things like software interpolators and bigger caches, for example.

 

[ugaboga232]

While I cannot find any benchmarks that have minimum fps, there is a wealth of information in that thread regarding the memory bottleneck which was determined to be nonexistant.


On the 5-way part, both the 770 and 870 are not fully utilized. This is easy to see in furmark where the temp on the rv770 series and rv870 series skyrockets as the entire card is being used. No game comes close to heating them that much. What this means is out of 800 possible shaders lets say 75% of them are being used. We have then 600 shaders working for the 4800 and 1200 for the 5800. This means the 5800 series is losing more performance and this could explain the lower than expected performance and how the extra shaders' difference is dulled between the oc'ed 5850 and 5870.

Edit: BFG said it quite well.

 

[Meaker10]

I'd like to point out AzN thinks crossfire doubles memory bandwidth so I think he needs to do some more reading....

 

[AzN]

Uh, because I actually tested a 5770 extensively and found it was slower than a 4850 sometimes? I’m not sure how anyone can claim the drivers are mature on that basis.

Come on now BFG. In all your tests in your own review. 5770 is beating 4850 in all tests except for 2560x1600 in Serious Sam 2 and COD2. I hardly call that needing driver optimization.

That and I found it was faster than a GTX285 at 2560x1600 with 8xAA sometimes, so I’m not sure how anyone can claim it’s starved for bandwidth.

Completely difference architecture that behave differently to some games. Now try that test with similar architecture like G92 G80, and the likes...

And there’s the fact that I actually did a bottleneck test on the 5770, and I found the core had the biggest impact, not the memory.

We’ve also seen Firingsquad showing the core affects performance more than memory on a 5870 by 2.5 to 1, and the 5770 is literally it cut in half.

Tell me, what tests have you run on a 5770? What evidence have you shown us that the core affects only affects maximum while the memory only affects minimums like you’re claiming?

Again showing only avg frame rates to come with that conclusion. I think I've pretty much answered this in my previous post.

The problem here is that you get an idea in your head and your refuse to accept anything that disproves it, even the most basic common sense. I’ve produced benchmarks that disprove your theories. What benchmarks have you produced on the 5770 to disprove mine?

This isn't anything new and I've been saying in our very first arguments about core needing right combination of bandwidth. It's just that the some of you can't accept when others try to show you things you don't want to agree with which is the reason why you tested your ultra core/sp/mem differences only reinforces what I've been saying about fillrate and memory. I don't need to write articles to show you when the data is already available on the internet. Which pertains to my other thread. http://forums.anandtech.com/showthread.php?t=2031140

Nope, because the previous architecture doesn’t have things like software interpolators and bigger caches, for example.

I can agree with that tis the same reason why RV8x0 is slightly faster than RV770 and some not. The difference? Mostly smoke and mirrors.

 

[AzN]

I'd like to point out AzN thinks crossfire doubles memory bandwidth so I think he needs to do some more reading....

lol. I'm sorry. You don't want to accept crossfire double bandwidth there's nothing I can do.

 

[T2k]

5870 is not new architecture. It is minor improvements over RV770 if anything at all.

Don't be silly. It is a new architecture as far as coding goes, of course it is.

 

[AzN]

Don't be silly. It is a new architecture as far as coding goes, of course it is.

whatever you want to believe.

What I consider architectural difference.

p4 to E6300. E6300 to I7. Not E6300 to E8400.

Same with GPU. G80, G92, GT200. Pretty much all the same architecture.