GTX480 Vs. HD5870 Crysis benchmark

[SRoode]

apoppin,

If you could, please downclock your memory to something like 1150 to 1175 MHz. Unrealistically high OCs to memory may be "stable", but bad memory calls usually lead to lower minimum frame rates because of the excessive (and I use that term for many reasons, one of them is that is the max you can move the slider in Afterburner on a 5850, not sure about a 5870) memory OC. The performance gain you get from the extra 10% of memory OC only results in a higher max frame rate of about 1%, but kills the min frame rate (by about 50%).

 

[SRoode]

PS - You may also be able to reduce your voltage if you do this...

 

[apoppin]

I want to share you an issue that I had recently. I modified the BIOS of one of my cards to lower the voltages at idle, and it would crank at its default vcore setting when it went 3D. Never gave me any issues with games but I found that it tended to stutter more ocassionally, FRAPS was showing like 40fps in Vantage and Metro 2033 but looked like 20fps and the frame rate was very irregular, plus the tearing was unbearable in some parts like if the cards weren't able to synchronize properly, and Furmark showed me some VRAM like anomalies, something that never happened with games.

I flashed the card to its default values and the stuttering went away, never happened again. It happened in Vantage, Crysis Warhead and Metro 2033, no game crashed or showed artifacts. While my GPU isn't the same as your HD 5870, I think that probably too low voltage may cause some unpredictable performance behavior like the once I experienced.

Thanks for sharing. And what does it have to do with my overclocking review of GTX vs. HD 5870?

i did not flash my card's BIOS, nevermind flash to to a wrong one; i did not have a low voltage, i did not experience stuttering; i did not experience unpredictable behavior, i did not experience anything except faster performance (and more heat and noise) in every single test i gave my cards.
Frankly, it scaled properly as a HD 5870 is supposed to scale with overclocking.

apoppin,

If you could, please downclock your memory to something like 1150 to 1175 MHz. Unrealistically high OCs to memory may be "stable", but bad memory calls usually lead to lower minimum frame rates because of the excessive (and I use that term for many reasons, one of them is that is the max you can move the slider in Afterburner on a 5850, not sure about a 5870) memory OC. The performance gain you get from the extra 10% of memory OC only results in a higher max frame rate of about 1%, but kills the min frame rate (by about 50%).

Absolutely not. Utter nonsense.

Both of my HD 5870s have no problem with 1300MHz on the memory; it is just as stable at 1200 MHz as it is at 1300 MHz

i test each clock independently of the other. When my diamond HD 5870 is at 975/1300 it is just as stable as it was at 800/1200

There is no min frame rate being killed. i overclocked it to its absolute limit on air cooling but i did not exceed its limits.

From what you are saying, i don't think some have a real clue about finding a truly stable overclock

 

[ugaboga232]

The 5870 isn't starved for bandwith. However, if you increase its speeds too much (seeing as your chip is "Average," i'd downclock the memory just to check), its ECC begins to drastically lower performance. It might be a small difference, but the gain from oc'ing the memory on 5XXX is almost nil.

 

[bunnyfubbles]

To me minimums are more important. But regardless, I think Crysis shows a weakness in Fermi - either that or Nvidia has some serious driver optimizations to do. But more than likely Fermi's texture management looks suspect.

When the frames are that low to begin with, I'd rather have the higher average. Then again, the numbers are so close here calling either card a winner is kind of pointless IMO.

 

[RussianSensation]

The 5870 isn't starved for bandwith. However, if you increase its speeds too much (seeing as your chip is "Average," i'd downclock the memory just to check), its ECC begins to drastically lower performance. It might be a small difference, but the gain from oc'ing the memory on 5XXX is almost nil.

Ya 5870 is primarily GPU limited. Not only that, but starting with 5xxx series, excessive memory overclocking does not produce any crashing so there is no way to check that you excessively overclocked the memory without checking benches. :\

 

[apoppin]

The 5870 isn't starved for bandwith. However, if you increase its speeds too much (seeing as your chip is "Average," i'd downclock the memory just to check), its ECC begins to drastically lower performance. It might be a small difference, but the gain from oc'ing the memory on 5XXX is almost nil.

Let me try one more time and please follow along.

i have *two* HD 5870s and they have been checked *thoroughly* - against each other
- to make absolutely positively with 100% certainty that they are scaling properly from 850 MHZ on the core all the way up to 925 MHz (where one stops on stock voltage) and from 1225 MHz on the vRAM clocks all the way up to 1335 MHz.

On the Diamond at 1301 MHz there is very slight instability that shows with an extreme core O/C; a couple of the benchmarks show very slightly lower performance. However, at 1304 MHz it exhibits some of what you are talking about including some GSoDs and it is completely useless for testing at that frequency or above.

Are we clear now? 1300 MHz is the "cutoff"' - the absolute pinnacle of this particular card's performance is at 975/1300 MHz

i spent a LOT of time testing my pair of HD 5870s and i am quite sure of their capabilities

 

[HurleyBird]

My performance scaling results sure won't be any difference than any other HD 5870 at 975/1300. And i am certain there was complete 100% stability across every demanding benchmark i could give it including maxed out details at 2560x1600. i am sure i did not have to raise the fan speed as far as i did.

Apoppin:

I'm pretty sure the Fermi does scale better than Cypress, but the results you obtained look somewhat abnormal, especially when you look at something like the 5970 toxic edition which is exhibiting great performance scaling (http://www.tweaktown.com/articles/3230/sapphire_radeon_hd_5970_toxic_4gb_video_card/index.html)

I can think of a couple of reasons why you might be getting lower results. I do believe you when you say that your card is stable, but it is true that memory clocked too high can lower performance while being stable due to error checking. You also needed a lot of volts to obtain your overclock, and don't forget that the 5870 *will* throttle to reduce loads on the VRMS (like in furmark). This might explain why your 5870 didn't scale very well (as in barely at all) in a number of games. Quick test might include looking at the VRM temps in GPU-Z for various scenarios. If the VRM temps are higher for the games that the 5870 scales poorly in, you've probably found your culprit. If not, I guess I'm wrong

 

[apoppin]

Apoppin:

I'm pretty sure the Fermi does scale better than Cypress, but the results you obtained look somewhat abnormal, especially when you look at something like the 5970 toxic edition which is exhibiting great performance scaling (http://www.tweaktown.com/articles/3230/sapphire_radeon_hd_5970_toxic_4gb_video_card/index.html)

I can think of a couple of reasons why you might be getting lower results. I do believe you when you say that your card is stable, but it is true that memory clocked too high can lower performance while being stable due to error checking. You also needed a lot of volts to obtain your overclock, and don't forget that the 5870 *will* throttle to reduce loads on the VRMS (like in furmark). This might explain why your 5870 didn't scale very well (as in barely at all) in a number of games. Quick test might include looking at the VRM temps in GPU-Z for various scenarios. If the VRM temps are higher for the games that the 5870 scales poorly in, you've probably found your culprit. If not, I guess I'm wrong

WHY the heck are you showing me 5970?
- that is no comparison at all to 5870, a single GPU.

Did you read my last post? The one before yours?

i tested everything that can be tested including running the vRAM at 1250 MHz to be sure. There is almost no performance difference between running 975/1250 MHz or at 975/1300 MHz (except the runs at 1300 MHz are consistently very slightly higher - that tells ME something)
- sheesh with all this nonsense i am getting from you guys, you'd think i had never overclocked before and if i had run the goddamn thing at 1250 MHz you'd be b!tching that i ran it too low.


i am beginning to understand that some of you guys simply cannot *handle* that GTX 480 scales better than HD 5870. It should. GF100 Fermi is 6 months late, hot and noisy .... and it is also new architecture.

Hmm .. i don't think i care to let you know the results of my 8xMSAA performance hit over 4xMSAA on the GTX 480 vs. HD 5870 with 22 games.
- from a preliminary look, you won't like those results either. Stay in the dark and believe whatever nonsense you want to.

 

[HurleyBird]

Aponnin:

First off, do away with the overly defensive attitude, please.

No, I didn't read the last post. Sorry, but I didn't feel like wading through 4 pages of bickering. Since you tested at 1250 and 1300 then I think you are right about RAM, or at the very least any performance decrease from error correction is outweighed by the performance increase from higher frequency.

I linked you the result of the 5970 toxic edition because:
a. its showing very good scaling with increased clock speed, which seems to be in contradiction of with data. (that's assuming the the 2gb frame buffer isn't playing too much of a role at those settings, eyefinity 6 and 2gb 5870 previews suggests that it shouldn't be.)
b. Who cares if it's crossfire and a different card? It's the same architecture, frequency scaling should not be much affected.

I also gave a potential reason for your (seemingly) abnormal results: throttling due to overheating VRMs. I also gave you a methodology to test if this theory holds ground, and conditions that would prove my theory wrong. Obviously, if you prove my theory to be wrong (and it's easy to do if it is), then you are actually strengthening your results. Does that sound like a rabid fanboy or someone who wants to help find the truth?

The only other rational solution that I can see for the difference in results is that there is something that kicks in around ~850 MHz that negatively affects scaling of the architecture. Sounds pretty far fetched to me, but in my experience the Cypress architecture is pretty quirky in a lot of ways.

 

[apoppin]

You keep asking me to run tests that i keep telling you - over and over - i have already run. i don't need to repeat what you are asking to do over again. It is done. The results are in. Both overclocks are 100% stable. The Radeon scales like Radeons do.

You'd be defensive if you kept being questioned over and over again about the same damn thing and no one bothered to pay the slightest attention to your answer.

The only other rational solution that I can see for the difference ...

Or it could JUST could be - as *shocking* as it appears to be to some of you ...
. . . GTX 480 just scales better than HD 5870

 

[cbn]

Or it could JUST could be - as *shocking* as it appears to be to some of you ...
. . . GTX 480 just scales better than HD 5870

http://forums.anandtech.com/showpost.php?p=29678326&postcount=57

Yep, Sylvanas did a good job explaining why this happens.

 

[cbn]

I linked you the result of the 5970 toxic edition because:
a. its showing very good scaling with increased clock speed, which seems to be in contradiction of with data. (

Yes, but you have to remember 5970 starts off with a 725 Mhz core. This is 125 Mhz short of a 5870's stock clock.

 

[apoppin]

http://forums.anandtech.com/showpost.php?p=29678326&postcount=57

Yep, Sylvanas did a good job explaining why this happens.

Yes. i demonstrated it. HD 5870 is finally at the end of its architectural life. It has a speed bump left perhaps in a 5890 but it will eventually become unbalanced if they push it much further.

5970 has a lower clock than 5870 and thus it can be pushed harder and will scale better as it hits it own extreme overclock on the core.

i know this is sad news for some but AMD knows this and they are hard at work on their new architecture. Fermi is brand new and Nvidia needs to work the kinks out of the high-end process. It will be interesting to see this happen.

 

[HurleyBird]

You keep asking me to run tests that i keep telling you - over and over - i have already run. i don't need to repeat what you are asking to do over again. It is done. The results are in. Both overclocks are 100% stable. The Radeon scales like Radeons do.

You'd be defensive if you kept being questioned over and over again about the same damn thing and no one bothered to pay the slightest attention to your answer.

Or it could JUST could be - as *shocking* as it appears to be to some of you ...
. . . GTX 480 just scales better than HD 5870

So, you did test the VRM temps and they weren't hotter in the instances where 5870 scaled poorly? I just did a quick one over of the thread, and unless I'm mistaken you never mentioned that you did. And even if you did, just saying so one more time can't be more troubling to you more than the long winded rant can it? Have some class, man.

And as far as stability goes, let me repeat myself one (hopefully) last time. I have no doubts that the card is 100% stable, however, if the VRMs *are* overheating and the chip *is* being throttled that wouldn't necessarily manifest itself as instability, although it would reduce performance.

As far as GTX 480 scaling better than HD 5870, again, I have no problem with that. Most of your 5870 OC results looks like the GPU is being severely bottlenecked by something else, like the CPU, or PCIe bandwidth, while the Fermi results don't. Fermi scaling better than Cypress is fine. Fermi scaling extraordinarily while Cypress' gains are simply minuscule simply looks fishy, especially when I have another example exhibiting pretty decent scaling.

Yes. i demonstrated it. HD 5870 is finally at the end of its architectural life. It has a speed bump left perhaps in a 5890 but it will eventually become unbalanced if they push it much further.

Except it doesn't -- or at least shouldn't -- work that way. Sure, as you increase the clock speed of one part the rest of the system becomes more and more of a bottleneck, everyone knows that, but you shouldn't see a giant nosedive after a specific frequency like we're seeing with your results vs. other people's unless something *very wrong* is happening. I mean, you have faster parts (5970 and GTX 480) that aren't yet causing the rest of the system to become a bottleneck, so it's obviously not that. Whether it is a flaw in your testing (either in the equipment or methodology), a driver bug, or just a very strange architectural quirk, I don't know, just that it is not normal behavior for a processor to be exhibiting. I'd like to find out, by first cancelling out the (no offense) seemingly most likely scenario, but you don't seem to be making that easy....

 

[apoppin]

i am not *repeating* any more tests just for you.

There were no bottlenecks with i7 at 3.80 GHz/ no PCIe issues with X58. Nothing unexplained as to performance and no vRAM overheating
- identical testbed for GTX 480 and HD 5870.

You can believe what you want to believe; you will anyway
--i am done wasting time. i have more benching to do and an article to write tonight.

 

[HurleyBird]

Sheesh, all the beating around the bush is getting old. I'm thinking your last post is code for "I didn't check to see if the VRMs were overheating and causing throttling". I can believe what I want to? If you would stop being difficult and come out and say something like "yes, I tested VRM overheating", "No, I didn't", or "No, but that's an idea I'll try" I'll believe you -- I'm actually a fairly trusting person!

I just don't see why you are being so avoidant, and rude, about such a simple thing.

 

[cbn]

Yes. i demonstrated it. HD 5870 is finally at the end of its architectural life. It has a speed bump left perhaps in a 5890 but it will eventually become unbalanced if they push it much further.

It will be interesting to see how Southern Islands turns out. Apparently that will use a refreshed Cypress core with the uncore from N. Islands.

Personally, I would rather see ATI release a smaller but complete version of N.Islands on 40nm instead.

P.S. Thanks for the testing and work you do Appopin.

 

[Nemesis 1]

Si at 40nm makes no sense . NI at 40nm makes sense. New arch on old process. AMDs been around the ball diamond. If they make same mistake as NV . New process with new Arch would be dumb as hell.

 

[MagickMan]

I found that OCing my 5850 RAM over 1200MHz causes a decrease in minimum framerates, and it gets worse the higher I push the clocks. Yes, I can set it to 1300, and the average will climb a little, but the Min always goes down. There's some powerful ECC going on, and it keeps the card from bombing when there's an error, but it comes at a cost.

Now, what this means for the 5870 is less clear, since they usually use the same RAM but at a higher voltage. Unfortunately, this particular voltage can't be modified in the BIOS.

 

[alcoholbob]

Lol this has gotten way off-topic.

I think the important message of this thread was: Fermi was a waste of 6 months if you are looking for a single card to run Crysis comfortably in DX10 VH @ 19x12. In fact, it does not better than the 5870.

All the other titles out on the market run blazingly fast already, to the point that typically the difference between a $300 card and a $600 card is laughably irrelevant. The actual trophy benchmark--Crysis--is massive fail for Fermi.

 

[Voo]

Sheesh, all the beating around the bush is getting old. I'm thinking your last post is code for "I didn't check to see if the VRMs were overheating and causing throttling". I can believe what I want to? If you would stop being difficult and come out and say something like "yes, I tested VRM overheating", "No, I didn't", or "No, but that's an idea I'll try" I'll believe you -- I'm actually a fairly trusting person!

I just don't see why you are being so avoidant, and rude, about such a simple thing.

He said that he tested with lower memory clocks and those results didn't have higher performance. So if you get higher results from higher clocks how exactly do you think the card is being throtteled? That could only be true for the last result because we don't know any higher values, but if it still gets better performance than a lower OC it's rather reasonable to assume that the card doesn't throttle.

 

[RussianSensation]

Yes. i demonstrated it. HD 5870 is finally at the end of its architectural life. It has a speed bump left perhaps in a 5890 but it will eventually become unbalanced if they push it much further.

http://www.firingsquad.com/hardware/ati_radeon_5870_overclocking/page12.asp

9% GPU overclock resulted in a 5% performance increase on 5870
9% memory overclock resulted in a 2-4% performance increase on 5870

This means a 9% overclock on GPU and Memory should produce a 7-9% performance increase overall. How is this poor scaling?

From a shader computational perspective, 5870 is 30% faster in shader ops over 5850 (850 x 1600 / 725 x 1440). If 58xx series scales poorly, why is it you can overclock 5850 to 950mhz and then match 5870's performance despite having 1440 shaders (i.e., 950 x 1400 / 850 x 1600)? 950 over 725 is a 31% overclock. Something doesn't add up for me. :hmm:

 

[HurleyBird]

He said that he tested with lower memory clocks and those results didn't have higher performance. So if you get higher results from higher clocks how exactly do you think the card is being throtteled? That could only be true for the last result because we don't know any higher values, but if it still gets better performance than a lower OC it's rather reasonable to assume that the card doesn't throttle.

There's a difference between error correction on the memory, which is not the same as throttling, and the VRMs overheating and throttling the shader units, like they already are known to do with fur mark. It seems like a likely scenario based on how little scaling Apoppin found in most of his games, and the fact that he had to give the card a lot of voltage to get his overclock.

 

[apoppin]

There's a difference between error correction on the memory, which is not the same as throttling, and the VRMs overheating and throttling the shader units, like they already are known to do with fur mark. It seems like a likely scenario based on how little scaling Apoppin found in most of his games, and the fact that he had to give the card a lot of voltage to get his overclock.

it is pathetic to watch you grasp at straws as you refuse to even acknowledge what i wrote before. But i will post it again for others:

There were no bottlenecks with i7 at 3.80 GHz/ no PCIe issues with X58. Nothing unexplained as to performance and no vRAM overheating
- identical testbed for GTX 480 and HD 5870.

i tested the vRAM independently of the core also. No errors to affect performance. Stable at 1300 MHz.

You can't seem to grasp that the card that you love so much scales less in performance benchmarks the more you push it; that much is true. But what you really don't seem to understand is that i really like my HD 5870.
i am impressed that a $400 video card puts out such great performance - quietly and elegantly.

However, you notice that MSI stops at 900 MHz on its Lightening 5870's core overclock. Why? Because it gets hot and noisy beyond that and it does not scale as well as it did up to 950 MHz; that is a good balance MSI gives for $500 and it pretty much catches the stock GTX 480 in close benches.

Some people like to push it even further to make a "gtx 480" out of a 5870 and will aim for 1000 MHz or more on the core. That is fine and they do get more performance but it is diminishing returns and they have to watercool or put up with noise.

Basically we can see from my review that a decently overclocked $400 HD 5870 *catches* a stock $500 GTX 480. What is wrong with that? We also see the GTX 480 on new architecture, scaling better - what a surprise.

And quit showing me HD 5850 anything.



i'm back to work.