HD58xx and memory bandwidth issues.

[cbn]

1600 stream processors @850 Mhz sounds like a ton of GPU processing power.....but then I look at the memory bandwidth for these cards and it is only marginally improved.

In what scenarios will the memory bandwidth of HD58xx be a limitation? Obviously when the memory bus is saturated GPU power figures less and less into total graphics processing time (which influences frame rates from the Video card side of things)

With HD4890 and HD4870 it seemed the bottleneck was GPU power http://www.anandtech.com/video/showdoc.aspx?i=3555

[Side note] With GT300 the relationship appears to be the same. 50% more transistors than HD58xx coupled to 50% more memory bandwidth. For this reason I wonder if some of the value priced products in both the GT300 and HD5xxx line-up will be more a bargain provided they come close to having the same memory bandwidth as their respective flagship counterparts.

 

[dguy6789]

A website tested and found that in almost all cases, the 5870 isn't memory bandwidth starved. I searched for a bit and couldn't find it. Hopefully someone will post it.

 

[SSChevy2001]

Originally posted by: dguy6789
A website tested and found that in almost all cases, the 5870 isn't memory bandwidth starved. I searched for a bit and couldn't find it. Hopefully someone will post it.

Here you go.

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.

http://www.firingsquad.com/har...verclocking/page12.asp

If you compare the 4870 vs 4890 you'll notice a decrease in memory / GPU clock. Now if you take that decrease and try to figure out what the 5870 should have, should look like the following.

( ( 975/850 ) - ( ( ( 900/750 ) - ( 975/850 ) ) * 8.5 ) ) * 1700 = 1185MHz

 

[cbn]

Originally posted by: SSChevy2001

Originally posted by: dguy6789
A website tested and found that in almost all cases, the 5870 isn't memory bandwidth starved. I searched for a bit and couldn't find it. Hopefully someone will post it.

Here you go.

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.

http://www.firingsquad.com/har...verclocking/page12.asp

If you compare the 4870 vs 4890 you'll notice a decrease in memory / GPU clock. Now if you take that decrease and try to figure out what the 5870 should have, should look like the following.

( 975/850 ) - ( 0.052941176470588 * 8.5 ) * 1700 = 1185MHz

If my math is right 1200MHz is more than enough.

Thank you.....but I wish that review also showed minimum frame rates.

It would seem to me memory bandwidth issues wouldn'ty really show up if looking at average frame rates. Seriously if memory bandwidth is already becoming an issue in average game play you know things would be absolutely terrible at peak Video card load.

 

[SSChevy2001]

Originally posted by: Just learning
Thank you.....but I wish that review also showed minimum frame rates.

It would seem to me memory bandwidth issues wouldn'ty really show up if looking at average frame rates. Seriously if memory bandwidth is already becoming an issue in average game play you know things would be absolutely terrible at peak Video card load.

The problem with minimum frame rates is that you need some sort of % to know how long it was there.

If it's only at it's lowest for 1% of the time then it doesn't matter.

 

[Forumpanda]

Yes, the very minimum frame rate could also be influenced by various OS background tasks or loading data from the hard drive.

I really wish those reviews which feature minimum frame rates would do multiple run through and discard some of the lowest measured FPS.
I'm sure someone with a bit of statistical background could come up with a reasonable way to describe 'playability' based on a series of FPS measurements over time.

As is running through a benchmark 1 time and measuring the very lowest FPS that got hit does not really tell me as a reader anything useful.

I would also think memory bandwidth issues are more likely cap the max FPS rather than cause minimum FPS, but maybe I am wrong, but I would just think that min FPS are more likely to be caused by a specific compute (either CPU or GPU) or load (hard drive) intensive scene, I would doubt bandwidth starvation would cause and FPS lower than 30.

I of course have no technical knowledge to postulate that, it is just my educated guess.

 

[cbn]

Originally posted by: SSChevy2001

Originally posted by: Just learning
Thank you.....but I wish that review also showed minimum frame rates.

It would seem to me memory bandwidth issues wouldn'ty really show up if looking at average frame rates. Seriously if memory bandwidth is already becoming an issue in average game play you know things would be absolutely terrible at peak Video card load.

The problem with minimum frame rates is that you need some sort of % to know how long it was there.

If it's only at it's lowest for 1% of the time then it doesn't matter.

http://www.xbitlabs.com/articl...n-hd5870_10.html#sect2

Look at this review.

HD5870 has much higher average frame rates.....but minimum frame rates are the same as HD4890.

I haven't looked at the other games but I would expect much of the same thing to be happening.

So if you are someone who is mostly concerned with how much a game slows down at load memory bandwidth of HD58xx seems to be pertinent.

 

[alcoholbob]

Looks faster than me at 19x12. At the lower resolutions all the cards are CPU bound.

 

[cbn]

Originally posted by: Forumpanda
yes, the very minimum frame rate could also be influenced by various OS background tasks or loading data from the hard drive.

I really wish those reviews which feature minimum frame rates would do multiple run through and discard some of the lowest measured FPS.
I'm sure someone with a bit of statistical background could come up with a reasonable way to describe 'playability' based on a series of FPS measurements over time.

As is running through a benchmark 1 time and measuring the very lowest FPS that got hit does not really tell me as a reader anything useful.

Yeah but looking through other comparisons it seems minimum frame rates are holding fairly consistent for HD58xx.

But I agree with you in that other factors can influence min. frame. For example in Crysis hard drive speed can affect frame rates.

http://www.xbitlabs.com/articl...on-hd5870_9.html#sect2

 

[cbn]

Originally posted by: Astrallite
Looks faster than me at 19x12. At the lower resolutions all the cards are CPU bound.

You are right about that. I guess I made a hasty judgement.

Wow...I guess this means HD4870 and HD4890 were grossly over-spec'd as far as memory bandwidth goes all along.

 

[netxzero64]

@astralite

then what is the minimum resoulution that a game will be GPU bound not CPU then?

 

[alcoholbob]

Originally posted by: Just learning

Originally posted by: Forumpanda
yes, the very minimum frame rate could also be influenced by various OS background tasks or loading data from the hard drive.

I really wish those reviews which feature minimum frame rates would do multiple run through and discard some of the lowest measured FPS.
I'm sure someone with a bit of statistical background could come up with a reasonable way to describe 'playability' based on a series of FPS measurements over time.

As is running through a benchmark 1 time and measuring the very lowest FPS that got hit does not really tell me as a reader anything useful.

Yeah but looking through other comparisons it seems minimum frame rates are holding fairly consistent for HD58xx.

But I agree with you in that other factors can influence min. frame. For example in Crysis hard drive speed can affect frame rates.

http://www.xbitlabs.com/articl...on-hd5870_9.html#sect2

Maybe you meant to link this?

http://images.anandtech.com/gr...090808151933/17331.png

 

[alcoholbob]

Originally posted by: netxzero64
@astralite

then what is the minimum resoulution that a game will be GPU bound not CPU then?

GPU bound/CPU bound is a bit of a misnomer. Games are always bound by *both* constraints. GPU bound means a game (sometimes only apparent after a certain resolution) displays far greater improvement with GPU speed than CPU speed.

However you will hit a wall of diminishing returns if you only increase one aspect, no matter how much a game leans toward "GPU bound" or "CPU bound." Once you reach that point, you will have to either increase CPU speed or GPU power to return to linear performance scaling.

Take Crysis for example. The game is considered heavily GPU bound because it takes significant increases in GPU power to increase framerate. However, minimum framerates seem to scale linearly with CPU speed up to 3.6ghz in Crysis (when you are testing with the same GPU). So the game is considered GPU bound, but you'd be really shortchanging yourself if you weren't running at a higher cpu clockrate.

 

[Idontcare]

Originally posted by: Astrallite

Originally posted by: Just learning

Originally posted by: Forumpanda
yes, the very minimum frame rate could also be influenced by various OS background tasks or loading data from the hard drive.

I really wish those reviews which feature minimum frame rates would do multiple run through and discard some of the lowest measured FPS.
I'm sure someone with a bit of statistical background could come up with a reasonable way to describe 'playability' based on a series of FPS measurements over time.

As is running through a benchmark 1 time and measuring the very lowest FPS that got hit does not really tell me as a reader anything useful.

Yeah but looking through other comparisons it seems minimum frame rates are holding fairly consistent for HD58xx.

But I agree with you in that other factors can influence min. frame. For example in Crysis hard drive speed can affect frame rates.

http://www.xbitlabs.com/articl...on-hd5870_9.html#sect2

Maybe you meant to link this?

http://images.anandtech.com/gr...090808151933/17331.png

Kinda amazing that some rate-limiting step in rendering the frames in Crysis involves not only stepping beyond the on-die caches in an effort to find that critical bit of data for the frame to be rendered but also that it has to step to the other side of the IMC and the ram is still not enough, it has to go all the way to the hard-drive to do whatever needs to be done at that critical moment in time during a frame rendering in order to get done what it needs to so it can then push the frame to the video screen.

Would be cool for Anand to add a test using a ramdrive on an x58 rig to just show us how much of a min-framerate barrier the HDD is putting up. (i.e. generate the best-case corner condition so we know how much more performance is being left on the table even with the Intel SSD's)

 

[alcoholbob]

Originally posted by: Idontcare

Kinda amazing that some rate-limiting step in rendering the frames in Crysis involves not only stepping beyond the on-die caches in an effort to find that critical bit of data for the frame to be rendered but also that it has to step to the other side of the IMC and the ram is still not enough, it has to go all the way to the hard-drive to do whatever needs to be done at that critical moment in time during a frame rendering in order to get done what it needs to so it can then push the frame to the video screen.

Would be cool for Anand to add a test using a ramdrive on an x58 rig to just show us how much of a min-framerate barrier the HDD is putting up. (i.e. generate the best-case corner condition so we know how much more performance is being left on the table even with the Intel SSD's)

I suspect he ran the "GPU_bench" timedemo which is a flyby over the island in Crysis. I noticed that the C-130 intro movie got SIGNIFICANTLY smoother even though the framerate counter remained largely the same when going from a VR 300 to an Intel X25-E .

What my guess is, since the camera is moving so fast across the map, the limitation becomes texture streaming speed.

Regarding RAM--remember, the developer does not know how much RAM the user has, as a result, the game will have a limited amount of the map pre-cached. The "seamless game world" that started with Oblivion and Dungeon Siege were really just a map with an invisible border (with only a low quality texture version of incoming area visible), and once you reached the border the game would pause for a second as the new area loads from disk. When these transitions occur on an SSD it's much faster.

 

[BFG10K]

Originally posted by: Astrallite

I suspect he ran the "GPU_bench" timedemo which is a flyby over the island in Crysis. I noticed that the C-130 intro movie got SIGNIFICANTLY smoother even though the framerate counter remained largely the same when going from a VR 300 to an Intel X25-E .
What my guess is, since the camera is moving so fast across the map, the limitation becomes texture streaming speed.

Yes, exactly. The island demo isn't indicative of actual gameplay because it moves through the level much faster than regular gameplay would, and hence falls victim to data streaming limitations.

I personally don?t use that benchmark for this reason but if you must use it, always use the second run since the data is more likely to be cached to some degree.