HD 5870 Discussion thread.

[Binge]

You know, some people just can't get over themselves you know
They always think the 5870 needs more bandwidth because the nVidia cards are supposed to have more.

I doubt that's the reason. I think the reason is because of the 4XXX series having this much bandwidth. The next gen should have more etc etc. There is a reason more bandwidth could provide a noticable (20%) improvement, but I don't want to start another argument.

 

[Zubasa]

I doubt that's the reason. I think the reason is because of the 4XXX series having this much bandwidth. The next gen should have more etc etc. There is a reason more bandwidth could provide a noticable (20%) improvement, but I don't want to start another argument.

No doubt that more memory bandwidth increases performance, but I guess we might never know how much a wider bus really helps.

 

[Bo_Fox]

I doubt that's the reason. I think the reason is because of the 4XXX series having this much bandwidth. The next gen should have more etc etc. There is a reason more bandwidth could provide a noticable (20%) improvement, but I don't want to start another argument.

No doubt that more memory bandwidth increases performance, but I guess we might never know how much a wider bus really helps.

QFT. I'd pretty much say the same thing.

 

[Bo_Fox]

Right, I'm just going to do some investigations into the effect of memory overclocking on my hd5870, the core clock on the card will be set to 900MHz for all the following results, i'll update this post as i do more benchmarks.

First up is left 4 dead.
Time demo was recorded for the first level of no mercy, from the rooftop start point to entering the safe room door, framerates recorded with fraps. Any aspect of human error has been included with results. settings: 1680x1050, maximum, 8xmsaa and no vsync.

900/1200: 118fps average +/- 2
900/1250: 117fps average +/- 2
900/1300: 118fps average +/- 2

So for the first test we see no gain from memory overclocking outside the bounds of any human error that could occur. From an 8% increase, we would at least expect to see a small increase in the region of 3-4 fps in a memory bottlenecked system. It is worthwhile to note that gpu usage was around 60% at most, while cpu usage was 50% at most, so neither were pushed to the limit, therefore the l4d test should be taken with a pinch of salt.

UNIGINE DX11 BENCHMARK

moving on to one of the more extreme gpu tests for the hd5870, this benchmark pushed the gpu usage up to 100% as expected, so the memory bandwidth should become more of a factor than in the l4d test. Settings used were 1680x1050, 4xMSAA, all settings highest and tessellation on. (Vsync turned off, ofc).

900/1200: FPS 31.9 [score 803]
900/1250: FPS 32.3 [score 814]
900/1300: FPS 32.8 [score 825]

this time around we see a linear increase when increasing the memory speed. However, the increase is only in the region of around 1% in each case. If the card were truly bottlenecked, we would expect to see a more drastic increase, but as it stands it does appear that the speed of the card seems to be pretty well matched against the memory speed. Of course, that may all change as ati improve the drivers to unleash more power from the card. Comparing the card to other generations is not feasible either, if the hd4890 was bottlenecked by it's memory bus, how can the hd5870 be so much faster if it is using the same memory? the increase in bandwidth is only around 33% but the increase in overall speed is around 60-80%. The numbers just don't add up, also remember that ati aren't idiots, if they had an epic card that would be severely bottlenecked by a 256 bit bus, they wouldn't used one. The card would cost more, that's for sure, but it would also be a helluva lot faster if it was severely bottlenecked. My experimenting with memory clocks has not yielded anything to indicated such a bottleneck thus far (and i have done others as well as the ones shown here), but as stated earlier, driver improvements in the future will likely prove whether or not the bottleneck exists.

However it could be entirely likely that the memory speed has a far lesser effect at the resolutions i am testing at, especially in l4d. If that is the case, i'm not too worried about the whole thing

Thanks, but did you get a 5870 for just 1680x1050?

You could say the same thing about a 4890 performing only ~20% better than a 5770 despite having 62% greater memory bandwidth. But for a 20-25% increase on the top-end segment, some people are willing to pay an extra $100. As the CPU's get faster, along with system memory, etc.. and we start playing newer and more demanding games, there might be greater need for more memory bandwidth.

To correct you, a 5870 has only 23% greater bandwidth than a 4890, not 33%.

 

[bobzilla2009]

Thanks, but did you get a 5870 for just 1680x1050?

You could say the same thing about a 4890 performing only ~20% better than a 5770 despite having 62% greater memory bandwidth. But for a 20-25% increase on the top-end segment, some people are willing to pay an extra $100. As the CPU's get faster, along with system memory, etc.. and we start playing newer and more demanding games, there might be greater need for more memory bandwidth.

To correct you, a 5870 has only 23% greater bandwidth than a 4890, not 33%.

right you are, i've changed it in the post (i was quite sleepy when i typed it) but i did buy a hd5870 for my native 1680x1050 monitor, but i move my computer around a bit, and tend to play on 1920x1080 hdtv's fairly often too

 

[bobzilla2009]

WTFBBQ!!!!!!


The increase was from the marginaly faster cache miss data delivery than from the extra bandwidth. XYZOMG!!!!


No can we please discuss the new hardware the card has?


can you run s set with tessellation on and off Bob?

yup, will do that now for you and post the results in a minute

 

[a_ump]

sweetness, i'm very curious to see the difference tesselation makes as well.

 

[bobzilla2009]

UNIGINE DX11 BENCHMARK (2)

moving on to one of the more extreme gpu tests for the hd5870, this benchmark pushed the gpu usage up to 100% as expected, so the memory bandwidth should become more of a factor than in the l4d test. Settings used were 1680x1050, 4xMSAA, all settings highest and tessellation on. (Vsync turned off, ofc).

900/1200: FPS 31.9 [score 803]
900/1250: FPS 32.3 [score 814] (4.2% increase in speed, 1.3% increase in framerate)
900/1300: FPS 32.8 [score 825] (4% increase in speed, 1.5% increase in framerate)

the story is much the same without tessellation:

900/1200: 51.5 FPS - 1297
900/1250: 52.5 FPS - 1324 (4.2% increase memory, 1.9% frames)
900/1300: 53.2 FPS - 1339 (4% memory increase, 1.3% frames)

Interestingly:

850 core/1300 memory provides a score of 31.4fps (791) with tessellation, meaning that a 5.8% increase in core clock (up to 900) yields a 4.5% increase in framerate. Therefore, the main 'bottleneck' atm is just the core speed and not the memory.

Based on the scope of tests i have used thus far, it would seem that the memory speed is not really an issue with the card. The meagre increases in speed compared to the overclocks speak for themselves, and i believe the difference in framerate with regards to core clocks shows that the core itself is not memory limited significantly, if at all. This could all change if driver improvements sends performance sky rocketing, and i will return if that is the case if anyone would like any other benchmarks, just say and i'll see what i can do.

 

[newfellow]

Well, about memory bandwidth.

Don't really see how higher bandwidth in HD5850 case for example would of helped nor on HD5870 in that manner, if you raise the 1000Mhz on HD5850 to 1200Mhz for example alone it does 0 fps increase to anything as far I can see. Raising core however drops damn huge increase, but well there's many kind of tests this could be argue. DDR is only 128-bit on motherboard so that's another one to consider moving data on system memory to GPU will always occure and including Vista/7 style Graphic memory interface it'll split anyway no application will detect that it's actually run on card.

I'd say only way to actually do an increase ala. microsoft style is to simply have GPU which has as much memory as half of system RAM (which is the amount forced by Vista/7 to be used as dedicated graphic data). Increasing RAM like now we have Triple-channels etc etc is worthless as it increases the dedicated memory & as long dedicated memory is larger then GPU memory it's always a bottleneck. Very odd concept they are building.

 

[bobzilla2009]

Well, about memory bandwidth.

Don't really see how higher bandwidth in HD5850 case for example would of helped nor on HD5870 in that manner, if you raise the 1000Mhz on HD5850 to 1200Mhz for example alone it does 0 fps increase to anything as far I can see. Raising core however drops damn huge increase, but well there's many kind of tests this could be argue. DDR is only 128-bit on motherboard so that's another one to consider moving data on system memory to GPU will always occure and including Vista/7 style Graphic memory interface it'll split anyway no application will detect that it's actually run on card.

I'd say only way to actually do an increase ala. microsoft style is to simply have GPU which has as much memory as half of system RAM (which is the amount forced by Vista/7 to be used as dedicated graphic data). Increasing RAM like now we have Triple-channels etc etc is worthless as it increases the dedicated memory & as long dedicated memory is larger then GPU memory it's always a bottleneck. Very odd concept they are building.

The various benchmarks i've done over the past two pages do agree somewhat with your statement of no increases with regards to higher memory speeds. However there are always some small gains to be had when the core speed has been increased too. I think that the cards are just not memory bottlenecked at all, and the evidence i've seen just acts to justify such a theory. When i have time i will continue my comprehensive testing by investigating the gains to be had when changing memory speeds with the stock core speed (probably in unigine again).

 

[20mmrain]

Okay I am now convinced that there is something holding this card back and I am now sure it is the drivers!
I just now got to finally game with my new 5870 for the first time. I am using 9.10 drivers and you get some weird results in Frames Per Second especially with Nvidia coded games. For instance I was playing Crysis and Crysis warhead and those games did fine and way better than before(even though it's a Nvidia game) But than I played Far Cry 2 and got some really weird results. Like for most of the time with that game I will be way above 60FPS but for a split second the whole time planing I would drop from 60FPS to 30FPS and it wouldn't even be a stressful part.
Now I do understand that FPS does drop from time to time but I've never seen a card have that big of a fluctuation. But next I would play some ATI coded games like HAWX and have no problems what so ever. Maybe it's my new lack of Physx that I'm noticing I don't know.
But regardless anyone else notice anything like this?

 

[Mussels]

you get crap like that in the nvidia games until ATI comes along and fixes it.

i suggest waiting for official cat 9.11/9.12 and seeing how things improve (its not like its unplayable as is, is it?)

 

[20mmrain]

you get crap like that in the nvidia games until ATI comes along and fixes it.

i suggest waiting for official cat 9.11/9.12 and seeing how things improve (its not like its unplayable as is, is it?)

Thanx for the input bud. Yeah I also noticed that when I turn off FRAPS or MSI afterburner screen info it skipping and weird lag trips goes away a little more too. But yeah that's what I figured with the whole ATI changing supporting that stuff So I'll hold in there of course 9.11 will be out soon any way.

 

[newfellow]

The various benchmarks i've done over the past two pages do agree somewhat with your statement of no increases with regards to higher memory speeds. However there are always some small gains to be had when the core speed has been increased too. I think that the cards are just not memory bottlenecked at all, and the evidence i've seen just acts to justify such a theory. When i have time i will continue my comprehensive testing by investigating the gains to be had when changing memory speeds with the stock core speed (probably in unigine again).

Yeah, It'll gain when there's data flow as high that it reaches usual RAM or constant data in GPU as long it doesn't swap to RAM. Games sure do this and some larger benchmarks. I kinda was too bold to say on about microsoft style and hell I might even be wrong on there many probably doesn't agree/think the same or see this point of view of how the whole system handles graphical dedicated memory for gamers completely wrong as designed for common functionality. Sad news is gamers are the ones usually with very much RAM, hehe, but system with 2GB DDRx + 1GB GDDRx(or similar split) will be fastest combination money can buy and there's absolutely nothing that can improve that, if there's no 2GB GDDR cards to go along(and this does not count the x2 cards either. There are "dual streamed" not exactly 2GB full. as same goes to crossfire) to 4GB of system RAM.

Unigine, indeed was quite damn nice both tropico and the heaven demos are something else, but I don't think they picture the idea of an unoptimized future titles nor the idea that there using only the GDDRx directly not the RAM to see real performing of the card.

 

[bobzilla2009]

Yeah, It'll gain when there's data flow as high that it reaches usual RAM or constant data in GPU as long it doesn't swap to RAM. Games sure do this and some larger benchmarks. I kinda was too bold to say on about microsoft style and hell I might even be wrong on there many probably doesn't agree/think the same or see this point of view of how the whole system handles graphical dedicated memory for gamers completely wrong as designed for common functionality. Sad news is gamers are the ones usually with very much RAM, hehe, but system with 2GB DDRx + 1GB GDDRx(or similar split) will be fastest combination money can buy and there's absolutely nothing that can improve that, if there's no 2GB GDDR cards to go along(and this does not count the x2 cards either. There are "dual streamed" not exactly 2GB full. as same goes to crossfire) to 4GB of system RAM.

Unigine, indeed was quite damn nice both tropico and the heaven demos are something else, but I don't think they picture the idea of an unoptimized future titles nor the idea that there using only the GDDRx directly not the RAM to see real performing of the card.

It is an interesting thought and that would probably require a change on the part of microsoft. However, since the effect is standardised across all cards people won't notice it as much until we do get 2GB cards. It'll be nice to see how that develops, since i'm pretty sure a 2GB hd5870 is coming out (the eyefinity edition with 6 ports?).

Unigine was not really meant to represent games of the future (although i'm sure some games will utilise it), it was just the higher end benchmark i chose, since most games do not push the gpu to 100%. I prefer using it to crysis because it employs newer rendering techniques. Also i trust the benchmark a bit more than the crysis one, since the standard beach test has quite variable results due to the constant usage of the HDD and processor because of the speed of the movement.

 

[newfellow]

Yeah, probably people won't notice at all before they need some excuse to make more money and actually introduce some fancy name to this, heh. I'm just surprice that high-end folks doesn't see this and there's incredible hype on more and more faster RAM. Which doesn't make any sense really as long it ain't as fast as.

what goes to Uniqine well neither that nor any game actually would utilize 100%. Hell, perhaps something like furmark might be capable of that, but then again we lose the idea of memory usage test as so small smount would be used, but hey, that's better than using half-half RAM/GDDR and think it's fast. Crysis is nice it'll do nice 1000MB of memory while at it on absolute top end +/- 100MB leak, lol.

What goes to HDD well precache figures that out through RAM and "new" RAID-0 in dream case good SSD has really damn nice read speed I just got picture from friend of my last week of RAID-0 Intel SSDs running on Intel Matrix ICH10R (with unofficial new drivers) read speed was 570MB/s with near 0ms access times. I mean damn. CPUs today well as long there's memory controller (even without) we're talking so fast hardware like minimum these days is 4 cores 4Ghz that's 16Ghz of processing speed on any decently made software, almost faster even in video encoding than today graphic cards think it'll manage some processing to keep on phase, heh. good saying would be on GPU processing 'you can't stream everything', but that's why 'where the hell are OpenCL apps' would come in picture.


ps. Sorry for a novel..

 

[bobzilla2009]

Crysis 'only' uses about 700-800MB of video memory on maximum for me. And indeed, no game does use maximum resources on the new cards yet, it'll be nice to see what games will do so, if they're coded somewhat efficiently at least [gta4 uses over 1GB of vram for example, but that's just unoptimised code and/or textures, they can't argue that liberty city should use more vram than crysis, unless they just went OTT with texture details, and tbh crysis's textures on the whole are much better looking, apart from the rocks anyway ]

And ssd's are awesome, i can't wait for the good ones to become mainstream ^^ my computer is entirely held back by the hdd with regards to boot up and loading (even on the WEI score my computer is 7.7-7.9 on everything but the hdd), so it'll be nice when a decent sized one (>320GB) can be had for sub £100.

 

[wolf]

@ bobzilla2009, I have been doing some testing over the past 24 hours and it backs yours exactly.

I've tested in Unigine Heaven, SF4 benchmark and DMC4 benchmark, core always at 950mhz, and memory speed varying from 1000-1300.

Heaven 1680x1050 - 4xAA - 16xAF MAX settings

950/1000 = 31.7fps - score 799
950/1100 = 32.6fps - score 822
950/1200 = 33.6fps - score 845
950/1300 = 34.4fps - score 865

1920x1200 4xAA - 16xAF, MAX - Posterization

950/1000 - 134.64
950/1100 - 139.37
950/1200 - 142.01
950/1300 - 145.44

DMC4 I had to do average of 3 runs across the 4 scenes and experienced some issues, so i wont clog up my post with those results, suffice to say it followed these results quite linearly.

my testing shows;

30% difference in memory bandwidth across the 256 bit bus results in a performance difference of ~9%

I dare make the assumption that even if ATi paired this card with 6.4gbs memory instead of 4.8, we'd see performance of around 10%, given my testing between 4gbs to 5.2gbs

I'd love to speculate how the card would perform with a 512 bit bus, but I honestly don't think I could do it justice. But I really think the choice for 4.8gbs memory was based on how cheap and abundant the memory chips are compared to faster clocked stuff, and the fact that performance on this GPU seems to have little to gain from the speed alone.

thoughts?

 

[bobzilla2009]

To be honest i think the ati engineers knew what they were doing, and the memory seems very balanced to me. In the end, if the core can't push enough data to warrant a 512 bus, why use it? Nvidia only really do it for marketing and small performance gains so they can claim top spot and charge 20% more for it ^^

 

[Zubasa]

To be honest i think the ati engineers knew what they were doing, and the memory seems very balanced to me. In the end, if the core can't push enough data to warrant a 512 bus, why use it? Nvidia only really do it for marketing and small performance gains so they can claim top spot and charge 20% more for it ^^

In fact the "almight" GTX 280 was like over 100% more for less than 20% better performance

Oh my word, $649 vs $299 what a fail nVidia
http://www.techpowerup.com/reviews/Diamond/HD_4870/1.html

 

[Steevo]

Doesn't it always seem that the ones who love their green camp and stuff are the ones always posting later a FS I'm broke thread.

I learned my lesson with my X1800XT and the raping I took on it. I waited untill the 2900 fiasco was over and the 3870's were a good price, then I waited untill the card i wanted was ready for this build, and now I am waiting for a eyefinity and already have a home for my 4850, and for the 3870 I will be taking back.



I believe ATI is suffering with yeilds on the 5870 due to the few people who have them, and the few and far between cards to be found.

 

[Benetanegia]

To be honest i think the ati engineers knew what they were doing, and the memory seems very balanced to me. In the end, if the core can't push enough data to warrant a 512 bus, why use it? Nvidia only really do it for marketing and small performance gains so they can claim top spot and charge 20% more for it ^^

In fact Nvidia uses wider memory bus and higher memory bandwidth as a whole because of GPGPU. There, in general computing memory bandwidth does make a huge difference. On top of that the high clocked memories that are used on desktop GPUs are not really suitable for HPC environment where reliability on 24/7 on small enclosures is a must. Hence bus width has to make up for the lack of ultra high clocks and that taking into account that genral computing needs muchmuch more bandwidth.

 

[grimeleven]

Not to bring back the hot debate "memory bottleneck" but i was reading some architecture documents and come to think about something. Many of you and reviewers noticed a 5-10% increase in FPS when memory was OC'ed.

5870 = 256bits (1200Mhz for 153.60GB/s)
5870 = 256bits (1350Mhz for 172.80GB/s) <--- = 11.25% more GB/s (+- 5-10% perfs)

Now what if it had a 512bit bus wide?

5870 = 512bits? (1200Mhz for 307.20GB/s) <--- = 200% more GB/s
5870 = 512bits? (1350Mhz for 345.60GB/s)

Wouldn't it make sense to expect more than 5-10% increase in FPS? i think yes
I think the issue has more to deal with "memory adressable" than the possible speed it operate.

Anyhow another good page to read.. http://www.beyond3d.com/content/reviews/53/7

 

[Bo_Fox]

Okay I am now convinced that there is something holding this card back and I am now sure it is the drivers!
I just now got to finally game with my new 5870 for the first time. I am using 9.10 drivers and you get some weird results in Frames Per Second especially with Nvidia coded games. For instance I was playing Crysis and Crysis warhead and those games did fine and way better than before(even though it's a Nvidia game) But than I played Far Cry 2 and got some really weird results. Like for most of the time with that game I will be way above 60FPS but for a split second the whole time planing I would drop from 60FPS to 30FPS and it wouldn't even be a stressful part.
Now I do understand that FPS does drop from time to time but I've never seen a card have that big of a fluctuation. But next I would play some ATI coded games like HAWX and have no problems what so ever. Maybe it's my new lack of Physx that I'm noticing I don't know.
But regardless anyone else notice anything like this?

Use D3DOverrider that is bundled with Rivatuner. Download the program, install it, and then find D3DOverrider in the Rivatuner menu in the Start menu in Windows. Tada, you have triple buffering, and the frame rates will no longer hitch/stutter (be halved into 30fps every now and then).

 

[Bo_Fox]

Not to bring back the hot debate "memory bottleneck" but i was reading some architecture documents and come to think about something. Many of you and reviewers noticed a 5-10% increase in FPS when memory was OC'ed.

5870 = 256bits (1200Mhz for 153.60GB/s)
5870 = 256bits (1350Mhz for 172.80GB/s) <--- = 11.25% more GB/s (+- 5-10% perfs)

Now what if it had a 512bit bus wide?

5870 = 512bits? (1200Mhz for 307.20GB/s) <--- = 200% more GB/s
5870 = 512bits? (1350Mhz for 345.60GB/s)

Wouldn't it make sense to expect more than 5-10% increase in FPS? i think yes
I think the issue has more to deal with "memory adressable" than the possible speed it operate.

Anyhow another good page to read.. http://www.beyond3d.com/content/reviews/53/7

Beautiful.

Hiya Steevo, about the latency issue, perhaps you are right that overclocking the memory does automatically increase the latency a notch. I wonder at what point (4.9GHz or 5.1GHz or what)? It is certainly possible that ATI's drivers increase the latency when overclocking, just like automatically increasing the Vgpu voltage when overclocking the core. Extensive testing/benching would need to be done, though.