ATI HD5XXX Series Discussion

[Fishymachine]

Originally Posted by Bo_Fox

Hey W1zzard, what do you think about doing an in-depth article on how the memory bandwidth is bottlenecking the 5870?

Regarding the memory bottleneck for a 5870, whenever I see a dual 4890 in Crossfire beating a 5870, I blame it on the reduced total memory bandwidth for a 5870 compared to two 4890's in total.

A 5870 theoretically should be at least equal to dual 4890's in specifications, save the memory bandwidth. It should NEVER be slower than dual 4890's if it were not bottlenecked by memory, simply due to the fact that crossfire scaling is not 100% efficient.

Once again, if I see a 2x 4890 beating a 5870 by say, 40% at a high resolution, I simply point my fingers at the memory bandwidth of a 5870. It is as logical as it can get.

Yes should see some improvements with 1600(6400)GDDR5 but all benches are done using beta drivers and as for the article's RV770 scaling i remember a 4850 being at 3870X2 levels ,especially with AA (625*10clusters=750*8clusters,and by the way a 850mil transistor 4770 also delivers more than a 2*666)

[Benetanegia]

Originally Posted by Bo_Fox

Hey W1zzard, what do you think about doing an in-depth article on how the memory bandwidth is bottlenecking the 5870?

I second that idea. Wizzard doing it would be perfect but, if a HD5870 owner wants to do a bit of testing it could also give us an idea.

What we have to do is something like this:

- Do some benchmarking at stock. IMO and in order to not make it too extense, Vantage and FarCry2 could be used, although the best would be an extensive testing, of course.. IMO Vantage is good to compare as long as you are within the same architecture. FarCry2 is perfect because it's very sensitive to memory. If there's a bottleneck here we will see it 10x amplified in comparison to other games.

- Overclock the memry to the highest possible. Reapeat benches.

- Leave memory overclocked and downclock the core to 775 Mhz. Bench. Then 700 Mhz, bench. And finally 600 Mhz, bench. 600 Mhz is the control for the other two, since at 600 Mhz it should be the core the slowest thing by far or AMD would have done the worst design in history*.

The idea is that if the memory is bottlenecking, the card would perform the same if we downclocked the card a bit. By benching at various levels we can see the scaling. Downclocking the core should give a reduction in performance comparable to clock reduction, if we don't see a linear scaling there's a bottleneck.

*I don't believe in that posibility, that's why I think the HD5870 is not bottlenecked by the memory. IMO the problem is somewhere else, like in the setup engine, or the dispatch processor. Why? Because improving those requires a lot of R&D, while improving memory bandwidth is easy. IMHO you don't make a 1600 SP beast if you aknowledge it will be bottlenecked by the memory. Memory can't be improved after the card has been released, but you can optimize with game profiles to make up for the "lacking" dispatcher.

[Blacklash]

Hopefully some notebooks will come out with something like a mobile 5850. That would be nice.

[Zubasa]

A mobile HD 5770 will be a good idea, it should be more compact and even more power efficient.

[newfellow]

yeah, the diet GPUs are coming been wondering when will they start marketing 57xx cards as extra crispy lite fat products for bet that marketing array covered.

I can already see the headlines "thin as paper clip, wouldn't you like your PC to feel green too" and primary sales at walmart between the 0% milk & 1% butter.

[Zubasa]

Some Quick Run on 3Dmark Vantage GPU Tests:

First Some using "High" settings:




Custom Setting "Extreme" @1080P and 8xAA 16xAF:






Some Results messing with the Memory Clocks:

"High":


1080P:

[Blacklash]

I'm not sure if this has been posted yet. You can use MSI's Afterburner overclocking tool to boost voltage.

http://event.msi.com/vga/afterburner/download.htm

Then you can use the AMD GPU tool available on this site to push up clocks.

http://www.techpowerup.com/downloads...r_HD_5870.html

Right now the AMD GPU tool doesn't actually change voltage on the 5870.

[inferKNOX]

Originally Posted by Blacklash

Hopefully some notebooks will come out with something like a mobile 5850. That would be nice.

That would be way too much power for such low rezos. It would be equivalent to Asus' silly lappy-melting move of putting an i7 in a laptop; just over the top.

Originally Posted by Zubasa

A mobile HD 5770 will be a good idea, it should be more compact and even more power efficient.

+1, or even a 5750 considering it's great performance. It's arguable that even that is too much though, but I think it'd be good.

[Thrackan]

Just ordered an XFX 5850

[inferKNOX]

I'm getting the XFX 5850 too! We'll be twins, lol!

[Thrackan]

Originally Posted by inferKNOX

I'm getting the XFX 5850 too! We'll be twins, lol!

I'm waiting for a confirmation, but they should be in stock

Website wasn't overloaded so I should be ok.

[Benetanegia]

Originally Posted by Zubasa

Some Quick Run on 3Dmark Vantage GPU Tests:

First Some using "High" settings:
http://img.techpowerup.org/091004/5870 GPU 850-1200.jpg
http://img.techpowerup.org/091004/5870 GPU 750-1200.jpg
http://img.techpowerup.org/091004/5870 GPU 650-1200.jpg

Custom Setting "Extreme" @1080P and 8xAA 16xAF:
http://img.techpowerup.org/091004/Settings.jpg
http://img.techpowerup.org/091004/58...850-1200 X.jpg
http://img.techpowerup.org/091004/58...750-1200 X.jpg
http://img.techpowerup.org/091004/58...650-1200 X.jpg


Some Results messing with the Memory Clocks:

"High":
http://img.techpowerup.org/091004/5870 GPU 850-1000.jpg

1080P:
http://img.techpowerup.org/091004/58...850-1000 X.png
http://img.techpowerup.org/091004/58... 850-900 X.jpg

Thanks a lot. It looks like there's no much bottleneck to me. I'll make a summary so it's easier to find the results and I'll add a score/clock ratio (GPU score/Mhz) to each of them:

Results by core clock. Stock Memory:
High:

850 Mhz -> 8111 GPU points -> Ratio: 9.54
750 Mhz -> 7162 GPU points -> Ratio: 9.55
650 Mhz -> 6421 GPU points -> Ratio: 9.87

Custom Setting "Extreme" @1080P and 8xAA 16xAF:

850 Mhz -> 6427 -> R: 7.56
750 Mhz -> 5705 -> R: 7.6
650 Mhz -> 5025 -> R: 7.73

Results by memory clock. Core clock 850 Mhz:
High:

1200 Mhz (4800 MT) -> 8111 -> Ratio: 6.76
1000 Mhz (4000 MT) -> 7780 -> Ratio: 7.78

Extreme:

1200 Mhz -> 6427 -> Ratio: 5.35
1000 Mhz -> 6260 -> Ratio: 6.26
900 Mhz -> 5985 -> Ratio: 6.65

Conclusions:

- By clock results show a very similar ratio across all GPU clocks, so we can conclude there's no significant memory bottleneck.

- By memory results corroborate the previous conclusion. As we move to lower memory clocks the performance-per-memory clock is much higher and that shows there's no bottleneck. If memory was a bottleneck we should see similar ratios as we moved to lower memory clocks.

At least in 3DMark Vantage, the HD5870 is not bottlenecked by the memory.

Thanks a lot for spending your time doing these benchmarks Zubasa.

[mR Yellow]

I just bought the 5870. They still didn't have stock of the 5850 here and couldn't stand using the onboard gfx any longer.

[inferKNOX]

Originally Posted by Thrackan

I'm waiting for a confirmation, but they should be in stock

Website wasn't overloaded so I should be ok.

I'm having it pre-ordered from Amazon since it's never there!
Meanwhile I'm selling my 4890 and expect to make some extra cash from it.

[Bo_Fox]

Originally Posted by Benetanegia

Thanks a lot. It looks like there's no much bottleneck to me. I'll make a summary so it's easier to find the results and I'll add a score/clock ratio (GPU score/Mhz) to each of them:

Results by core clock. Stock Memory:
High:

850 Mhz -> 8111 GPU points -> Ratio: 9.54
750 Mhz -> 7162 GPU points -> Ratio: 9.55
650 Mhz -> 6421 GPU points -> Ratio: 9.87

Custom Setting "Extreme" @1080P and 8xAA 16xAF:

850 Mhz -> 6427 -> R: 7.56
750 Mhz -> 5705 -> R: 7.6
650 Mhz -> 5025 -> R: 7.73

Results by memory clock. Core clock 850 Mhz:
High:

1200 Mhz (4800 MT) -> 8111 -> Ratio: 6.76
1000 Mhz (4000 MT) -> 7780 -> Ratio: 7.78

Extreme:

1200 Mhz -> 6427 -> Ratio: 5.35
1000 Mhz -> 6260 -> Ratio: 6.26
900 Mhz -> 5985 -> Ratio: 6.65

Conclusions:

- By clock results show a very similar ratio across all GPU clocks, so we can conclude there's no significant memory bottleneck.

- By memory results corroborate the previous conclusion. As we move to lower memory clocks the performance-per-memory clock is much higher and that shows there's no bottleneck. If memory was a bottleneck we should see similar ratios as we moved to lower memory clocks.

At least in 3DMark Vantage, the HD5870 is not bottlenecked by the memory.

Thanks a lot for spending your time doing these benchmarks Zubasa.

3DMark Vantage is probably the least likely program to be bottlenecked by the memory, since it would be bottlenecked by the GPU processing power for the next few generations.

[1Kurgan1]

Originally Posted by Bo_Fox

3DMark Vantage is probably the least likely program to be bottlenecked by the memory, since it would be bottlenecked by the GPU processing power for the next few generations.

Got to agree here, a run of Furmark would probably be the best bet as that completely eliminated the CPU and is free.

http://www.ozone3d.net/benchmarks/fur/

[u2konline]

So i guess in time, they will make non power connector DX11 cards right?

[inferKNOX]

Good question since even the 5750 has a 6-pin connector...

[Blacklash]

MSI's Afterburner is nice. I'm up to 1030|1300 on my HD 5870 so far. That's worth 18.1k GPU using the performance setting in Vantage.

http://img5.imageshack.us/i/van1kgpu.jpg/

[Benetanegia]

Originally Posted by Bo_Fox

3DMark Vantage is probably the least likely program to be bottlenecked by the memory, since it would be bottlenecked by the GPU processing power for the next few generations.

I agree to an extent, that's why I also suggested FarCry2. That's why I especfied "At least in 3DMark Vatage..." too.

Anyway, games are mostly GPU limited too, especially the new ones. Memory overclock yields very little performance improvements and we also have the GTX260 and GTX260 216 to prove that point. Both GTX260 have the same memory bandwidth but the 216 one is significantly faster in Wizzard's performance charts, where they say that the HD5870 is being bottlenecked, and for the most part the performance difference between both 260 flavors matches the added GPU power.

Sure, what is enough for one architecture might not be enough for another, but there's one more "proof" that we can add to the mix: the HD4850 with just half the bandwidth of the HD4870, is a lot faster than half a 4870. All the performance difference can be explained by the clock difference.

Another proof is that if the HD5870 was bottlenecked by the memory, we should see some variance on it's advantage over the other cards as we moved to least demanding settings. The performance difference or relative difference is mantained or scales in the same form that other cards do, so nothing in Wizzard's benchmarks points out to the HD5870 being memory bottlenecked.

IMO "192SP vs 216SP GTX260 in all games" + "HD4870 vs HD4850 in all games" + "Zubasa's HD5870 scaling in Vantage" + "no extrange scaling as more demanding settings are used in Wizzard's review" + "you don't make a chip so big when you know it will be bottlenecked because it's useless" already makes a knowledge base big enough to make me think that it's not bottlenecked.

Of course I would appreciate some game testing, but it's probably asking too much, so I play with what I have.

[inferKNOX]

Heads up, the:
GIGABYTE GV-R587D5-1GD-B Radeon HD 5870 (Cypress XT) 1GB 256-bit GDDR5 PCI Express 2.0 x16 HDCP Ready CrossFire Supported Video Card w/ATI Eyefinity - Retail
is available at NewEgg at the moment for anyone that wants one (and it comes with a free DiRT 2).
Wow, these 5 series cards are rarer than chickens teeth!

[jagd]

5750 and 5770 started to seen in lists at Europe €119 for 5750 and €149 for 5770

http://www.fudzilla.com/content/view/15838/1/

[Thrackan]

That XFX 5850 blocks some of my SATA ports, gonna see tonight if I can fit in a hooked connector.

One heck of a card though, feels a lot heavier than my 4850. Ran a quickie on Vantage when I got home late last night (1:30) and I didn't even notice the score. Did run it a lot more smoothly though.

Gonna try a 3DMark06 run as well, since I remember what scores I had with my 4850 on that.

[kylzer]

Hello =D

[Thrackan]

Originally Posted by kylzer

Hello =D

http://i38.tinypic.com/f3uy4j.jpg

http://i34.tinypic.com/34oq1py.jpg

I see you finally went "expensive"?

I already love my 5850 and I've hardly let in run yet