Exhaustive Performance Anaylasis of ATI 4800 platform

[Truckchase!]

As some guys on this board know, I've developed an automated testing suite of games, and I've taken to benching on ATI's latest offering. My focus will be on the following:

- overall performance
- Performance hit for anti aliasing
- Performance hit for anisotropic filtering
- architecture bottleneck analysis (Gpu vs. mem)

My first test subject is a HIS 4850 with stock cooling. For the sake of comparing to ATI's last gen, I used a heavily overclocked 2900xt. This should be relevant enough to be interesting vs. the 3870 as well.

For anyone who's interested in the raw data, please inquire and I'll send it to you if interested. It's just to much to post here. I have run my full test suite @ the following resolutions/aa/aniso: 1600x1200@ 4/16, 4/0, 0/16, 0/0, 1024x768@ 4/16. The games include Quake2,Quake3,UT2k4FlyBy,UT2k4BotMatch,Comanche4,F arCry,Half Life 2,Serious Sam 2,Doom 3,Half Life 2 Lost Coast,Quake 4,Splinter Cell Chaos Theory,Supreme Commander,Company of Heroes DX9,Company of Heroes DX10,Crysis DX9 High,Crysis DX10 High,Crysis DX10 Very High. Crysis results are not complete due to test reliability issues. (thanks to copy protection bombing) Cards tested sofar are 2900xt@918/918 and 4850@625/993, 700/993, 625/1115, 700/1115. All tests run on all cards at all resolutions. The underlying system is entirely consistent for all runs. (See sig) Let me know if you would like any of that data presented in any special way and I'll format it as such.

First, overclocked results of the overclocked 2900xt vs. the overclocked 4850.

NOTE: This first test isn't fair. The 2900xt is overclocked (the core at least, which is the restricted part on this card) by 24%, while the 4850 by only 12%. This just gives a good general idea of what the new 4800 series is capable of. (which as we all know is quite a bit)


The reason this data should be interesting to most is the drastic performance impact of Anti aliasing (I'll directly statistically correlate this in a post to follow) on the previous generation of cards vs. this generation. The 4800 series handles AA much better.

More to follow....

[Truckchase!]

In this post I'll examine the potential for a memory bottleneck on the 4800 series. As I stated earlier:

here's a different take on how to tell what impact mem bandwidth is having...

1> benchmark 4850 @ stock speeds
2> oc the GPU and vMem by a static percentage (15%) and bench
3> OC just the GPU and Just the vMem by the same static percentage and benchmark each oc seperately.

Compare the results to see how large of an impact the vMem overclocking has on performance vs. the gpu. In all actuality, you need only to do step 1 and 3, but 2 is always fun to compare against to make sure there isn't an odd unknown variable.

Here's an example from when I did this on a 2900xt:
http://www.xtremesystems.org/forums/showpost.php?p=2381435&postcount=31


I've done this at 1600x1200 with AA on/of and Aniso on/off. Here are the results from the two extremes on a 4850:

Note that the 4850 is the more effective card of the two to check for memory bandwidth bottlenecks in this fashion because of the significantly lower overall bandwidth. That being said, I'll still test the 4870 later. :D

Quick note on the graphs... stock speeds are 100%, this shows the performance relative to that for ONLY gpu overclocked and ONLY mem overclocked. All overclocking for this test was by 12%.

This platform definitely looks to be more bandwidth restricted than the 2900, which is no surprise being that the 2900 had a full 512bit interface. I unfortunately don't have 3870 numbers to compare to, but I suspect that the results would be similar to that of the 2900 since the overall performance characteristics are very close. There is certainly potential justification here for GDDR5, as the bottlenecks on some games are close to being evenly split across the memory and gpu.

More analysis to come on these numbers...

[Truckchase!]

Here I'll look in further detail on the impact to framerates by AA and ANISO vs. the previous generation. This new generation handles AA MUCH better, as outlined in the graphs below....

Graphs updated 7/1 due to remove Splinter Cell due to invalid AA data and for clarity

[Truckchase!]

Edit: Added 4850 vs. 4870 info

For this test, I ran the 50 vs. the 70 at the same clocks. (700/995) Obviously the effective clock of the 70 is double that of the 50 due to GDDR5 vs. GDDR3. The first graph I'm posting is comparing the two with 4xAA and 16x Aniso enabled. The results aren't as dramatic with those features disabled. (I'll show that later) Too tired now.... going to bed...

[mAJORD]

Excellent work. Ati really did an awesome job fixing AA performance. Couldn't ask for more there.

Obviously a comparison to a GTX260 would be great, but if you don't have one it aint gonna happen.. someone donate the man one ;)

[keiths]

Graphs are hard to follow, axis need labels. Are the first one's all FPS for y? % of performance vs noaa/ansio is awkward and nonsensical. "X aa/ansio vs no aa/ansio" and including "100%=no aa/ansio" would be clearer. Hats off for the exhaustive testing, though this reinforces my opinion that showing the actual results, the more specific the better, is better.

[Truckchase!]

Graphs are hard to follow, axis need labels. Are the first one's all FPS for y? % of performance vs noaa/ansio is awkward and nonsensical. "X aa/ansio vs no aa/ansio" and including "100%=no aa/ansio" would be clearer. Hats off for the exhaustive testing, though this reinforces my opinion that showing the actual results, the more specific the better, is better.

The third set of graphs has been updated with better titles, and I also removed Splinter Cell because it was not applicable for that test. I'll update the first set after I finish tweaking the Crysis benchmarking routine. Unfortunately all my Crysis data using aniso was not valid because it seems the command line and/or config file aniso options have no effect. I've changed the code to set aniso via the driver for that game, and I'll upload new graphs once I have the updated results. (with the fps label for y)

Thanks for the feedback!

[FlawleZ]

Good work! Great to see a thorough testing especially the AA impact since this was such an issue with the previous two generations. However, I do think you could have made the graphs a bit easier to read.

Overall great job though! Thanks for posting!

[alexio]

Good stuff. Thanks :)

Suggestion: Put it up agains the HD4870 when you get it and play a lot with different clocks.

[Truckchase!]

Good work! Great to see a thorough testing especially the AA impact since this was such an issue with the previous two generations. However, I do think you could have made the graphs a bit easier to read.

Overall great job though! Thanks for posting!

You're right... this is something I need to improve on. I present data like this for my career as well and I've always struggled to present it in a fashion that is readable to someone who didn't collect it in the first place. I'll add to and update the graphs after the weekend.

Good stuff. Thanks :)

Suggestion: Put it up agains the HD4870 when you get it and play a lot with different clocks.

Working on it now. :D

[Truckchase!]

Added another analysis set, see 4th post.

[biohead]

basically 4870 is roughly 5-10% faster than 4850 at those clocks. but i've seen some other comparisons as well with a few heavier games and higher core clocks on both cards. then the gap broadens. bandwidth. just bandwidth.

[ymer]

yea gddr5 really makes a difference

[aldamon]

What about minimum FPS?

I see the GDDR5 makes a small difference in average FPS, but is the difference in minimum FPS large enough to go from unplayable to playable under certain conditions?

[binormalkilla]

Great review man....hats off to you. I love the in-depth analysis of AA/AF, as well as the comparison with the 4870.

[L7R]

Thank you for comprehensive review. Just great.

[Burn]

First off, Truckchase, thank you for doing this comparison. It has helped me immensely. Second off, could you possibly do a test in both overclocked cards? If I were to buy a 4870, I certainly wouldn't keep it at stock settings. Thus, a comparison between 2 overclocked cards IMO would provide the best knowledge about how they perform head-to-head. What do you think?