HD 5850 overclocking gains in the Heaven 2.1 benchmark: avg up 30% for 38% clock bump

[dug777]

I used the latest heaven 2.1 benchmark tool, with everything set as it defaulted in the heaven options menu (I noticed tesselation was 'normal') and 1900x1200 screen resolution.

Specs of computer are : Q6600@3.2Ghz/Asus P5Q PRO/4GB DDR2/XFX 5850 BE.

First run was done at stock 5850 speeds, so 725/1000:

Avg fps: 28.4

Second run was done at 1000/1225:

Avg fps: 37.0

Not too shabby at all, that's a 30.3% increase for a clock bump of 37.9% and a memory bump of 22.5% if I have done my maths correctly

Having done that, it occurs to me that it would be interesting to repeat those with tesselation cranked right up and see if scaling is similar. Those will be up shortly

EDIT: same settings as above, but 'extreme' tesselation selected, and the scaling has improved!

First run was done at stock 5850 speeds, so 725/1000:

Avg fps: 17.3

Second run was done at 1000/1225:

Avg fps: 22.9

Not too shabby at all, for the extreme tesselation that's a 32.4% increase for a clock bump of 37.9% and a memory bump of 22.5% if I have done my maths correctly

 

[Qbah]

Great results, dug! Keep the numbers coming

 

[AndroidVageta]

It is actually pretty impressive to see how well the FPS improve with the overclocking...I could easily see a re-worked 5870 core made into a 5890 that could be the fastest single GPU graphics card.

Either case, cool numbers Dug, Ive never seen overclocking performance bumps like that!

 

[Lonyo]

Any chance of doing just core and just mem? Be interesting to see if the mem has any impact at all on its own (I would assume not particularly).

 

[MagickMan]

it impacts performance a few % (1000 > 1225), but not nearly as much as core clock.

 

[Cookie Monster]

This should be expected since the tesselator found on cypress runs at the same core clock as the rest of the chip.

 

[happy medium]

Your scores are 1 fps faster then a 5870. nice!

This benchmark shows 5770's in crossfire beat up on a 5870 in this test.
I wonder why?

Of course the gtx series mops the floor in this test.

http://www.geeks3d.com/20100527/tes...ion-gtx-480-vs-gtx-470-vs-hd-5870-vs-hd-5770/

 

[dug777]

Your scores are 1 fps faster then a 5870. nice!

My maths says 22.9-21.1 = 1.8 FPS

Plus I'm running 1920x1200

 

[dug777]

Any chance of doing just core and just mem? Be interesting to see if the mem has any impact at all on its own (I would assume not particularly).

As Magick Man says, it was pretty limited on its own, although it was still significant (I think it might have gone up 1 fps or less for the normal tesselation runs I did, but I didn't write them down I am afraid).

 

[Qbah]

Downloading Heaven 2.1 now. My HD5850 is running at 850/1150 stock voltage (card came with 765/1125). Tried 880 but it just wouldn't happen. Since it's a Toxic, I can't change volts, but no biggie, the card is mighty fast anyway Did a few Crysis Warhead tests and I'm getting 30AVG with 20+MIN on the snow benchmark levels - that's all Enthusiast AAx4 1920x1080 The game is very smooth! CPU is humming at 3.4GHz, stock volts too I can finally enjoy Crysis ! My old HD4870 was struggling at 1680x1050 no AA and Gamer / High...

The card itself is inaudible at idle which I absolutely love. Hits 70C tops at full load with the OC. However the fan ramps up a bit and I can hear it a bit. But compared to my Xbox it's like a whisper in my ear (Xbox is a jet that's taking off D.

I'll post my numbers once Heaven lands on my HDD

EDIT: So I got 32.8FPS AVG (826 points). All default settings. 1920x1080.

 

[Sylvanas]

Downloading Heaven 2.1 now. My HD5850 is running at 850/1150 stock voltage (card came with 765/1125). Tried 880 but it just wouldn't happen. Since it's a Toxic, I can't change volts, but no biggie, the card is mighty fast anyway Did a few Crysis Warhead tests and I'm getting 30AVG with 20+MIN on the snow benchmark levels - that's all Enthusiast AAx4 1920x1080 The game is very smooth! CPU is humming at 3.4GHz, stock volts too I can finally enjoy Crysis ! My old HD4870 was struggling at 1680x1050 no AA and Gamer / High...

The card itself is inaudible at idle which I absolutely love. Hits 70C tops at full load with the OC. However the fan ramps up a bit and I can hear it a bit. But compared to my Xbox it's like a whisper in my ear (Xbox is a jet that's taking off D.

I'll post my numbers once Heaven lands on my HDD

EDIT: So I got 32.8FPS AVG (826 points). All default settings. 1920x1080.

Good results! Try Heaven 2.1 in OpenGL 4.0 mode with Tess enabled with the 10.6 drivers. See what the performance disparity is between DX and OGL since 10.6 seems to have brought some improvements to GL.

 

[dug777]

Downloading Heaven 2.1 now. My HD5850 is running at 850/1150 stock voltage (card came with 765/1125). Tried 880 but it just wouldn't happen. Since it's a Toxic, I can't change volts, but no biggie, the card is mighty fast anyway Did a few Crysis Warhead tests and I'm getting 30AVG with 20+MIN on the snow benchmark levels - that's all Enthusiast AAx4 1920x1080 The game is very smooth! CPU is humming at 3.4GHz, stock volts too I can finally enjoy Crysis ! My old HD4870 was struggling at 1680x1050 no AA and Gamer / High...

The card itself is inaudible at idle which I absolutely love. Hits 70C tops at full load with the OC. However the fan ramps up a bit and I can hear it a bit. But compared to my Xbox it's like a whisper in my ear (Xbox is a jet that's taking off D.

I'll post my numbers once Heaven lands on my HDD

EDIT: So I got 32.8FPS AVG (826 points). All default settings. 1920x1080.

Sounds solid mate, and that's a solid speed bump for stock voltage

 

[Qbah]

So I ran Heaven in OpenGL and... ouch. Same settings, just a different renderer and I got 20.0 FPS (505 points). Now here's the weird part:

MSI Afterburner log for DX11:

MSI Afterburner log for OpenGL:

What a difference in GPU load, eh? The two "spikes" for OpenGL were for Scene 8 and Scene 22. The common thing for both scenes? A distant shot of the floating islands. So in OpenGL the GPU was loaded around 60-70% on average. In DX11 mode the GPU was loaded to the max the whole time.

Another thing is CPU load. Don't have any screencaps for those, but in DX11 all 4 cores were loaded at ~10-15%. In OpenGL one core was loaded at like 40-60%. The other 3 were pretty much idle.

 

[Scali]

What a difference in GPU load, eh? The two "spikes" for OpenGL were for Scene 8 and Scene 22. The common thing for both scenes? A distant shot of the floating islands. So in OpenGL the GPU was loaded around 60-70% on average. In DX11 mode the GPU was loaded to the max the whole time.

Another thing is CPU load. Don't have any screencaps for those, but in DX11 all 4 cores were loaded at ~10-15%. In OpenGL one core was loaded at like 40-60%. The other 3 were pretty much idle.

Those two generally go together. Apparently in DX11 you are completely GPU-limited.
In OpenGL there is considerably more CPU overhead, and as a result the GPUs are occassionally waiting on the CPU to send them new commands.

DX11 also has some opportunities for multithreading, where OpenGL is pretty much a single-threaded API, like DX10 and earlier versions.
So in general it's not that strange to only see one core loaded with a graphics benchmark.
In Heaven there is virtually no interaction, no AI, no physics, animation or anything that the other cores could be processing, unlike in games. So it's pretty much all just sending OpenGL commands to the hardware from a single thread, nothing to do for the other cores.

One of Microsoft's design goals with DX10 (and DX11, which is based on that same design) was to reduce the CPU load, and make the API and driver as efficient as possible. These results indicate that they've done a good job at that.

Having said that, nVidia's OpenGL drivers appear to be quite a bit more efficient, and the gap with DX11 isn't that large.

 

[Qbah]

Those two generally go together. Apparently in DX11 you are completely GPU-limited.
In OpenGL there is considerably more CPU overhead, and as a result the GPUs are occassionally waiting on the CPU to send them new commands.

DX11 also has some opportunities for multithreading, where OpenGL is pretty much a single-threaded API, like DX10 and earlier versions.
So in general it's not that strange to only see one core loaded with a graphics benchmark.
In Heaven there is virtually no interaction, no AI, no physics, animation or anything that the other cores could be processing, unlike in games. So it's pretty much all just sending OpenGL commands to the hardware from a single thread, nothing to do for the other cores.

One of Microsoft's design goals with DX10 (and DX11, which is based on that same design) was to reduce the CPU load, and make the API and driver as efficient as possible. These results indicate that they've done a good job at that.

Having said that, nVidia's OpenGL drivers appear to be quite a bit more efficient, and the gap with DX11 isn't that large.

Well, since neither the GPU nor CPU (even one core) were loaded to the max, it's just a bad ATi's OpenGL driver? In DX11 it is obvious my GPU was working as best as possible. But what's stopping it to do the same in OpenGL? The driver or API only I would assume, as nothing else changed.

Let's hope ATi sorts those things out before there's anything I'd like to play and it's in OpenGL only (like iD Software's Rage).

 

[Scali]

Well, since neither the GPU nor CPU (even one core) were loaded to the max, it's just a bad ATi's OpenGL driver?

Well, it's a single-threaded serial mode of operation...
So you get this cycle:
CPU sends commands to GPU -> CPU waits for GPU to finish -> GPU finishes -> GPU waits for CPU to send commands -> CPU sends commands to GPU -> etc.

(in between, there is some queuing going on, commands from the application can be preprocessed for the GPU and queued to be sent when the GPU is ready to receive them).

So they are constantly waiting for eachother, and as a result, they will rarely both reach 100% usage at the same time... Apparently in DX11 you are able to reach about 100% usage of the GPU because the CPU overhead for sending the commands is negligible, and the GPU barely has to wait.
The commands themselves are the same for the GPU, it doesn't 'know' whether it's doing D3D or OGL. So it must be the OpenGL driver that just takes more time preparing these commands... and since nVidia seems to do a better job at it, it's probably a poor implementation from ATi.

I have been working on a small OpenGL project myself (http://sourceforge.net/projects/bhmfileformat/), and I found pretty much the same thing... nVidia cards just work faster in OpenGL... I found another thing: the difference between ATi and nVidia seems quite a bit larger in Vista/Windows 7 than in XP.
So it could be interesting to run the Heaven benchmark in XP and see if it improves OpenGL performance.

 

[Kenmitch]

Any improvements with 10.6 for you?

I loaded them last night and 3dmark Vantage gained about 110pts. for the gpu. Heaven 2.1 benchmark looked alot smoother. Didn't compare my results yet tho. Gonna try and play around tonight and do some more testing if time permits.

 

[dug777]

Any improvements with 10.6 for you?

I loaded them last night and 3dmark Vantage gained about 110pts. for the gpu. Heaven 2.1 benchmark looked alot smoother. Didn't compare my results yet tho. Gonna try and play around tonight and do some more testing if time permits.

With extreme tesselation and at the same overclocked settings my average was exactly the same but my minimum improved by 9.2%

 

[Kenmitch]

With extreme tesselation and at the same overclocked settings my average was exactly the same but my minimum improved by 9.2%

So it looked alot smoother also to you?

With the 10.5's it kinda was a little jerky at times for me.

I only ran this and vantage last night. But did notice both to be alot smoother looking visually.

 

[Sylvanas]

So I ran Heaven in OpenGL and... ouch. Same settings, just a different renderer and I got 20.0 FPS (505 points). Now here's the weird part:


What a difference in GPU load, eh? The two "spikes" for OpenGL were for Scene 8 and Scene 22. The common thing for both scenes? A distant shot of the floating islands. So in OpenGL the GPU was loaded around 60-70% on average. In DX11 mode the GPU was loaded to the max the whole time.

Another thing is CPU load. Don't have any screencaps for those, but in DX11 all 4 cores were loaded at ~10-15%. In OpenGL one core was loaded at like 40-60%. The other 3 were pretty much idle.

Nice results Qbah, it puts it as plain as day the differences between OpenGL and DX these days- Dx11 looks to be clearly the overwhelming API of choice for developers. Whether it's inefficiencies in the API itself or drivers (probably a bit of both) OGL is still playing catch up. Perhaps if some games based on id's tech engine 5 show up we may see more more action on the OpenGL driver front, but it's pretty barren at the moment.

 

[dug777]

So it looked alot smoother also to you?

With the 10.5's it kinda was a little jerky at times for me.

I only ran this and vantage last night. But did notice both to be alot smoother looking visually.

Yes, with the 10.6s it didn't do the very noticeable hitch it had done with the 10.5s a second or so after you loaded up the program and the splash screen disappeared (as you started down that cobblestone street with the dragon courtyard on your right).

 

[dug777]

Nice results Qbah, it puts it as plain as day the differences between OpenGL and DX these days- Dx11 looks to be clearly the overwhelming API of choice for developers. Whether it's inefficiencies in the API itself or drivers (probably a bit of both) OGL is still playing catch up. Perhaps if some games based on id's tech engine 5 show up we may see more more action on the OpenGL driver front, but it's pretty barren at the moment.

At 19x12 and extreme tesselation I get 22.9 avg for DX11 and 17.1 avg for OpenGL, a whopping 34% hit for OpenGL. In the latter benchmark my GPU usage peaked at 75%

What I find interesting is that compared with happt medium's little link earlier, I give the stock 5870 a handy kicking in DX11, but utterly obliterate it in OpenGL (on both counts especially when you consider I am running 19x12). I presume that suggests that the drivers aren't as good as divvying up the workload to take advantage of the extra SP units in the 5870 and/or the API itself doesn't lend itself as easily to that process, although it's a comparison that needs to be treated with some care as they were done with the 10.5s and I used the 10.6s.

 

[Qbah]

In OpenGL I'm getting roughly 60% of DX11's performance. Which is pretty much in line with my average GPU usage. So if the card was used fully in OpenGL, I'm sure the performance would be similar.

Also, remember Anand's Steam on Mac vs PC performance tests? The Mac is running at around 60% of Windows performance (OGL vs DX). Looks familiar? If there's a tool under OS X to measure GPU load, this could be easily proved (or it would at least be a supportive argument). Anyone here with a Mac wants to volunteer? It looks like utilization is similar under Windows, regardless of the speed of ones GPU.

EDIT: Once I'm back home I'll run some tests in Source. I think there's a way to force OpenGL in Windows. Perhaps it's only Heaven that acts weird in OpenGL?

 

[Scali]

EDIT: Once I'm back home I'll run some tests in Source. I think there's a way to force OpenGL in Windows. Perhaps it's only Heaven that acts weird in OpenGL?

Nah, it's not just Heaven.
This blog also indicated performance problems in OpenGL, from one of the Unity3D engine developers:
http://aras-p.info/blog/2007/09/23/is-opengl-really-faster-than-d3d9/

I think it's a bit ironic that the anti-MS brigade has been feeding us for years with this nonsense that OpenGL was as good an API, or even better than D3D... but once you finally get a chance to compare the two, OpenGL falls flat on its face.
I've dissected a recent popular pro-OpenGL FUD blog a while ago: http://scalibq.spaces.live.com/blog/cns!663AD9A4F9CB0661!287.entry

Oh, and no, you can't use OpenGL in the Windows version of Source.
People seem to confuse it with the original HalfLife, where there were three renderers implemented: D3D, OGL and software.

 

[Scali]

I found another thing: the difference between ATi and nVidia seems quite a bit larger in Vista/Windows 7 than in XP.
So it could be interesting to run the Heaven benchmark in XP and see if it improves OpenGL performance.

Seems that I was onto something here.
I just ran it in Windows XP and Windows 7 (same machine, same settings, with a Radeon 5770, Cat 10.6)...
Windows 7 DX11: 28.1 fps
Windows 7 OpenGL: 18.2 fps
Windows XP OpenGL: 25.5 fps

So although DX11 is the fastest of the three... it looks like there's some kind of performance issue specific to Windows 7 (and probably also Vista) in ATi's OpenGL driver. It looks like nVidia doesn't suffer from this problem, that would explain why their OpenGL and DX11 results are quite close together, much like my numbers here in XP vs DX11.