In this article is a look at the state of the open-source Nouveau Gallium3D driver on low-end NVIDIA GeForce graphics hardware. In particular, a $10 USD NVIDIA retail graphics card is being tested under Ubuntu Linux on both Nouveau and the proprietary NVIDIA driver and is then compared to a wide range of other low and mid-range offerings from NVIDIA's GeForce and AMD's Radeon graphics card line-up with a plethora of OpenGL benchmarks.
So... at what clock was the card running on mesa anyway?
It looked like dmesg was reporting the correct clock speed, or at least the same as the binary blob, unless I was reading it wrong... But I'm guessing that the Gallium driver wasn't using floating point textures or S3TC, and there are probably plenty of other optimizations/extensions that could explain the difference in performance.
So nexuiz is broken, yes ? Well, WORKSFORME, so unless you file a bug report it's NOT going to get fixed.
As for the clock speed of the 9600 GSO, yes, you're reading it wrong:
[ 15.202433] [drm] nouveau 0000:03:00.0: 3: core 500MHz shader 1250MHz memory 900MHz voltage 1000mV fanspeed 100%
[ 15.202446] [drm] nouveau 0000:03:00.0: c: core 399MHz shader 810MHz memory 399MHz voltage 950mV
Doesn't look like anyone had mentioned this yet, so I registered so I could point it out. This card is not a 96 shader 9600GSO, it's the later 48 core variety. The stock core is 650 and memory is 900. The memory buss should be 256 bits (it's a cut down G94a/b die), but it looks like this model was further cost reduced by only populating half of the buss with higher density (currently cheaper) parts. A modern comparison for this card might be a GT520 which has same # of cores, but faster ones, though it has half yet the memory width and bandwidth (sharing the 900MHz memory clock with this part). It does have fewer ROPs and TMUs. The 9600GSO 512 has a 48:24:16 core:TMU:ROP layout while the GT520 has a 48:8:4 layout. The GT520 may have more highly featured/optimized TMUs and ROPs, though--plus other improvements made in the (almost) three years between their releases.
I started reading this article at work. The graphs on pages 6-10 look awful in Internet Explorer 8 on Windows XP Professional. The names of the cards run into each other and they also obliterate the beginning of the standard error (SE) labels. It looked so awful that I gave up. I then tried to read this article on my mobile phone during the bus trip home, but the same poor display of the graphs occurs on Opera Mobile 10.00 on Windows Mobile 6.1 Professional. I gave up again. When I got home I tried again with Firefox 9.0.1 on Ubuntu 10.04 and the graphs looked much better. I then tried using Firefox's zoom in feature and the graphs still looked just as good.