Benchmark CaveatsI have said this before, and I will undoubtedly repeat myself in the future: The Mac platform has incredibly lousy benchmarking tools. There is no objective way to truly test all the features of the 8500 versus other cards, and there is no avoiding the fact that CPU speed is the determining factor in many tests. On the PC side they avoid this problem by 1) using specialized benchmarking apps in addition to games and 2) using astoundingly fast systems that have been aggressively tweaked for gaming performance.
Sadly, until dual 1 GHz Mac systems grow on trees, IMG does not have access to that level of hardware. Coupled with the fact that no current (or indeed future) Mac OS games use any of the special features of the Radeon series, these benchmarks present a totally incomplete picture of the 8500.
As with the original Radeon, the 8500 is largely a CPU-bound card. At resolutions under 1280x1024 the top speed of a given game is determined by the CPU power of the machine, not by internal limitations. Thus the results you see here in my performance tests may vary radically from your own experience with a faster or slower machine. I have adjusted many of my usual benchmarks in an attempt to limit the influence of CPU speed on the numbers, and I think it makes the current advantages (and limits) of the 8500 fairly clear.
Quake 3 Arena 1.30This tired, outdated 1999 engine is rapidly losing favor as a PC benchmark – after all, when they can run the game at 180 fps at 1600x1200, where else is there to go? The game uses absolutely no second- or third-generation card features, and does its lighting and geometry tessellation in software (no hardware assist). In short, it is outdated and doesn’t even make the 8500 break a sweat – but hey, you use what you have available.
I’ll do the OS 9 numbers first. I used Quake 3 Arena version 1.30 (the last non-beta version) with the built-in "four" demo. I threw away the .cfg file to start fresh, but I also went to Game Options and turned OFF Marks On Walls, Dynamic Lighting and shell casings – these effects are rendered totally by the CPU (not the video card) and skew the results by 8-10 fps. Unfortunately, the 1.30 version of Q3A has a bug in the OS 9 build which seems to limit the top resolution to 1152x870 While the chart says 1280x1024, the max rez is actually 1152x870; if you choose a higher setting, the game will act as if the change has indeed taken effect, but bringing down the console shows that the screen resolution has not changed. 1280x1024 and 1600x1200 are not being tested at all, which basically renders the benchmarks completely useless when evaluating the 8500's power. C'est la vie!
The first test begins with the Normal default setting, increasing only the screen resolution for each test. As you can see, the 8500 doesn’t even blink at high resolution, with less than 3 fps difference between the 640x480 benchmark and the max benchmark.
The GeForce 2 MX initially leaps ahead, which you might find slightly surprising -- after all, this is a much weaker piece of hardware. However the NVIDIA drivers have been very well optimized for Quake 3 Arena, and the results don't surprise me at all. At low resolutions, the GFMX benefits tremendously from its own simplicity -- the fact that it only has two texture pipelines and a single texture unit mean that the path through the GPU pipeline is very simple, and there are no "states" (instances when data has to wait for a system on the card to purge or be ready to accept new data) slowing things down.
In contrast, the complex pathways of the 8500 are like a maze of tunnels full of gated paths. The added complexity allows for many complex operations to occur in parallel, but Quake 3 Arena doesn't test that complexity in the slightest -- at most two textures are being drawn at once and lit by a single source of light. Thus the complexity can act like a roadblock at low resolutions, as data is forced to wait for multiple states. But who really cares about low resolutions?
On High Quality settings, the numbers begin to get interesting. Just for kicks, I set the texture and geometry settings to maximum (fourth click) to really test that card memory bandwidth. As you will see, the GeForce 2 MX dies off quickly, and even the Radeon begins to crumble as the resolution climbs. I was a little surprised at the size of the frame rate drop between 1024 and 1280, but overall these are excellent results for the 8500. With some tweaking I was able to play Q3A at high rez with nine bots and maintain an average frame rate well over 50 fps on any map – let me tell you, that game has never looked better. For a 733 Mhz machine with no L3 cache, I feel those are excellent results – and I bet the drivers have plenty of room for optimization.