Tests conducted

Altivec Fractal Demo IP: Start application, draw first picture, then refresh once and get the score. This test shows how significant the difference between a G3 and a G4 can be, when the G4 uses Altivec (Velocity Engine). Although not actually a benchmarking application, I consider it a rather theoretical value.( this is the non-carbon version)

CineBench 2000: Shading the test scenes (polygons, wireframe and textures) using the Cinema4D engine and Raytracing (Rendering) the test image in single CPU mode. Since CineBench is based on the Cinema4D 3D application, this test is pretty much a possible real-world situation. I didn't test the OpenGL shading, as the Voodoo3 doesn't work in this mode. This test is unfortunately also graphics card dependent, which turns out to show the Beige G3 perform slightly worse than it would with an ATI card. Accelerated for SMP and Altivec.

Gauge Pro: was used to show the impact of a new CPU on the memory performance. This is a rather theoretical result (as you'll rarely have to know what RAM performance you get), but the more the better, that's for sure.

TechTool Pro 2.5.1: Doing the Whetstone, Dhrystone, Cycle, Sieve and Sort tests. These are all specific CPU task of the IPU and FPU units, I don't think this application supports Altivec in this rather dated version I have. Totally theoretical stuff.

iTunes 1.0: Converting a 87.7MB AIFF file off the harddrive (8 min 31 sec of music) to an MP3 file using 96kb/s mono and 192kb/s normal stereo stream settings (highest quality). This is absolutely a real-world situation, as you surely know and is also accelerated for Altivec. Harddrive dependencies can not be ruled out here, but they are small.

SoundJAM MP 2.5.2: Converting the same file as above, same settings too, but also using Digital Vibrance Control on highest quality. Accellerated for SMP and Altivec and aparently the core technology on which iTunes itself is based. Still, it's a bit faster.

Stuffit Expander 5.5: Again, not the most up to date version, but since this test is also harddrive dependent, it's not very comparable anyway. Unstuffing a 2.6MB .sit.bin file containing 483 small files. (G-News archive) Tested with a stopwatch, so there's an additional source of mistake.

Quake 3 Arena 1.27h: Using both the default .cfg file with normal settings @ 640x480, 16bit and my custom .cfg I use for actually being able to play it. Here the Rage 128 quickly hit a wall and stopped scaling together with the CPU, which is sad. The results using the custom .cfg are more informative there. Used the UCGuide.pk3 containing the converted demo001 from version 117. For more details, download the .cfg file I used and see for yourself. I didn't want to test with Locki's .cfg, although that would nicely relieve the graphics card factor, but then again removes the real-world experience, as Locki is really ugly (the game using his cfg, not the guy). This is real-world. Quake 3 is not altivec enhanced so far, but OpenGL makes certain use of it.

Quake 2 1.0.3: using the latest version here, running demo1.dm2 with normal settings @ 640x480. I thought this wouldn't be very demanding for a graphics card, but the Rage 128 hit a big wall there, don't ask me why. I'll also include the scores for the crusher demo, as the CPU is more demanded there. This is real-world. This is again using OpenGL.

Unreal 2.2.4b6: Doing the castle flyby 3 times and taking the average (as it tends to get worse). Minimum and maximum scores included as well. Settings were highest quality 640x480 with all extra options turned on (all that improve image quality). Running in Rave and Glide respectively. Again very dissapointing with the Rage (what did I do wrong?).

Unreal Tournament 436: Taking the average of one cityintro demo, running Rave and Glide mode respectivey. Graphics mode was 640x480 both texture details medium, minimum desired framerate set to 45 FPS. Not a very good test for this game, real-world performance is better.

MacBench 5.0: Running the CPU and FPU tests. Not optimized for G4.
This again is pretty much theoretical stuff.

Photoshop 5.5: Yeah I sort of had to do this, right? I'm telling you, it's a pain in the ass. I had quite some trouble getting this to work properly. Running a test called PS5Bench Basic, which generates a 10MB file and applies 21 filters to it. Some of these are optimized for Altivec, some are not. The bad thing is that if you run the test step by step without break, the application is forced to its knees and ends up being horribly slow. Thus going back to the finder after each test and letting the OS recover a bit is a good thing to do. Each filter is applied 3 times, but I only took the fastest time, since each run can vary b up to 200% from another. Doing these test convinced me that testing in Photoshop is actually a pretty worthless thing to do, as it heavily depends on how you do it etc. Photoshop was given 200MB of RAM. Altivec plug-ins were installed of course. Again this can be true to real-world, but unfortunately doesn't have to be, as you'll see in the detailed chart.

Last but not least I also tried the G4TimeDemo from Altor Systems, but since this only runs properly on the ATI card, I only tested this with the B&W G3. This is very much optimized for Altivec, and thus represents only a rather theoretical comparison, at least for current games.


page8: enough talk: the benchmarks>>>