Accelerate Your Mac! logo
The Source for Mac Performance News and Reviews

Initial Tests: Dual G5 2GHz with Radeon 9600 vs 9800
compared to MDD G4 Dual 1.25GHz w/9800

Published: 9/17/2003
(Updated 10/4/2003 for Jedi II and LW 3D Tests)
(Updated 10/15/2003 for beta G5 Cinebench2003 Tests)

= Game Tests | Apps Tests | Benchmarks =

Intro: (If you want to skip the chit-chat and go direct to the tests, click here)
I want to preface this article to note this is just the results of initial tests I've done in the first day of having the Dual G5. There's a lot more I'd like to do but after so many requests to compare performance of the 9600 vs 9800 Radeon (retail) card, I wanted to post this as soon as possible. (The advantage of the retail Radeon 9800 Pro over the BTO 9800 Pro card is the fact only the retail model can use the ATI displays 3D/GL Overrides feature. The downside is that the retail card does not have an ADC port for those that have or plan to buy an ADC display. A $99 adapter would be required for that. The BTO card is an 8x AGP version, where the retail is 4x AGP, but I don't think that is a real factor in performance based on 8x vs 4x AGP PC tests.)

Notes on Retail 9800 Pro card G5 Install:
The retail 9800 Pro card requires a power supply connection (standard 4-pin molex connector) and there's no spare PS connector in the G5. The G5 hard drives are SATA and do not have a standard molex connector so the only existing one is connected to the Optical drive. In the first install (rushed that day to get tests done for posting) I routed the 9800's Y adapter power cable up the rear area to the HD cage area - but I've found a better solution that has less clutter and does not require a 2nd Y adapter. Here's the better routing solution:

  • Remove the Optical drive (see owner's manual for details, one cable disconnect and flipping two levers at the drive side) and reroute its Power Supply connector to the lower area (the PCI/AGP card slot area) by passing the Optical drive power cable down through the slot where the IDE cable passes through (there's clearance hole in that area adjacent to the cable slot) Then the 9800's "Y" adapter cable will reach the connector. (Note - you can also try just routing the 9800's Y adapter cable up into the optical drive area instead and connect from there.)
  • After connecting the 9800's Y adapter cable to the optical drive P.S. connector, pass the "Y" adapter's drive connector up through the same area and connect to the Optical drive. (Not needed if you can route the Y adapter cable into the optical bay area to begin with.) (It's a little more difficult to install the Optical drive back this way, but not that bad.)
  • To secure the "Y" adapter/PS cables I plan to use cable ties and adhesive-backed cable clamps in the PCI slot area flat surfaces. (This will keep the cables away from PCI slots and reduce clutter in that area.)

First Impressions on Dual G5:
It feels very fast of course. My first impression was how fast it booted to OS X (granted the disk isn't very full or fragmented). The time from the end of the gray screen at boot (which BTW ends much faster than a MDD G4) to having the OS X desktop ready to use feels like about 10 seconds although I've not timed it. Starting the MDD and the G5 at the same time shows the MDD is still at the gray screen when the G5 is ready to use.
The system is quieter than even a rev B (revised PS/Fan) MDD model, although when placed on a hard desk surface (i.e. desktop) you can hear the drive seeking and the fans. Fans get louder of course when running taxing apps or 3d games, but still an improvement over the MDD G4s in my opinion. Unlike past Superdrives I've owned, this one (Pioneer DVR-106 in mine, although some G5s have Sony DW-U10) actually seems to spin up to high speed with CDs. Especially when on a desktop/hard surface, this means some additional noise. Also lately I've noticed a distinct "hum" from the upper area (not the lower area power supply and not from the superdrive - it seems to come from either the upper hard drive fan or the hard drive.) I've not noticed any of the 'beeping' or 'chirping' noises some other G5 owners have reported. (See my G5 Noise reports/Tips page for those reports and a tip that helped most.) Any computer should be kept away from dusty floors, but with the G5's open vent area and no filters, I'd pay closer attention to this.
My wish-list items include a 2nd Optical drive bay and more internal drive bays. Also I wish they'd include at least a 2-button mouse, but I already own a Logitech (under $20) 2-button mouse w/scroll wheel. As a FYI - if you didn't order the BTO option for internal bluetooth module, it's not end-user installable per the manual, so you'll need to use a USB port Bluetooth adapter. The front mounted Firewire 400 port (and USB 2.0 port) are welcome additions. (One FW800 port on the rear as well as another FW400 port and additional USB ports.)
The built-in speaker also sounds better than the MDD G4 model in my opinion, but of course you'll want to use external speakers for quality audio (and the G5 has digital/optical audio In/Out as well).

After seeing the specs on the G5 when it was announced I wondered if the faster bus and memory (dual channel) design would better 'feed' a fast graphics card. As seen so many times in the past with 3D/First Person games, the previous Mac designs just don't come close to saturating a modern graphics card or allow it to reach its maximum potential. I hate to keep repeating it but since some still buy a new graphics card and are disappointed that it's not a lot faster at lower resolutions than their previous card, I'll say again that performance depends on several factors such as:

  1. The system's ability to feed it data
  2. The application/game code (efficiency and support of card features)
  3. The quality/efficiency of the drivers
  4. The resolution and quality settings used
    (at lower settings it may be no faster than an older card, due to item 1)
  5. The limits and features of the card design
    (higher resolutions/settings demonstrate this)

The good news is that the G5 DP system seems to be a clear improvement as far as item 1. And as more software is optimized for the G5 (none tested here are), performance should also improve. I have Photoshop 7 (and the G5 plugin update) but have not had time yet to test that app, but it's on the to-do list.

Although there's not as many tests here as usual, it's the best I can do in the first day of use with the G5 (and preparing for Hurricane Isabel). I show game tests first since there's been less of that posted to date for the G5. I show both a lower resolution (more CPU bound) and higher resolution (to stress the graphics cards) to illustrate the performance difference in the 9800 vs 9600 card in the G5 and how two systems with the same card compare. In many cases like 2D tests, the results are more CPU bound than video card.

System Details: (All OS X updates applied before testing)

  • PowerMac G5 Dual 2GHz OS 10.2.7 (build 6S80), boot rom 5.0.2f2, 1.5GB RAM, 160GB HD, Superdrive (Pioneer DVR-106), Airport Extreme card, with OEM ATI Radeon 9600 Pro and Retail Radeon 9800 Pro cards tested.

  • MDD (Mirror Drive Door) Dual G4 1.25GHz OS 10.2.6, 1.5GB RAM, 120GB HD, Superdrive (retail DVR-105), Airport card, Retail ATI 9800 Pro graphics card.

A few non-game tests include results from a Digital Audio Dual G4/533 (10.2.6) with 1GB of RAM and an ATI 8500 graphics card and a 1.4GHz CPU upgrade.

= 3D Game Tests =

Jedi Knight II MP Demo Tests

This game is very CPU/Bandwidth bound (i.e. often delivering similar FPS rates across resolutions with high end cards) and I was interested to see if the Dual G5 would 'feed' the 9800 Pro card better than previous Macs. (From the results it clearly does.)

Jedi Knight II non FSAA

Quake3 Arena Tests

This old game is still popular and one of the few on the Mac that seems to scale well. All tests used the latest Quake3 1.32 "G4" version. Sound was enabled for all tests but music volume was set to 0 and all game options were on. First a test with Quake3's "high quality" settings + high geometric detail and maximum texture quality. (r_smp set to 1 to use dual CPUs, and my typical "s_chunksize 4096" vs default 2048 was used..)

Quake3 max quality Tests

(Note - the standard "High Quality" settings w/o max texture quality/high geometric detail would show a bit higher rates than the above "max" settings - for example with std "high quality" only settings and no config file tweaks I get 340FPS at 1024x768 and 223.2 FPS at 1600x1200.) Of course at the maximum resolution (2048x1536), the MDD and G5 delivered literally the same rates with the 9800, but that's because the card not the CPU/system is the limiting factor at that resolution. You can see the 9600 running out of breath at higher resolutions compared to the 9800. I didn't run the 2048x1536 max quality/settings test with the 64MB 9600 card, due to crashes with 64MB cards at that resolution.

Although personally with a high end system and graphics card, I'm more interested in high quality settings performance, I also ran a test with the "Fastest" setting (low res/16bit/ugly), just to see what speeds would be seen on the G5 vs G4. (Use of custom configs like boli's also can boost framerates at the expense of image quality.)

Quake3 Fastest Tests

Note: after a reader request, I ran 1600x1200 tests on the Dual G5/9800 combo with sound disabled and standard "High Quality" settings (which have medium Geometric detail and Texture quality slider down from Max. The "Max Quality" tests I ran had HQ + high Geometric detail and Texture quality slider set to max, which takes a bit of a performance hit). The dual G5 2GHz w/9800 pro card (w/sound disabled and std HQ settings) delivered 227.4 FPS at 1600x1200 and 400.1 FPS at 640x480.


Unreal Tournament 2003 Tests

I have reams of test data with the 9800 Pro vs GF4 Ti vs 9000 vs 8500 cards from tests in the MDD Dual 1.25GHz to graph, but with the G5 only being here for a day, I had to limit the number of tests. All UT2003 tests used max image quality settings and the benchmark procedure mentioned on the UT2003 demo tests entry page.
Since the game is already a CPU hog, I modified the INI file to set Audio channels to 4. (The flyby test defaults to no audio, but the botmatch test does have sound enabled.) I used the Asbestos benchmark, since it was less CPU bound than the outdoor Antalus level (in both flyby and botmatch). With options like FSAA enabled, you can see the performance difference between the 9600 vs 9800. (Note: "8x AF" in the graphs means 8x Anisotropic Filtering was enabled.)
Longer bars/higher scores are faster of course. The Average framerate is graphed but the Min/Max rates are shown in parenthesis.

UT2003 1024x768 no FSAA

UT2003 1024x768 4xFSAA 8xAF

UT2003 1600x1200 no FSAA

UT2003 1600x1200 4xFSAA 8xAF

The 9600 Pro (64MB) card could not run this test at 1600x1200 - attempting to do so resulted in the app quitting with a memory allocation error.

Note: Since only the retail ATI 9800 Pro can use the ATI 4.1 displays control panel 3D/GL overrides feature (for selecting FSAA, Anisotropic filtering, etc.).


Return to Castle Wolfenstein Tests

I used the same RTCW Multiplayer demo tested used in CPU upgrade reviews, sent by a reader. (Note: With RTCW v1.41, you have to change the "59" in the demo filename to "60". Once the demos folder is installed, just open the console using ~ and type Timedemo 1 and Demo demo0001.) All tests used High Quality game settings, audio enabled, but music volume set to 0.

RTCW MP Tests

Quite a difference with the G5 and the 9800 delivers much better performance than the 9600 at 1600x1200. (At 2048x1536, even the 9800 card is maxed out.)

One thing I wanted to mention is that until the G5, I've seen some unexpectedly low FPS rates (with the instantenous FPS counter enabled) in single player RTCW. I keep a saved game to test for this (from one of the first levels, where you're in the room that has the holes in the floor with solders in the lower room). Standing in the corner of that room even without any action happening shows dips into the 20s, with averages in the 38-45FPS range (at 1024x768/HQ). With the 9800 pro in the MDD 1.25GHz DP, there were peaks in the 65FPS range but often dipping to 40-45FPS in firefights. With the G5 DP (even with the 9600 card) - the FPS meter was usually pegged at 80-90FPS and never dropped anywhere near as low as it did with previous Macs.


Unreal Tournament X PR3 Tests

In some ways I still prefer the original Unreal Tournament over the 2003 version although it's not anywhere near as pretty. It's not a great test of video card performance typically but it's a popular game and people will want to know kind of performance to expect. UT X PR3 was set to Maximum texture/skin detail and MinFPS = 0, audio on and music set to min volume. I ran tests at 4x FSAA, 1024x768 mode (highest possible with super-sampling, although the 9800 supports Multi-sampling some older cards don't) and also at 1600x1200 (w/o FSAA).

UT PR3 FSAA Tests

UT PR3 1600x1200 Tests


= A Few Apps Tests =



Lightwave 3D 7.5c Tests

LW Hummer Scene Playback

I finally made time to install Lightwave 3D 7.5c on the Dual G5 (w/9800 Pro retail card) and ran the "Hummer" scene preview playback test I've used in some past graphics card reviews. I'm not an experienced LW 3D user but I was impressed to see the preview playback on the Dual G5 was literally twice as fast as with a MDD Dual G4/1.25GHz with the same graphics card installed.

    Playback Hummer Preview (full size)
  • Dual G5 2GHz/9800 retail: 22.9 sec.
  • Dual G4 1.25GHz/9800 retail: 46.3 sec.

(Different graphics cards were tested in the MDD G4, the fastest at the Hummer playback was the GF4 Ti at 43.8 seconds.)


Quicktime to MPEG4 Conversion

As in all the CPU upgrade tests this year, I used a very large 1.92GB (GigaByte) Quicktime movie and timed how long it took to export (convert) to MPEG4 using the default settings in Quicktime 6 Pro (w/all updates). I've included results from some past reviews of older Macs and CPU upgrades as well as the MDD G4 and G5. (Shorter bars are faster of course.)

QT to MPEG4 Export Test Results

QT6 Pro export shows no real benefit from dual CPUs. High end apps like Final Cut Pro do however, although I can't afford them for benchmarking right now. (I do have Final Cut Express and as soon as I can create a test project I'll include results with that program.)


iMovie 3 Export Tests

I used the same test I have since iMovie was first released - stacking the 6 tutorial file clips end-to-end (no transitions) and timed how long it took to export the movie using the standard "CDROM" settings. (By using the tutorial with no variables like transitions, it's something everyone can easily test with their own systems.) Shorter bars are faster.

iMovie3 QT Export Test Results

As you can see from the results, this is another app that doesn't take advantage for dual CPUs, and isn't optimized for the G5 (yet at least).


Appleworks 6.2 Scrolling Tests

With the desktop set to 1600x1200, millions colors I used AppleWorks 6.2.7 to test scrolling times from top to bottom of a 100 page, multi-column newsletter with images. As usual, in the same system, the differences with different graphics cards is small (within the margin for error of starting/stopping a stopwatch).

Appleworks scrolling test

2D performance isn't really graphics card bound (at least with a modern graphics card).


PSBench7 (Photoshop7) G5 Results/Comparisons: (from a reader mail, I've used PSbench in past reviews/tests but not yet run it personally on a G5)

"Hi Mike,
Over on the Ars forums Ps7Bench results for a dual G5 have been posted. They are better than I thought they would be. Out front by a *very* nice margin. Some of the PS pros say that some of the tests in this bench are hardly ever used (e.g. watercolor filter). Hopefully a new PS bench will appear.
-ADDRAGYN
PS7Bench
Ars thread

Scores (Normalized, higher score is better)

2x 2000 G5 OSX 10.2.7 (G5 plugin) 555 (energy settings highest perf)
2x 3060 Xeon (OC'd 2400)          488
2x 2930 Xeon (OC'd 2400)          471
   3200 P4 (800MHz)               427
   3000 P4 (800MHz)               405
   3495 P4 (OC'd 3.06)            386
   3060 P4 XP Pro (533 FSB)       358 HT
2x 2200 Xeon PC 800 RDRAM CPQ Evo 357 HT
2x 1500 G4 (OC'd 1420)            348
2x 1333 G4 DDR OS9.2 (oc'd 1.25)  346
   1800 G5 OSX 10.2.7 (G5 plugin) 344 (energy settings highest perf)
2x 1420 G4 OSX 10.2.4             338
2x 2400+Athlon MP                 338
2x 1250 G4 OS 9.2.2j              337
   3200+Athlon XP                 332
   1800 Opteron(dual-chnlDDR 333) 332
2x 1333 G4 DDR OSX10.2.2(oc 1.25) 326
   1800 OPteron(singl-chnlDDR333) 320
   3000+Athlon XP                 318
2x 1250 G4 OSX 10.2.5             318
2x 1250 G4 DDR OSX 10.2.1         316
2x 1800 Athlon MP                 312
   2800+Athlon XP Barton          298
2x 2000 P4 Xeon                   286
2x 1200 G4Powerlogix(867MHzG4/QS) 285 upgraded
2x 1533 Athlon MP                 285
2x 1533 Athlon MP                 283
   2530 P4 mobile (OC'd 1400)     282
   2700 P4B (OC 2400, 600 MHz FSB)280
2x 1466 Athlon XP                 279
   1600 G5 OSX 10.2.7w/G5 Plugin  276 *MacNNscores
   2666 P4 (DDR 333)              269
2x 1000 G4 DDR 10.2               267
   2400+Athlon XP                 262
2x 1000 G4 OS9                    260
2x 1000 G4 OSX 10.1.5             254
   2400+Athlon                    252
   2400 P4B (800MHz)              251
   2400b (sis 648 DDR400)         251
   1600 Centrino IBM T40          250
   2400 P4 (533MHz bus)           249
   2400 P4 B                      241
   2340 P4 (overclock)            239
   1600 Centrino Dell D800        236
   2400 P4                        234
   1800+Athlon XP (1533 MHz)      226
   1577 oc'd Athlon XP (Lestat)   221
2x 1000 G4 OSX 10.2.2 (upgraded)  218 ?!(dual 533 logic board)
   1548 Athlon XP                 214
   1670 Athlon XP (2000+)         213
   1667 Athlon XP                 211
   1400 Athlon XP 1600+ xp pro    200
1x 1533 Athlon MP                 197
   1300 Centrino Sony VAIO Z1A    196
   1000 G4 17" Powrbk OSX 10.2.6  196
   2000 P4 Xeon                   194
   1400 Athlon XP 1600+'98SE      191
   1000 G4 OSX TiPbk 10.2.2       185
2x  533 G4 OSX 10.1.5             175
2x  533 G4 OS 9.2.2               174
   1800 P4                        173
   1200 AthlonMP                  168
   1508 Celeron (overclock)       167
   1400 PIII Tualatin             160 **?
2x  550 G4 OSX 10.2.3 (OC Cube)   160 **?
2x  500 G4 OSX                    152
2x  450 G4 OS9                    151
   1333 Athlon TBird              147
2x  450 G4 OSX 10.1.5             143
    800 G4 Pbook OSX  1MB L3      135
    733 G4 (miro7)                134
    667 G4 PBk OS9 noL3           127
    667 G4 PBk OSX 10.2.3 no L3   125
    466 G4 OS9                    123
    667 G4 OSX TiPBk 10.1.5 noL3  121
    866 PIII                      114
    466 G4 OSX 133 MHz bus        112
    550 G4 Powrbk OS9*            104
    500 G4 Pbook (OC'd 400)       103
1x  450 G4 OSX 100 MHz bus        101
   1000 Athlon TBird (PS6.01)     100
    550 G4 Powrbk OSX*             95
    933 Transmeta Crusoe Sony      78
    700 G3 iBook                   74
    600 G3 iBook OS 9.2.2j         70
    233 PII                        30


= Benchmarks =

Since I prefer real-world tests over pure benchmarks, I didn't focus on them for this first posting.


CineBench 2003 Tests

I used Maxon's Cinebench 2003 benchmark (available here) with the desktop set to 1600x1200, millions colors. I've included both the render score results scores as well as the Framerates for the two 3D scene fly-by tests. (I didn't spend time to graph the software/GL SW/GL HW scores yet, but the scene flyby tests I think better illustrate the differences in performance.)
(Updated for Beta G5 version released Oct. 15th, 2003)

Cinebench 2003 results

I also graphed the detailed results from CineBench regarding the Hardware OpenGL framerates for the Scene1 and Scene2 fly-bys.

Cinebench Scene FPS rates

(I know XBench 1.1 and later were tweaked for the G5 a bit but I've not tested with v1.1.3 yet.)


= Return to Home Page =

Copyright ©, 2003, all rights reserved.
All brand or product names mentioned here are properties of their respective companies.