Click for SMOKING FAST 6G SSDs!
Click for SMOKING FAST 6G SSDs!


A Click shows your site support to my Sponsors

Accelerate Your Mac! - the source for performance news and reviews
The Source for Mac Performance News and Reviews


First Look: XLR8's G4 400MHz CPU Upgrade
How does it compare to the G3?
Published: 9/02/1999
(Updated 9/4/1999 for SoundJam 1.1 tests)
G4/400 vs G3/400 in Benchmarks and Real World Performance Tests
Benchmarks | Apps Tests  | Game Tests | Observations  | Summary | G4 Specs | Related Links

Introduction:
In some ways testing a G4 CPU with the current MacOS and applications is like running a V8 engine on 4 cylinders; you're not using 1/2 the horsepower of the processor. Test results in current apps with no Altivec support were disappointing based on the 'supercomputer' comments I'd heard on the G4. The only software I had that used the G4's main claim to fame (the Altivec instruction set that Apple calls the 'Velocity Engine') was XLR8's PowerFrax fractal generator and Soundjam MP v1.1 which did show substantial performance increases with the G4.

Even I was puzzled by some of the results in non-Altivec applications and I retested repeatedly to ensure they weren't a fluke. Perhaps disabling 'Speculative Processing' hurts G4 performance more than it does for the G3. Enabling 'Speculative Processing' on the G4 caused boot errors in my tests, at least with this beta control panel/extension on this 9500 motherboard Mac. G4 systems from Apple will have it enabled I'm sure and will help performance in most apps I'm sure. Photoshop was one application where I've seen better performance with it disabled however (based on previous tests in CPU Upgrade reviews).

Things look much brigher for the G4 in the future however, with OS 9 said to have some core Altivec support and with Quicktime (and QD3D/OpenGL I hope) also set to use the 'Velocity Engine'. Adobe's recent press release indicates their G4 core/lighting plug-ins can boost Photoshop 5.5 performance in some functions by 10 to 15 times. That claim is impressive and I was very disappointed to not be able to obtain the plug-ins for this article. I'm hopeful performance in Apple's G4 systems (Sawtooth) will show better results than these tests with an older Mac since that system has an improved bus that takes advantage of the new 'maxbus' feature of the G4 CPU. I have a G4/AGP 500MHz system on order and will update this page with results when it arrives.

As you'll see in the tables below, Altivec enabled applications (PowerFrax and Soundjam) showed good gains from the G4, but most existing applications (without Altivec support) perform almost identically on a G4 as a G3 CPU of the same speed. Games were one exception where I consistently saw approximately 10% better framerates with the G4 CPU. The biggest surprise for me was the memory bandwidth benchmark; where the G4 produced substantially higher rates in the memory bench test than the G3 even on this older Mac's 50MHz system bus and old PCI/memory controller. I had assumed the G3 had already saturated the older Mac's memory bus.

As I mentioned before, I was confused at some of the test results but repeating the tests confirmed they were consistent. Memory bandwidth tests showed the G4 as a clear winner even for data larger than the backside cache, yet MacBench CPU and Bytemark Integer tests both showed the G4 a hair slower than the G3. Rendering applications like Bryce2 and Infini-D showed literally identical performance with either processor. I can only guess the beta control software, disabled Speculative Processing, or some errata (bug/error) in this particular early G4 chip may be responsible. This particular G4 chip was made in in May of 1999 (date code 9919, stepping 2.2). Current G4s with later steppings (revisions) may perform differently. I hear that Apple's shipping G4s have v2.6 of the G4 and that 500MHz models may have v2.7.


Test System:
Since both my B&W G3s had firmware that would not allow the G4 CPU to boot (v1.1f1 and v1.1f4), I used my loaded Genesis (9500/Tsunami motherboard based) system. If I had more time to keep the G4 I'd have tested in the Beige G3 and 9600/350 as well (and perhaps the PowerCenter Pro and PowerTower Pro - but that would take more time).

  • Daystar Genesis (9500 Mb based):
  • 512MB RAM (matched pairs of 64MB FPM Dimms)
  • Seagate Cheetah UW SCSI 4.5 GB HD (ST34501W)
  • 2 ATTO PCI SCSI cards
  • 3Dfx Voodoo3 2000 Graphics card (beta2 ROM/drivers)
  • OS 8.1, 4MB Disk Cache, VM off, QT 2.5, QD3D 1.5.3 [No Libmoto]

For these tests I used a 3Dfx Voodoo3 2000 graphics card. There were two reasons for using the Voodoo3 - a) the Orion's drivers caused hangs with every OpenGL game tested with the G4 CPU and b) it is the only card with enough fill-rate performance to not be a bottleneck for a fast CPU. As noted in my video card reviews, most other Mac video cards just run out of breath at 1024x768 in 3D games. The Voodoo3 doesn't.

G4 and G3 zif pix

During all tests the only thing that was changed was the CPU card and control panel/extension. XLR8 provided an early 1.4.1b0 version for use with the G4 upgrade.

Speculative processing was left at the default disabled setting (per XLR8's recommendation). Enabling it may have helped some tests such as MacBench. With this early beta G4 control software, enabling Speculative Processing caused problems which prevented me from testing with that feature on. Unless noted otherwise, both the G4 and G3 tests had Speculative processing disabled.

Most of you will be disappointed (as I was) in the fact that most of the applications I tested showed no real gain with the G4 over a G3 of the same CPU speed. Even MacBench's CPU score was literally identical (oddly the G4 scored 2% less than the G3 in repeated tests). Macbench's FPU score did favor the G4 by 11% however.


Benchmark Tests

PowerFrax:
XLR8 sent a revised version of their PowerFrax fractal explorer program which has an option to enable Altivec support. Enabling Altivec resulted in a consistent 3.7X-4X reductions in times to generate the zoomed image.


MacBench 5.0 Tests:
I found it hard to believe the G4 scored 2% less than the G3 of the same speed, even in a benchmark like Macbench. However I repeated the tests 3 times and the results always showed the G4 trailing by 2-3%. Perhaps the beta control panel/extension, disabled Speculative Processing or even this particular G4 CPU stepping (chip revision) are reasons for the G4's disappointing score. As XLR8 commented in their white paper, for non-Altivec software the G4 performs similarly to the same speed G3. The G4 did consistently score higher on FPU tests however.



ByteMark DR/3:
I've never been a fan of ByteMark as noted in my other CPU upgrade reviews. However some readers will want to see the results so I show them here. ByteMark is not an indicator of total system performance and results don't necessarily equate to real world applications performance. Speculative processing was off for both the XLR8 G3 and G4 tests.

ByteMark DR/3 Results
Upgrade/System
Integer
Floating Point
XLR8 G4 400
(Genesis/9500)
400/200/50
12.81
9.98
XLR8 G3 400
(Genesis/9500)
400/200/50
12.94
8.59
Blue & White G3/400
(400/200/100)
OS 8.6
12.90
8.71


I can only guess the beta XLR8 software or this G4's particular stepping (cpu revision) or just standard deviation is the reason for the lower integer score. Note the G4 scored about 15% faster in Floating Point math than the G3.

Memory Bandwidth Tests:
I ran tests with a small 'Memory Bench' utility (can't remember where I obtained it last year) as Michiro Isobe (author of PPC Checker) asked that I try to verify the reports of improved memory pipelining in the G4. As shown below, the G4 at the same clock speed and with the same backside cache size/speed was dramatically faster in this test than the G3 CPU. What surprised me was the rates were still higher even for data exceeding the size of the backside cache (below the red line in the tables).

In the tables below, scores above the red line are for data sizes that fit in the backside cache - below the red line indicates main memory bus performance.

Memory Bench shows the G4 as a clear winner, with the highest rates I've seen from any older Mac. For rates from the 100MHz memory bus B&W G3/400 see this image from my Interleaved RAM article.

Applications Performance

SoundJam MP 1.1 G4 Tests: I downloaded the Altivec enabled SoundJam v1.1 demo and ran tests converting a CD audio file (Scorpions "Winds of Change" song) with the G4 and G3. Both cards were set to the same CPU/Cache/Bus speed (400/200/50) so the difference in times were a direct result of the Altivec instructions. The Genesis has an Apple 24X CDROM drive and the song was 5 min. 13 seconds in duration. [Update: Due to reader requests, I've added results of MP3 encoding with a Dell Pentium III 500 (Win98) using RealPlayer's free Jukebox application.] The results are shown below:

SoundJam MP v1.1
(Convert Audio Cd Track to MP3)
CPU
Rate
CD Drive
XLR8 G4 400
(Genesis/9500)
400/200/50
5x
Apple 24X SCSI
XLR8 G3 400
(Genesis/9500)
400/200/50
3.1x
Apple 24X SCSI
Dell PIII 500
(Win98/RealPlayer Jukebox)
500/250/100
8.2x
Kenwood 52X
updated ROM
Dell PIII 500
(Win98/RealPlayer Jukebox)
500/250/100
6.4x
Toshiba DVD M1202

[Update: A B&W G3/400 owner with 32X CDROM drive reported a rate of 3.1X. He also noted only 1 second difference between encoding times from CD or the Hard Drive.]

A better/faster CDROM drive may have improved times, but personally I was impressed with the gain provided on this early (May 1999) stepping G4 CPU. As I mentioned previously, sources tell me early G4 CPUs have errata (term for bugs in the chip design) that probably doesn't exist in the current G4s which may offer even better performance.


Results of rendering applications surprised me, as even without Altivec I expected the stronger FPU of the G4 to deliver some benefit, however the completion times were literally identical to the same speed G3. This backs up XLR8's claim that for non-Altivec enhanced apps the G4 delivers about the same speed as a G3 CPU.

This page lists test results in common Mac benchmarks and applications like Photoshop 5, Bryce 2, Infini-D 4.01, After Effects 3.1, Premiere 4.2 and popular 3D games like Quake2, Q3test 1.08 and Unreal v224b7.

Infini-D 4.01 Tests

I used the same Infini-D 4.0 tutorial file rendering test as I've done in past reviews using the 'Chapter 7 completed' scene. Rendering quality was set to Ray Trace, medium anti-aliasing, shadows on, patch detail low. Infini-D was allocated 40MB of RAM for the tests. The Genesis tests used OS 8.1 , the B&W G3 used OS 8.6. Speculative Processing was disabled for the Genesis tests which hurts Infini-D scores based on past experience (appx. 1 minute increase in this test).

Infini-D 4.01 Make Movie Tests
(Minutes:seconds)
Upgrade/System
Time
Notes
XLR8 G4 400
(Genesis/9500)
400/200/50
40.34
Spec Proc. OFF
XLR8 G3 400
(Genesis/9500)
400/200/50
40.49
Spec Proc. OFF
Blue & White G3/400
(400/200/100)
39.02
OS 8.5
OS 8.6 was slower


Bryce 2 Tests

I also ran tests in Bryce 2, using the samples scenes in the KAI folder called "Magical Knight I" and "Alexandria II". Bryce 2 was left at the default memory allocation of 16980k, antialiasing was on, screen resolution was set to 1024x768, thousands colors. Genesis tests used OS 8.1, the B&W G3 used OS 8.6. Bryce2, like most rendering apps, is primarily CPU speed dependent and does not usually show any advantage to faster bus speeds.

Bryce 2 Rendering Tests
(Minutes:seconds)
Upgrade/System
Magical Night I
Alexandra II
Notes
XLR8 G4 400
(Genesis/9500)
400/200/50
3:57
2:55
Spec. Proc. OFF
XLR8 G3 400
(Genesis/9500)
400/200/50
3:51
2:51
Spec. Proc. OFF
Blue & White G3/400
400/200/100
3:53
2:50
OS 8.6

Again I was surprised that this CPU/FPU bound program didn't benefit from the G4.

After Effects 3.1 Tests

The following is a chart showing the time to render a special effects movie (10MB file size) in AE with the G4/400 and G3/400 with a B&W G3/400 for comparison. Resolution was set to 1024x768, thousands colors as was common on all but the Photoshop 5 tests.

After Effects 3.01 Make Movie Tests
(Minutes:seconds)
Upgrade/System
Time
Notes
XLR8 G4 400
(Genesis/9500)
400/200/50
3:30
Spec Proc. OFF
XLR8 G3 400
(Genesis/9500)
400/200/50
3:33
Spec Proc. OFF
Stock B&W G3/400
400/200/100
3:34
OS 8.6


Photoshop Tests:

I also ran a complete PS5Bench (21 filter test) series. System settings were per PSBench requirements - 1024x768, millions colors, Interpolation set to bicubic (better) and Photoshop should be allocated enough RAM (I used 80MB) to avoid any swap file activity from the 10MB test image filter actions (according to Adobe disk activity still happens however). Virtual Memory was off as with all my tests. I never touched Photoshop's history settings (default of 20 with 'automatically create first snapshot' enabled). I was unable to obtain the G4 Altivec plug-ins for Photoshop 5.5 but will update this page with those tests if they are received before I have to return the G4.

All results below were with Photoshop 5.02, which favored the G4 by a small margin. Numbers in the chart are seconds to complete the filter action, so lower scores are better.

PS5Bench Test
G3 400
400MHz CPU Speed
1MB Cache at 200MHz
50MHz Bus
512MB RAM
Spec Proc OFF
Voodoo3
OS 8.1 VM OFF
G4 400
400MHz CPU Speed
1MB Cache at 200MHz
50MHz Bus
512MB RAM
Spec Proc OFF
Voodoo 3
OS 8.1 VM OFF
Rotate 90
0.9
0.6
Rotate 9
3.6
3.4
Rotate .9
3.3
3.2
Gaussian Blur 1
1.5
1.3
Gaussian Blur 3.7
4.1
3.0
Gaussian Blur 85
5.4
4.0
Unsharp 50/1/0
1.5
1.4
Unsharp 50/3/7/0
4.6
3.4
Unsharp 50/10/5
4.7
3.5
Despeckle
2.5
2.4
RGB-CMYK
5.1
5.1
Reduce Size 60%
1.4
1.3
Lens Flare
4.9
4.7
Color Halftone
6.0
5.9
NTSC Colors
4.0
4.0
Accented Edges
10.8
10.7
Pointillize
17.2
17.4
Water Color
22.9
22.4
Polar Coordinates
6.3
6.0
Radial Blur
37.4
36.2
Lighting Effects
9.4
9.2
Total Time (seconds)
157.6
149.1
System
G3 400
400MHz CPU
Speed
1MB Cache at 200MHz
50MHz Bus
512MB RAM
Spec Proc OFF
Voodoo3
OS 8.1 VM OFF
G4 400
400MHz CPU Speed
1MB Cache at 200MHz
50MHz Bus
512MB RAM
Spec Proc OFF
Voodoo 3
OS 8.1 VM OFF

Note: My stock Apple B&W G3/400 (OS 8.5.1) scored 148.6. A Pentium III 500Mhz (500/250/100) Windows 98 system with 256MB SDRAM took 165.7 seconds to complete this test. Adobe's Pentium III Photoshop SSE update was applied prior to the test.

See my Photoshop5 Performance page for more results including Pentium III/Dual Pentium II


Game Performance Tests:

Games did show approx.. a 10% boost as I had suspected. I used a 3Dfx Voodoo3 2000 PCI card since most other Mac video cards are the bottleneck with a fast G3 CPU. Remember that regardless of your CPU speed if your video card can only produce 30 frames per second a faster CPU card won't help. The Voodoo3 has the highest fill rate of any Mac card I'm aware of and framerates depend on the fill rate of the card (and a CPU that can keep the card 'fed' with data).

Tests were done in Quake2, Q3test 1.08 and Unreal v224b7. Since the video card is a major factor in the speed (and quality) of 3D games, I used a Voodoo3 which unlike most other Mac video cards, can't seem to be saturated even with the fastest CPU.

Quake2 Performance: The following table illustrates Quake2 performance with the upgrade. Included in the scores for reference is the same Voodoo3 card installed in the 66MHz PCI slot of a Blue and White G3/400.

Quake2 Demo 1 Results
Using multitexture RefGL.lib
System/Upgrade
640x480
800x600
1024x768
Stock B&W G3/400
(400/200/100 w/Voodoo3)
Card in 66MHz PCI slot
54.1
53.8
52.5
XLR8 G4 400
in Genesis/9500
(400/200/50 w/Voodoo3)
47.6
47.6
47.3
XLR8 G3 400
in Genesis/9500
(400/200/50 w/Voodoo3)
43.5
43.5
43.3


Q3Test 1.08: I used Q3Test v1.08 (latest currently available) for all tests. Graphics settings in the game were :

  • 16-bit color
  • Lightmap lighting
  • Geometric detail high
  • Texture quality 16-bit
  • Texture detail 1 notch from max
  • Texture filter - bilinear

As shown in the table below, the G4 delivered appx. 10% faster framerates than the G3 in the same Genesis system.

Q3Test 1.08 640x480 800x600 1024x768
B&W G3/400
w/Voodoo3 2000
in 66MHz PCI slot
40.0 39.3 36.7
XLR8 G4 400
in Genesis/9500
(400/200/50)
w/Voodoo3
33.7 33.6 32.3
XLR8 G3 400
in Genesis/9500
(400/200/50)
w/Voodoo3
30.8 30.7 30.1


Unreal 224B7: The table below summarizes results from the 3rd cycle timedemo of the 'Castle flyby' scene. (All tests used the same detail settings using our standardized Unreal.ini files for RAVE and 3dfx cards.)

Unreal 224B7 Castle Flyby Tests
(High Quality Detail Settings)
System
640x480
1024x768
B&W G3/400
(400/200/100)
w/Voodoo3
Card in 66MHz PCI slot
50.1
40.5
XLR8 G4 400
in Genesis/9500
(400/200/50)
w/Voodoo3
44.35
38.12
XLR8 G3 400
in Genesis/9500
(400/200/50)
w/Voodoo3
39.23
34.97

.


Observations:

  • The G4 runs a bit hotter than the G3. XLR8's control panel indicated 48C temperatures in the Genesis. Often these software temperature readings are not accurate but the G4 ran a higher indicated temp. than the G3.
  • Don't expect the G4 to be as clockable as the G3. From what I hear from several sources in the industry, there may not be as much headroom in the G4, at least in the current steppings of the chip.
  • This particular G4 chip was made in in May of 1999 (date code 9919). Current G4s with later steppings (revisions) may perform differently.

For B&W G3 owners with updated ROMs, XLR8 is working on a fix for the ROM check to allow these systems to use G4 upgrades.

Summary:

Until more software is available that takes advantage of Altivec extensions in the G4 CPU, don't expect performance to exceed that of a G3 of the same speed. Games are one area where there was a small (10%) consistent boost without Altivec. These test results here are hardly an indicator of the G4's potential with software that uses its new instructions.

With OS 9 said to have core Altivec support (hopefully both Quicktime and Quickdraw3D/OpenGL), a G4 CPU should show significant gains for multimedia functions (video, audio) in the future. And with Apple's aggressive pricing on the G4 systems, look for 3rd party software support to increase, as having 'Optimized for Altivec' will surely be a way for companies to increase the performance and attraction of their products.

If you're looking to buy a new system, then a G4 is the logical choice. For owners of G3 CPUs, I'd suggest waiting until a) more software uses Altivec and b) prices for G4 upgrades drop. Since the G4 supports up to 2MB of backsidec cache (twice the G3 limit), this feature should be of benefit to older Macs as well. Although there are no announced G4 systems or upgrades with more than one processor, unlike the G3 the G4 was designed with multiprocessor support. Imagine OS X with two or four G4s...

PowerPC 7400/G4 Specs: Based on Motorola info as of this date. Note that a 500MHz is due in October and my sources indicate it may have a 10x or 12X bus/CPU ratio.

PowerPC 7400/G4
based on 8/99 specs

CPU speed - Internal

350 MHz
400 MHz
450 MHz
(500MHz soon)

CPU/Bus Ratios

x3, x3.5, x4, x4.5, x5, x5.5,
x6, x6.5, x7, x7.5, x8, x9
(my sources mentioned 10x or 12X in next rev)

Bus Interface

64-bit

Bus Protocol
MPX/60x
Instructions per Clock

3 (2+branch)

L1 Cache Size

32 Kbyte inst
32 Kbyte data

Backside L2 Cache
Support

512 Kbyte, 1 Mbyte,
or 2 Mbyte

Backside Cache Ratios
(CPU speed/cache speed ratio)

1:1,1.5:1, 2:1, 2.5:1,
3:1, 3.5:1, 4:1

Typical/Max.
Power Dissipation

400 MHz:  7.9W/15.5W
450 MHz:  8.9W/16.3W
[Updated]

Die Size

83 sq mm

Package

360 CBGA

Process

0.15 micron 6LM CMOS

Transistors

6.5 million

Voltage

1.8 / 2.15V internal,
1.8 / 2.5 / 3.3V I/O

SPECint95 (est.)

21.4 @ 450 MHz

SPECfp95 (est.) 20.4 @ 450 MHz
Other Performance 825 MIPS @ 450 MHz
Execution Units

Integer (2)
Floating-Point
Vector

Branch
Load/Store
System



= Other Site Topic Areas =
Mac Models | CPU Upgrades | Storage | Audio | Video | Archives | Search

- Top of Page -

= Back to Home Page =


Copyright © 1997-2012, all rights reserved.
All brand or product names mentioned here are properties of their respective companies.
Site terms and conditions of use.