Click for Portable Storage Solutions!
The Source for Mac Performance News and Reviews
Review: XLR8's MACh Speed G4zTM
G4 CPU Upgrade for AIO/Beige/B&W G3s
|Benchmark tests were run with MacBench 5.0 as it is the accepted Mac standard. I've also included somewhat baffling results with three memory bandwidth benchmarks to see what the G4 CPU offers in that area, and to check Apple's claims of 2x higher bandwidth in their Sawtooth G4 systems. As you'll see below - there were some very odd and interesting results from the memory bandwidth tests. [Update: After this review was published I received a reply from XLR8's Director of Software Engineering Chris Cooksey (below) which has his explanation for the memory bandwith test results I observed with and without their software active.]
I'll also show results with some of my old favorites that I've not used in some more recent reviews - RaveBench and Walker QD3D Rave benchmarks (often used in my video card reviews). Often the video card is the real bottleneck to performance gains in tests like this, but there was some improvement from the G4 CPU seen in these tests with the same Rage128 graphics card.
BYTEmark tests are not shown since it returned a 0 score on the Sawtooth G4/AGP system. I'm not a big fan of BYTEmark as it is not a good indicator of other important system performance areas like disk, memory and graphics card. BYTEmark did run on a G4 upgraded Genesis system and results were shown in my First Look at the XLR8 G4/400 upgrade back in August.
Throughout this review I use the convention of showing CPU, Cache and Bus speeds (in MHz) in the format of (cpu speed)/(cache speed)/(bus speed); so 400/200/100 would mean a CPU speed of 400 MHz, a backside cache speed of 200 MHz and a system bus speed of 100MHz. The default settings for the MACh Speed G4/400 ZIF as shipped is 400/200/100 (it's set for a B&W G3's 100MHz bus speed, but can be easily set for Beige or AIO Mac's 66MHz bus speed within seconds).
Remember that Macbench Graphics scores will vary depending on graphics mode and the installed video card. Since all these Macs have graphics cards of widely differing performance capability, this is not a reflection on the CPU card under review. In Macbench 5.0 a 1000 score is the baseline based on performance with an Apple Beige G3/300 running millions colors, 1152x870, so consider this when evaluating any scores at lower resolutions and color depths.
Be aware that each card and system has some tolerance variation, so these speeds may not be attainable with every card/system/hardware combination. Due to the risks involved I do not recommend overclocking but do test for maximum speeds (if the upgrade is adjustable) during my reviews. The specific CPU chip on each card, your motherboard components, installed hardware and RAM mix can affect maximum reliable speeds. Never buy a CPU card assuming you will be able to overclock it or repeat anyone else's results (for instance reports in my Rate Your CPU Upgrade database). Remember to consult the card vendor and your system dealer for warranty impacts from overclocking any product (all disclaimers apply).
MacBench 5.0 Tests: The graph below compares the follow configurations and systems
All systems had the same OS version, 256MB of RAM, VM Off, and OpenGL 1.1.2 with its ATI driver update.
A red dot in the graph below indicates the upgrade was running beyond its rated speed. This speed was reliable in my particular system with this particular sample during my testing but these results may not be repeatable in your Mac. I do not recommend you overclock any CPU card due to the risks involved (see my FAQ's Overclocking Risks topic area). The main reason I set the upgrade to 450MHz was to allow a 1:1 MHz/MHz comparison to the Apple G4/450's performance. I did not run the upgrade long enough at 450MHz to satisfy myself it was 100% reliable at that speed, but during all the applications tests for this review I saw no errors. Don't assume that a G4/400 will always run at 450MHz however.
MacBench 5.0 Performance
Notes: Some explanation of the MacBench graph and test components.
G4timedemo is an benchmark for showing the benefit of Altivec support with G4 CPUs. I ran tests each of the system/upgrade combos and graphed the results. As you can see, a G4 CPU w/Altivec support does make a huge difference here (60% gain with a G4 CPU of the same speed). All tests were run with the desktop video mode of 1024x768, thousands colors and G4timedemo set to 'amazing' (max) quality.
G4timedemo results seem to vary a few FPS from run to run. I repeated the tests several times (each after a clean reboot) and noted the best score of 3 runs. According to some of my previous test data, OS 8.6 reported faster results with this benchmark.
RaveBench 1.1.1 Results:
As I've done with many Video Card reviews, I used Villagetronic's RaveBench 1.1.1 to test QD3D Rave performance. For an explanation of RaveBench's tests, see my Illustrated Guide to RaveBench published in 1998. (RaveBench is supplied with VT's video cards in the past, it's not available elsewhere that I know of and I can't distribute it so please don't ask).
I'm not sure why Intersections and Triangles were a bit lower with the Rage128 AGP card, but the variation is only a very small percentage of the total and not really significant.
Lightwork's Walker 3D Viewer is another QD3D/Rave benchmark I've used in past video card reviews. I tested two scenes the lowest and highest polygon (complexity) scenes and noted the minimum and maximum framerates seen during spin tests. Note the minimum framerate shows how well the system/video card handled the most complex part of the scene. Maximum framerates are listed in parenthesis in the graph below.
Memory Bandwidth Tests:
I want to caution readers to not be too concerned with the results below. Although there were some strange results in many cases, rely more on the real world applications tests on the next page of this review rather than the results of pure benchmarks like these. Despite the oddities shown below, the XLR8 control panel and extension (v1.42b0) was used in all applications and game tests and performed well. Real world application results matter far more than benchmarks.
I used 3 tools to test for memory performance. Memory Bench, Stream and Newer Tech's new Gauge Pro. In the images below you'll see results of a stock B&W G3/400, the B&W with the XLR8 G4/400 upgrade (at both 400 and 450MHz speeds) and a new Apple G4/450 AGP (Sawtooth) system. All systems were running OS 9 and had 256MB of PC100 RAM. The Apple G4/450 system had all 2-2-2 clock SDRAM. The B&W G3 had 3-2-2 RAM. Although I doubt it mattered, when a G4 CPU was present the 4 Altivec OS 9 extensions were enabled.
Since the Apple G4/450 AGP (Sawtooth) system uses the 'maxbus' feature of the G4 that is not available on any previous motherboard design, I was curious to see what effect this would have on memory bandwidth performance. Keep in mind that memory bandwidth test results don't have a direct relationship to real world application performance due to the efficiency of the L1 and L2 caches. This is why a 50MHz system bus speed older mac with the same speed CPU upgrade compares well in most applications tests to a 100MHz bus speed more modern Mac.
What I was most interested in was something I had seen in my first look at G4 CPU upgrades. With a 50MHz bus speed mac I had noticed a huge difference in L1 cache memory performance according to a memory bandwidth benchmark. This made no sense to me, as the G4 internal L1 cache should have no handicap from the slower system bus. I did see much higher L1 cache speeds with the G4/400 in a B&W G3 (100MHz bus) than when tested in a Genesis (50MHz bus). This makes no sense unless there is some quirk in Memory Bench (entirely possible).
Memory Bench Results:
First compare the Results of the B&W G3/400 before and after the G4/400 CPU Upgrade.
Note the massive increase in L1 cache speeds on data sizes up to 16K with the same speed G4 vs the G3 CPU (as much as 3x improvement at the same clock speed):
As I noted before, tests without the XLR8 extension active resulted in 159.6MHz cache speeds rather than 200MHz, but showed higher results for some reason. In real world apps tests like Infini-D rendering, performance was better with the XLR8 software enabled (400/200) vs 400/159.6 with no XLR8 software enabled. (Now you see why my reviews take so long to produce, as I run many tests and document many more variations than most other reviews on the net or in print.)
Memory Bench seems sensitive to repeat runs (results varied) so I always reboot before each run to ensure as consistent a memory/cache state as possible. In one test without the XLR8 extension active, opening the XLR8 Control Panel and setting the cache from 156.9 (default cache speed w/o the XLR8 extension active) to 199MHz (2:1 ratio) resulted a drop in rates to those seen with the extension active. So it seems just opening the control panel affected results the same as having the extension active. [XLR8's reply below addresses this observation] Remember that this is just one small, unusual benchmark and as mentioned, application tests were better with the extension/control active in a spot test of Infini-D rendering times.
The next results compare the B&W with XLR8 G4/450 vs an Apple G4/450 Sawtooth which uses the G4's 'MaxBus' design. Note the Sawtooth's maxbus design has higher L2 cache and main memory transfer rates than the B&W G3's motherboard, even when running the same CPU, Cache and Bus speeds. However as shown on the Applications tests page, you won't see much advantage from this with the majority of today's applications.
Note that although I didn't see the 800MB/sec that Apple claims - their comments of 2x higher bandwidth appears to be true according to the results above. Since the bus width and speed is the same as the B&W G3, the G4/Sawtooth 'maxbus' feature must be responsible for the higher IO rates.
Stream is one of the most popular and widely known memory tests. The images below show comparisons of the B&W G3/400 vs a G4/400 upgrade in the same system and then a B&W G3 w/G4/450 compared to the Sawtooth Apple G4/450 system.
Now compare the B&W G3 with G4/450 CPU vs the Sawtooth G4/450 AGP system
Again the Sawtooth's maxbus design shows a large advantage in _potential_ performance. However as shown on the Applications Tests page, the huge advantage in memory bandwidth is not really a factor in many applications tests. As mentioned previously, the efficiency of the L1/L2 caches isolate the system bus speed in most cases to where lower main bus speeds are not often a factor in real world application performance. In blind tests, you'd never tell the difference in these systems.
Newer Tech's Gauge Pro Results:
Newer Tech's latest Gauge Pro utility also includes a memory bandwidth test feature. What's interesting is that without the XLR8 extension active, the B&W G3 with XLR8 G4/400 upgrade showed significantly higher speeds than even the Sawtooth G4/450 system. Puzzling... with the XLR8 extension active results were lower as I'll show below.
Now the same test without the XLR8 Extension Active:
What is really odd is that just opening and closing the XLR8 control panel (no extension active, and making no changes to the settings) resulted in Gauge Pro showing the 95MB/sec scores again. [Again XLR8's reply after this review was posted provides answers for the reduction in memory bandwidth benchamrks with their control or extension active.]
Now Sawtooth G4/450 AGP System Results:
Same B&W G3 with XLR8 G4 @450MHz but without the extension active.
I'm baffled at why the Sawtooth G4/450/225 with its 'maxbus' design scored lower than the G4 upgrade at a slower speed (without any XLR8 software active) according to GaugePro. The Sawtooth's design delivered substantially higher scores in Memory Bench's tests as far as main memory bus speed and L2 cache speed however. Again these 'pure' benchmarks are interesting, but not directly related to real world application performance as shown on the next page of the review.
Update: After this review was posted, XLR8 replied to my questions on the lower memory bandwidth results with their extension or control panel active. Chris Cooksey XLR8's Director of Software Engineering replied:
"...the differences in performance you are seeing with the extension in and out, are because the extension is deliberately setting bits in some of the low level control registers to correct certain errata in the 2.6 (and earlier) revs of the G4. If I do not set these bits the machine will not run reliably.
|Summary: In other than some memory bandwidth tests, G4 upgrades score about as well as Apple's latest Sawtooth G4 system (and for B&W G3 owners, are far less expensive). Benchmark performance is interesting for technical reasons, however I prefer to rely on real world application performance since that's what really counts in the final analysis. Application and 3D Game tests are covered in the next page of this review. Or you may use the links below to jump to a specific page.|
- or -
Copyright © Mike, 1999.
Users of the web site must read and are bound by the terms and conditions of use.