When I had the opportunity to get a G4 533MHz CPU card with heatsink from a Digital Audio G4, for only $40, I jumped on it. I originally did not think it would fit into my G4 correctly, but I read that Digital Audio CPU's could be put into older G4's by removing the motherboard's ATA-33 connector (which I didn't really need anyway).

I was wrong. It fit. Turns out the G4 466 and 533MHz CPU's are 7410's, similar to the old 7400's, and on the same size card. It's the 667 and 733MHz 7450 models that are too big to fit in a Sawtooth. I was also able to use the DA's larger heatsink, because my G4 is in an ATX case.

So, after installing the CPU, I started changing the multiplier. The 533MHz CPU only ran at 400MHz, at first, because its multiplier was 4x. This makes 533MHz in a 133MHz bus DA, but only 400 in my Sawtooth. The procedure for changing the multiplier can be found elsewhere Ð it is beyond the scope of this article. I used auto rear window defroster repair paint.

I reached 600MHz, and was quite happy. Then I tried 650, and was even happier! 650 seemed okay, until it froze after about 5 minutes of Folding@home. Hmph. On rebooting, I got a kernel panic. Sigh, no good. I let it sit and cool, and on subsequent boots, it would alternately kernel panic, and freeze shortly after reaching the desktop. If I was quick, I could use PowerLogix's CPU Director to back the L2 cache down to 260MHz, where it was stable, but benchmarked about the same as 600MHz with 300MHz cache. The L2 cache chips are Samsung, and end in -QC27, therefore they are 275MHz rated.

I re-visited Michiro Isobe's fantastic G4 speed settings article (how does he do it?), available at: http://www.xlr8yourmac.com/G4ZONE/sawtooth/SawtoothCPUdesign.html. (this and other similar articles are linked on the Systems page) There, I read the procedure for L2 cache voltage boosting. The formula looks like this:

L2OVdd = 1.260 x (R51 + R52)/R51

I plugged this into Pacific Tech's Graphing Calculator for OS X, and started playing around with some values. Here's what I found:

A 100 (Ohm) resistor at the R51 position will yield a setting of about 2.82v, up from the original 2.5v. I just happen to have a 100 (Ohm) resistor - they're very common. Also, this setting is within Samsung's 2.9v spec for these chips per Samsung's K7A403609A datasheet PDF file.

Soldering onto such small solder pads is difficult, but by no means impossible. It does require a certain amount of practice. I just popped the original 124 (Ohm) resistor off with a fingernail (quick & dirty, but it works), cleaned off the solder pads, then soldered the two leads of my 100 (Ohm) resistor onto the pads. Time for a test.

Success! Booting and totally stable at 325MHz L2 cache, for about 20 minutes with Folding@home running. Then, it froze. Rats. At least this time I could reboot without panicking. But it would seemingly either freeze or panic after about 15-30 minutes of use. Still no good. I booted into OS 9, ran Gauge Pro's Memory Testing, and sure enough, there was some memory corruption going on.

There was one last thing to try, I figured. Perhaps the cache chips were overheating with the voltage boost - after all, they're back-to-back, and un-cooled. So, I grabbed an old heatsink from an Intel 440BX Northbridge, cut it (roughly) to the appropriate size, lapped it with some 320-grain sandpaper, and used (non-conductive) thermal paste to create a good contact between the heatsink and the cache chip on the underside of the card. I then used hot glue (wonderful stuff) on two of the corners to ensure that the heatsink would not fall off the card when the paste got warm and softened up.

I used this particular heatsink on the back of the card because it's just the right height to clear the motherboard. I put some thin cellophane packaging tape on the motherboard directly underneath, just in case there might be an accident.

I decided to boot the machine with this new attachment and see what happened. Success again! It now appears to be totally stable at 650MHz, with a 325MHz L2 cache. It passes Gauge Pro's Memory Testing in OS 9 perfectly after 20+ passes, and I am yet to have a freeze or kernel panic in OS X. This is a 21.875% overclock on the CPU, and a 18.182% overclock of the cache chips - all in all, pretty respectable.

I am surprised that adding this small amount of cooling to only one cache chip allowed for so much more stability - then again, it was nearly stable to begin with, and I am operating in a pretty cold environment, in a case with good airflow. I am currently working on an aluminum shim to attach the cache chip on the front side of the card to the CPU's main heatsink. This should help it survive in temperatures above my basement's current chilly 60°F. (The perfect overclocking environment) The CPU's heatsink usually reads about 76.5°F via an external thermometer, running Folding@home 24/7. There is an 80mm fan running at ~7v screwed to the heatsink, blowing air across from front to back, through a homemade air duct.

XBench Scores:
Results			70.91	
	System Info		
	Xbench Version	1.1.3
	System Version	10.3.7 (7S215)	
	Physical RAM		896 MB
	Model			PowerMac3,1
	Processor		PowerPC G4 @ 650 MHz
	Version		7410 (Nitro) v1.3
	L1 Cache		32K (instruction), 32K (data)
	L2 Cache		1024K @ 325 MHz
	Bus Frequency	100 MHz
	Video Card		ATY,R200
	Drive Type		ST380021A
	CPU Test		77.57	
	GCD Loop		69.36		2.71 Mops/sec
	Floating Point Basic	96.87		350.33 Mflop/sec
	AltiVec Basic		53.54		1.56 Gflop/sec
	vecLib FFT		77.64		1.21 Gflop/sec
	Floating Point Library	122.56		4.91 Mops/sec
	Thread Test		60.34	
	Computation		41.78		564.02 Kops/sec, 4 threads
	Lock Contention	108.57		1.36 Mlocks/sec, 4 threads
	Memory Test		77.88	
	System			81.08	
	Allocate		477.64		311.56 Kalloc/sec
	Fill			69.53		553.42 MB/sec
	Copy			48.73		243.63 MB/sec
Stream			74.93	
	Copy			74.47		544.37 MB/sec [altivec]
	Scale			73.99		546.05 MB/sec [altivec]
	Add			73.82		472.42 MB/sec [altivec]
	Triad			77.54		473.80 MB/sec [altivec]