Sapphire Radeon 9600XT 128Mb AGP: Advanced Overclocking

All right, it has been written a lot on how to do overclocking with voltmodding properly, now it's time to tell you what exactly has been done with this Sapphire Radeon 9600XT, and how fast it runs after all. So, the graphprocessor has been fed with 1.6V right from the start (two 1.0kΩ resistors), what allowed to advance right for 641.25MHz (here and forth all frequency increments are in quarters of the master clock which is 27.00MHz, i. e. in 6.75MHz quanta). However, almost all memory overclocking which has been obtained prior to voltmodding vanished somewhere. No matter how fast or slow the graphprocessor is clocked, the memory suffers greatly. This is a clear sign of overvoltage, so the RV360's stabiliser has been set for 1.45V (2.2kΩ upper and 2.7kΩ lower resistor). Memory performance has recovered to 371.25MHz, and the graphprocessor has been able to reach 614.25MHz. That's +60.75MHz from 553.50MHz obtained prior to voltmodding. Not bad, but who wants more? 1.53V (2.0kΩ upper and 2.2kΩ lower resistor) — 634.5MHz, but the memory has started to play tricks again, though not so hard as with 1.6V before. It seems that this RV360 isn't from the best batch definitely, and TSOP-II memory packaging isn't the best thing for high clock speeds, but we've got what we've got. It's time to fight for memory performance with increased voltage. Setting up Vdd for 2.75V (2.0kΩ lower resistor) and Vddq for 2.73V (2.2kΩ lower resistor). The memory has become stable at 364.50MHz, but the radiators on the back side of the card have been quite warm under heavy load. No more memory volts, let's play with the timings.

Unfortunately, even with the timings increased the memory has stuck at 364.50MHz. So, it has been only left to tune them properly for the best performance and reliability.

	tCAS	tRCDrd	tRCDwr	tRP	tRRD	tR2R	tWR	tR2W	tW2R	tW2Rsb	tRAS	tRFC
before	4	5	5	7	3	2	3	7	2	3	10	17
after	4	5	4	6	2	2	4	7	2	4	10	18

After all, we've got the graphprocessor at 634.50MHz and the memory at 364.50MHz (729.00MHz effective). A good overclock, but not that much to be really proud of. All right, let's check with 3DMark 2001SE v330 and 3DMark 2005 v1.2.0 how much the card has gained in means of performance if compared to the non-overclocked one. The card has been plugged into an ASUS SK8V mainboard (VIA K8T800 based) with an AMD Athlon 64 FX-51 2.2GHz@2.44GHz and two 512Mb Mushkin ECC registered DDR SDRAM sticks at 222MHz (444MHz effective) in so-called dual-channel mode. Windows XP SP1 with DirectX 9.0c and ATI Catalyst 5.10 drivers have been installed.

3DMark 2001SE v330 (1024 × 768 × 32bpp)
	Fill Rate		High Polygon Count		Nature
	S-texturing	M-texturing	1 light	8 lights	BF	16x AF	16x AF 4x MSAA
before	1020 Mtex/s	1970 Mtex/s	48.8 Mtr/s	13.3 Mtr/s	51.3fps	35.5fps	24.1fps
after	1280 Mtex/s	2490 Mtex/s	61.9 Mtr/s	16.8 Mtr/s	64.5fps	44.7fps	30.3fps

3DMark 2005 v1.2.0 (1024 × 768 × 32bpp with TF)
	Return to Proxycon	Firefly Forest	Canyon Flight	Overall Score
before	9.4fps	6.0fps	9.4fps	2030
after	11.3fps	7.2fps	11.4fps	2440

Well, there is some nice 20% to 25% improvement in performance on average. Not bad definitely. Although there have been no visual artifacts detected, the author has followed his regular practice of reducing clock speeds by one or two quarters of the master clock before putting the card to everyday's use or selling it away, so 621.00MHz graphprocessor clock and 357.75MHz (715.50MHz effective) memory clock together with the modified timings have been flashed into the card's BIOS. Soon after, this master-piece has been dumped to the author's junior brother for further testing with various gaming crap. No complaints after a week or so.