1. Foreword. The First Impression.
2. Cooling System. Basic Components.
3. Setting Up and Overclocking.
4. Synthetical Tests (3DMark 2001SE and 2005).
5. Direct3D Games (Far Cry, Serious Sam 2).
6. OpenGL Games (Quake 4, Prey).
7. Conclusions.

The testing job has been performed on the following hardware:

• processor: AMD Athlon 64 FX-51 (Sledgehammer, rev. C0, 130nm SOI) 2.44GHz (11x222MHz)
• mainboard: ASUS SK8V (VIA K8T800, Socket 940, BIOS rev. 1002)
• memory: Mushkin High Performance 2x512Mb PC3200 DDR SDRAM ECC registered 222MHz
  (tCAS=2, tRCD=3, tRP=2, tRAS=6, tRC=8, tRFC=14, tWR=2, tR2W=1)
• HDD: Western Digital 800JB (IDE, 80GB, 8Mb of cache memory)
• DVD-RW drive: TEAC (actually, Lite-On) DV-W516G (IDE, 2Mb of cache memory)
• sound card: Turtle Beach Montego II PCI (Aureal Vortex 2)
• PSU: InWin 430W (IW-P430A2-0 — 32A at +3.3V, 35A at +5V, 18A at +12V)

Nothing outstanding these days, but about 3 years ago this configuration has been on the top of gaming performance. However, most of AGP-based systems aren't any better, so why not? Let's see how fast it handles modern tasks.
 
As you may have guessed already, this hardware has been overclocked. Standard settings of the processor are 2.20GHz (11x200MHz) and 1.50V core voltage. This is the fastest processor of the K8 core rev. C0, so it hasn't been an easy task to make it running significantly faster with neither water nor phase-change cooling. Firstly, all major capacitors related to the processor stabiliser have been replaced: 3 Nippon Chemi-Con KZE 1200µF/16V liquid electrolytes for 3 Sanyo WG 1800µF/16V liquid electrolytes with 1 Fujitsu RE-SU 330µF/16V solid electrolyte installed additionally; 6 OST RLX 1500µF/6.3V liquid electrolytes for 6 Rubycon MBZ 2200µF/6.3V liquid electrolytes with 3 Sanyo SP 560µF/4V solid electrolytes and 3 EPCOS 220µF/10V tantalum ones installed additionally. However, it has been a question of about 50 MLCCs (10µF ones mostly) installed wherever possible to make the processor really happy with 1.70V core voltage. Although there were several thick copper wires placed on the mainboard's back side to decrease voltage losses even further. The mainboard's BIOS doesn't allow to go over 1.70V, and the author hasn't got neither time nor intention to play with earth levels of the processor's PWM controller to work around this obstacle. By the way, many of the other electrolytic capacitors on the mainboard have been also replaced with better ones. Secondly, Thermaltake Big Typhoon air cooling system (copper base with 6 heat pipes coming to 140 aluminium ribs and a 120mm dual ball bearing fan at 1300rpm) chosen for this processor has undergone an upgrade with a Titan TFD-12025H12B 120mm dual ball bearing fan at 2200rpm, and the resulting 2-fan monster has been taken together with 70mm screws and 5mm fluorine-plastical pads. Still quiet, but much more effective now. Thirdly, the processor's heat spreader has been removed, and the cooling system's bottom surface has been polished thoroughly in order to achieve a proper contact. As a result, a significant decrease in the maximal core temperature (from 58 °C initially to 52 °C with two fans and to 42 °C with the bottom polished and no heat spreader; that's in winter, add up to +15 °C for summer). In fact, there is no sense in removing a heat spreader and attaching a rough radiator's surface afterwards. Fourthly, some memory timings have been reduced from those set by default, even the overclocking hasn't been an problem — those W942508CH 5ns memory chips by Winbond are good ones, it hasn't been necessary even to increase their voltage from default 2.6V. Despite all efforts, the HyperTransport bus speed has got to be reduced in the BIOS from 800MHz to 600MHz (666MHz with the overclocking), otherwise the whole system would hang up deadly from time to time even under low software activity. Both ECC mode of the system memory and Cool'n'Quiet mode of the processor have been disabled to maximise performance. VIA KT8T800 system logic doesn't support AGP/PCI lock feature, unlike K8T800 Pro (in theory at least), thus speeds of the AGP and PCI buses have risen from 66MHz and 33MHz to 74MHz and 37MHz respectively, though it hasn't caused any trouble.
 

(the testing system)	(the mainboard's back)

 
A non-localised 32-bit Windows XP Professional SP1 with DirectX 9.0c (4.09.0000.0904) has been installed, also NVIDIA ForceWare 93.71 and ATI Catalyst 7.1 drivers with their settings as the following:
 

NVIDIA ForceWare 93.71
Anisotropic filtering	Application-controlled
Anisotropic optimisation	Off 1
Anisotropic sample optimisation	Off 1
Antialiasing settings	Application-controlled
Conformant texture clamp	On
Extension limit	Off
Force mipmaps	None
Gamma corect antialiasing	Off
Multiple graphics card acceleration	Single display performance mode
Negative LOD bias	Off
Texture filtering	High quality
Transparency antialiasing	Off
Trilinear optimisation	Off 1
Triple buffering	Off
Vertical sync	Force off
1  After switching Texture filtering to High quality these optimisations get disabled.

 

ATI Catalyst 7.1
Anisotropic filtering	Application managed
High quality anisotropic filtering	Disabled
Anti-aliasing	Application managed
Adaptive anti-aliasing	Disabled
Mipmap detail level	High quality
Catalyst A. I.	Disabled 2
Wait for vertical refresh	Always off
2  This setting disables internal filtering optimisations.

Only control panel options of both drivers have been tuned. No manual system registry tweaks or any third-party software have been involved in order to change any hidden options. Therefore, all video cards by NVIDIA and ATI about to be tested have been set up to compete fairly.
 
Sapphire Radeon X1950 Pro 512Mb AGP appeared to be almost non-overclockable. The maximal clock speed with reliable perfomance have been estimated at 594.00MHz of the RV570 and 361.13MHz (1444.50MHz effective) of the memory. Such little improvements (+2.3% and +2.9% respectively) make no sense of testing the card overclocked. A few more words about the cooling system. Temperature of RV570 running at the default clock speed has been able to reach 78 °C (20 iterations of the Nature test from the 3DMark 2001SE suite). No need to comment on this any further.
 
BFG GeForce 7800GS OC 256Mb AGP has been chosed as the primary competitor. See BFG GeForce 7800GS 256Mb AGP: Review and Testing to learn more about this card. In brief, it's based upon a NVIDIA G70 graphprocessor which has been manufactured using a 110 nm technological process. There are 16 pixel/texture pipelines, 6 vertex pipelines and 8 ROPs active. All of them are clocked at 398.25MHz, no geometric delta. A 256-bit wide data bus is used to access 256Mb of GDDR3 SDRAM clocked at 313.88MHz (1255.50MHz effective). The card has been overclocked to 480.94MHz for the G70 and 355.73MHz (1422.90MHz effective) for the memory. No cooling system change, voltage modifications or adjustments of the memory timings, just some regular overclocking.
 

(click to enlarge, 113Kb)

(click to enlarge, 109Kb)

Not so much time ago Sapphire Radeon X1600 Pro 256Mb AGP has been the fastest AGP solution from ATI/AMD to feature support for SM3.0. It's based upon an ATI RV530 graphprocessor which is manufactured using a 90 nm technological process with low capacitance dielectrics. There are 4 texture pipelines with 4 ROPs, 12 pixel pipelines and 5 vertex pipelines. The graphprocessor is clocked at 499.50MHz by default. There are 256Mb of DDR2 SDRAM clocked at 202.50MHz (810.00MHz effective) accessed through a 128-bit wide data bus. The card has been overclocked regularly to 580.50MHz for the RV530 and to 216.00MHz (864.00MHz effective) for the memory.
 

(click to enlarge, 99Kb)

(click to enlarge, 115Kb)

Finally, the author has dug up a Sapphire Radeon 9600XT 128Mb AGP. This 4-year old card will participate in our testing to show what it can do these days. Believe or not, but at the moment of writing such a brand new card could be purchased with no trouble. There is an ATI RV360 graphprocessor comprising 4 pixel/texture pipelines with 4 ROPs and 2 vertex pipelines, which has been manufactured using a 130 nm technological process with low capacitance dielectrics. There are also 128Mb of GDDR SDRAM accessed through a 128-bit wide data bus. Just like the RV530 reviewed above, this RV360 is clocked at 499.50MHz. The memory runs at 297.00MHz (594.00MHz effective). The card has been overclocked regularly to 553.00MHz for the RV360 and to 371.25MHz (742.50MHz effective) for the memory. See also Sapphire Radeon 9600XT 128Mb AGP: Advanced Overclocking
 

(click to enlarge, 129Kb)

(click to enlarge, 128Kb)

Let's get back to Sapphire Radeon X1950 Pro. This card has been pushed right to its limits. There is a lot of evidence to prove this statement including poor overclockability mentioned above. In particular, after passing through all tests, the author has decided to replace the cooling system with Accelero X2 by Arctic Cooling tuned a little for better efficiency. It has done the job really well: even at 50% fan duty cycle the RV570 temperature hasn't got over 50 °C. The bad thing is that Accelero X2 doesn't fit the design as it should. While the original cooling system provides adequate cooling to power FETs (those two VT1165SF), Accelero X2 can only offer some light air-flow in their direction. It needs to mention that these silicon packaged and BGA mounted power FETs can get very hot in run-time. Just believe, you don't want to put your finger on them even for a few seconds. So, several days after the card stopped to function properly in 3D, video overlay has gone as well, though it has been just fine in 2D. Numerous artifacts have appeared in a few seconds after launching any 3D application with further loss of video signal at all. VPU Recover has been unable to help. Reliable performance has been achieved again about at 200MHz of the RV570. A quick investigation has shown that one VT1165SF gets very hot even at low loads while the second remains just warm. Even more, the first FET has lost one of its corners. It seems that there has been a microcrack, and high temperatures have accelerated the lethal end. The diagnosis has been confirmed by a multimeter: Rds(off) of the failed FET has figured into 1.8kΩ while of the second FET — 4.5MΩ. In fact, the stabiliser has lost one of its channels with a result of much higher voltage pulsations on the output and half as low current limit allowed. That's why the graphprocessor could operate reliably at so low clock speed only. Of course, the dead FET should be replaced, but it would be nice to find out its specifications first. To prevent the second FET from following the first one on its way to computer gods, a construction shown below has got to be built and installed quickly. Appears to be not bad though. At least, it hasn't got any worse since. By the way, it doesn't seem to be possible to disable the failed FET because after loosing input voltage in the channel (+12V indeed) the stabiliser would disconnect all outputs. No documentation, no help.