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

Those who aren't familiar with electronical components (and don't want to get to) may skip over the following paragraphs which are mostly for hardware engineers and those who find themselves involved into repair jobs. So, the power supply stabiliser of the graphprocessor is built upon a 2-channel Intersil 6568 (ISL6568 &mdash PDF datasheet, 518Kb) controller located on the back side of the card. Indeed, this device uses PWM (Pulse Width Modulation) instead of linear regulation. Every channel is built of of three high-power N-channel field-effect transistors (FETs): one Infineon 59N03 (BSC059N03S &mdash PDF datasheet, 104Kb) and two Infineon 32N03 (BSC032N03S — PDF datasheet, 99Kb), also of a fast diode (40V, 5A) and a 1.0µH choke. There are two solid electrolytic capacitors installed on the input, Nippon Chemi-Con PS 330µF/16V, and three of the same kind on the output, Sanyo SVP 1200µF/4V. That's good to see them because capacitors with solid electrolyte are much better than with either hybrid or liquid one. In brief, they feature very low impedance (especially good at 200KHz to 500KHz range) and high rated ripple current, what allows for well smoothed output in real life. Additionally, these capacitors don't suffer from electrolyte dissipation, so their life span under normal temperature conditions (say, below 65 °C) is unlimited practically. However, there are some drawbacks as well when compared to products with either hybrid or liquid electrolyte: lower capacitance at the same physical dimensions and rated voltage, also significantly higher price. One must say that hybrid electrolyte is a compromise between solid and liquid ones as it includes both of them. However, it isn't of widespread use. In general, one aluminium capacitor with solid electrolyte if used in a low voltage PWM stabiliser may replace 2 to 3 capacitors with either hybrid or advanced liquid electrolyte given the same physical dimensions and rated voltage. Regular liquid electrolyte found often in no-name "low ESR" capacitors is of no comparison at all. The following table shows how well Nippon Chemi-Con PS 330µF/16V stands against another popular product of this manufacturer, KZG 330µF/16V which features advanced liquid electrolyte.
 

	Impedance at 100kHz, mΩ	Rated ripple current at 100kHz, mA RMS	Estimated lifetime at 65 °C, years
PS 330uF/16V	14	5050	22.83
KZG 680uF/16V	26	1540	3.65

Although electrolytic capacitors are the most popular for smoothing purposes, there are also tantalum capacitors with either standard or polymer cathode, niobium and niobium oxide capacitors, some kinds of film capacitors which feature low impedance over wide frequency range and long life span. However, they're outnumbered largely by multilayer ceramical capacitors (MLCCs) which demonstrate excellent resistive characteristics and the longest possible life span, though their capacitance is small. At the same time, manufacturers of modern high-end video cards tend to prefer MLCCs of 10µF to 22µF to solid electrolytic capacitors even if exchange ratio has to be a dozen to one. It needs to mention that the video card being reviewed hasn't got many large MLCCs to offer: there are 10 soldering places available on the PCB and related to the graphprocessor stabiliser, but only 2 of them are populated. Of course, it isn't that critical but does make a difference. By the way, this stabiliser receives +12V from and only from the auxiliary power supply connector, so the graphprocessor wouldn't even start if you forget to plug a cable in. The video card accommodates a piezoelectrical buzzer to make an untolerable noise in this case. That's nice to see that power isn't drawn from the connector directly, but through a LC-filter of a 1.0µH choke and an ELNA 47µF/16V capacitor to cut off some high-frequency noise produced mostly by high-voltage transistors of a power supply unit.
 
There are other two stabilisers available to feed the memory and HSI respectively. Each of them is based upon Intersil 6549 (ISL6549 — PDF datasheet, 323Kb) chip which consists of a PWM controller able to drive one channel of two FETs and of a linear regulator to take care of another FET. The PWM controllers only are enabled in this design. The memory stabiliser is built of two N-channel FETs by Alpha & Omega, AO4410 (PDF datasheet, 42Kb; placed between earth and the middle point) and AO4420 (PDF datasheet, 44Kb; placed between the middle point and the output), also of a fast diode (30V, 5A) and a 1.5µH choke. There are two Nippon Chemi-Con PS 330µF/16V capacitors installed on the input and one Sanyo SVP 1200µF/4V on the output. That should be well enough if to consider that power consumption of all memory chips is several times less than of the graphprocessor. Accordingly to the datasheet, that's approx. 11W in burst mode at 1250MHz and 1.8V supply. Although there is one thing to note. This stabiliser is fed from the auxiliary power supply connector through a LC-filter of a 1.0µH choke and an ELNA 47µF/16V capacitor very much like the previous one, but it's hooked up to a +5V line instead of a +12V one. The HSI stabiliser consists of two N-channel FETs as well: an unknown one labelled as RLA130 (placed between earth and the middle point) and an AO4422 by Alpha & Omega (PDF datasheet, 44Kb; placed between the middle point and the output), also of a fast diode (40V, 5A) and a 1.5µH choke. It accepts +3.3V from the AGP slot through a 1.0µH choke with a solid electrolytic Sanyo SVP 180µF/16V located after it. The output of this stabiliser is smoothed by a solid electrolytic Sanyo SG 510µF/4V. Every of these two stabilisers has been honoured with one large MLCC instead of five expected.
 

(click to enlarge, 186Kb)

It's weird a little to see 16V capacitors where 10V or even 6.3V would be satisfactory. That's a regular consumer-level hardware, not a server or military equipment where such overkills are usual matters. Whatever the reason is, let it be.
 
The graphprocessor, memory and HSI are supplied with 1.44V (1.4V is standard), 1.82V (1.8V is standard) and 1.43V (1.4V is standard) respectively, nothing is wrong there. Power supply quality has been evaluated with a portable 1-channel 12MHz oscillograph, Velleman HPS40. The following measurements have been performed under a heavy load (the Nature test from the 3DMark2001SE suite) to illustrate one of the worst possible situations.
 
 

(G70 power supply quality)

(memory power supply quality)

(HSI power supply quality)

So, HSI power supply quality is the best because AC part is less than 0.1% of average DC voltage received. The graphprocessor and memory are supplied significantly worse because their AC parts are 1.0% and 0.7% respectively. Those missing MLCCs would improve the situation for sure, but even now it may be considered as good.