As we know, ATI had a perfect reference point in front of them when they were launching RADEON 8500 graphics solution. GeForce3 Ti500 from NVIDIA was a very good chance to assess their powers.
Naturally, they had to provide the newcomer with the highest core and graphics memory frequencies possible, in order to compete worthily with such a serious rival. Eventually, ATI RADEON 8500 reached 275MHz/275MHz (550MHz) chip and memory frequencies, while according to some unofficial sources, RADEON 8500 based cards were originally expected to arrive with 220MHz/250MHz (500MHz) frequencies.
This fact gives us a serious cause for concern. Actually, we as well as ATI get every reason to be quite doubtful about the overclockability of RADEON 8500 chips. Remember that we observed no breakthrough in our tests of NVIDIA GeForce3 "overclocked" by NVIDIA engineers to GeForce3 Ti500 (see our article called "NVIDIA GeForce3 Ti500 and Ti200 Based Graphics Cards Extreme Overclocking").
Vcore and Vmem Increase
We managed to find 7 individual regulators on ATI RADEON 8500: 5 of them were responsible for powering the graphics core and memory. They are exactly the ones we are interested in greatly. In comparison with NVIDIA GeForce3 based cards, which are usually equipped with only 2 regulator chips, increasing Vcore and Vmem on RADEON 8500 seems a really difficult task. But as we found out later, it will be no obstacle for a true overclocker :)
As far as we understood, the core of RADEON 8500 is powered by 3 voltage regulators. Let's consider them one by one:
1.3V voltage regulator. It is based on LM2636 chip from National Semiconductor. This microchip serves as a controller for direct voltage pulse converters and voltage regulators. It is mainly used to CPU voltage regulators. The Vout of the regulator is determined by a 5-bit code output to the VID0..VID4 Ins of the microchip and can vary between 1.3V and 3.5V.
In our case the In of VID4 received a logical 0, i.e. VID4 is grounded, and none of the other Ins is involved. However, since all these Ins are connected with the power pole via resistors inside the chip, they have a logical 1. According to the VID code the default Vout of this regulator for RADEON 8500 is 1.3V.
We changed the input VID code by connecting VID1 and VID2 to VID4 and got 1.6V, as it follows from the specs:
1.6V voltage regulator. Here we have CS51031 chip from On Semiconductor acting as a controller for pulse regulators. The Vout is determined by resistors marked as R1 and R2 on the photo. According to the manufacturer's specs, it can be calculated with this formula: Vout=1.25x(1+R2/R1).
The R1 and R2 resistors from the divider are located at the rear side of the card right in front of the chip. We shunted R1 with an accessory 2.7kOhm resistor and got 1.95V Vout for this regulator.
1.9 voltage regulator. This regulator is built of discrete components. One of them is a powerful field-effect transistor from Fairchild Semiconductor, FQD20N06. The other one is marked as "31L", but we failed to identify it. Nonetheless, we found out that R1 and R2 resistors determine the Vout of this regulator. We shunted R1 with an accessory 2.7kOhm resistor and the regulator Vout turned 2.1V. The too dense components mounting around this regulator on the front side of the card inspired us to track the location of the conducting paths and to put the resistor onto the rear side of the card. We soldered one of the pins to a metal spot on the rear side of the card (on the photo it is marked with a red cross), and the other one - to the minus output of a huge electrolyte capacitor located near this regulator.
3.5V voltage regulator (VDDQ) of graphics memory. Like the previous regulator, this one is based on discrete components arranged in the same way. The Vout here is determined by R1 and R2 resistors. We shunted R1 with an accessory 10kOhm resistor, so the Vout became equal to 3.15V.
3.3V voltage regulator (VDD) of graphics memory. This regulator uses CS51031 controller from On Semiconductor. The Vout is determined by R1 and R2 with the resistances obtained from the following formula: Vout=1.25x(1+R2/R1). We shunted R1 with an accessory 10kOhm resistor and got 4.1V Vout.
Cooling System Improvement
In order to cool down the graphics core we took Mini Copper Orb cooler from Thermaltake. We decided on this solution because the design of the graphics card doesn't allow any standard coolers like Blue Orb or CPU coolers with a wide flat foot.
The relatively small copper foot of Mini Copper Orb was an ideal choice, but before installing it on the card we were to get rid of some of its components:
We also removed an outstanding band, which rimmed the foot and didn't let the cooler get tightly pressed to the core.
Furthermore, we sacrificed two old coolers for Slot1 CPUs to construct the heatsinks for the graphics memory. We sew away the side parts of those heatsinks and fastened them to the graphics memory chips.
Each of these heatsinks could cover two memory chips at a time, but unfortunately it was too early to install them. RADEON 8500 has some capacitors between the memory chips, so they appear in the way when you try to install this self-made heatsink. For this reason we had to drill out special holes in the heatsinks for those components:
After we introduced a new cooling system, the card got the following outlook:
We overclocked RADEON 8500 core up to 325MHz, but the card didn't work very stably at this frequency and hung after working for long with heavy workload. Accordingly, we had to decrease the core frequency. The ultimate top frequency of the core, at which the card worked was stably, totaled 320MHz.
As compared to the nominal core frequency of RADEON 8500 (275MHz), the gain amounted to 45MHz or 16%. In fact, the figure is not that impressive against the overclockability of NVIDIA GeForce3 core which provided a 70MHz gain to the nominal 200MHz or 35%.
It feels as though ATI originally clocked its RADEON 8500 chips to nearly the highest frequencies possible, whereas the overclocking potential of NVIDIA GeForce3 hadn't been entirely exhausted until GeForce3 Ti500 arrived. At the same time, GeForce3 Ti500 chips boast lower overclockability than RADEON 8500 chips: the difference makes about 12% (they could be overclocked from 240MHz to only 270MHz).
As for the graphics memory, it worked at 730MHz-740MHz in overclocked mode, though we decreased the working frequency down to 704MHz (!) for the sake of absolute stability. So, 704MHz - that's what the highest acceptable frequency for the graphics memory proved to be.
We should pay tribute to ATI: that was the fastest graphics memory we have ever tested, though some graphics cards happen to be equipped with equally fast or even faster memory with 3.5ns or 3.6ns access time. However, it was only RADEON 8500 graphics memory that took advantage of the card's quality design to overcome the 700MHz bar.
The last thing we'd like to mention here are the criteria used to assess the card's stability at higher frequencies. A stability criteria for us was the card's ability to pass a complete set of 3DMark 2001 tests in 32bit color modes with 1024x768 to 1600x1200 resolution without any artifacts.
In order to run all these tests we configured the following testbed:
- AMD Athlon XP 1500+ CPU;
- Gigabyte GA-7VTX (VIA KT266A) mainboard;
- 2 x 128MB PC2100 DDR SDRAM CL2by Nanya;
- IBM DTLA 307020 20GB HDD.
- Driver version 7.20 for Windows 9x;
- 3DMark 2001;
- Quake3 Arena v1.27;
- Windows 98 SE.
We tested ATI RADEON 8500 in the following modes:
- At nominal frequencies - 275MHz / 275MHz (550MHz);
- At 300MHz / 300MHz(600MHz) like ATI FireGL 8800;
- With an ultimately overclocked core - 320MHz / 275MHz (550MHz);
- With an ultimately overclocked graphics memory - 275MHz / 352MHz (704MHz);
- With both core and memory overclocked to their maximum - 320MHz / 352MHz (704MHz).
For this Quake3 Arena test we took demo127.dem and ran it in 32bit color mode with the maximum image quality settings:
It appeared more fruitful to overclock RADEON 8500 graphics memory in Quake3 Arena: the performance almost doubled, which can be clearly seen on the diagrams. In 1024x768 mode the performance is limited by the CPU and the overall system performance, so there turns no significant performance gain with an overclocked graphics card.
These gaming tests were run in High Detail modes. Color depth, Z-buffer and texture quality were set to 32bit:
In Car Chase test the results remain almost unchanged as the resolution grows, while the random elements introduced by the developers lead to a notable divergence in figures. These facts make it impossible to analyze the performance gain provided by overclocking RADEON 8500 in this test.
Dragothic includes a good deal of geometric calculations. The reason is that the dragon and rider models are drawn very thoroughly and comprise a big number of polygons. Besides, their animation is implemented with vertex shaders. That is why in Dragothic RADEON 8500 core reveals higher overclockability than the graphics memory. Overclocking the core and graphics memory to their maximum we get an almost 20% performance gain.
It's noteworthy that at resolutions ranging from 1024x768 to 1600x1200 we observed very little change of the performance gain. This implies that the load is mostly laid upon the graphics card, therefore this test depicts the performance of the graphics card, but not that of the CPU or memory subsystem.
In Lobby we can trace the same tendency as in Quake3 Arena: the performance in 1024x768 mode is restricted by the system speed, but as the resolution gets higher the performance gain increases. Subsequently, it turns out more fruitful to overclock the graphics memory.
The core overclocking in this test is most efficient at 1024x768. But at higher resolutions higher graphics memory bandwidth is needed, that is why it appears the overclocked memory that ensures most of the performance gain.
The Nature test loads the graphics card really a lot. This can be seen well on the diagram showing the performance gain during overclocking: as the resolution grows, the performance gain stays practically the same.
In our ATI RADEON 8500 and RADEON 8500LE Graphics Cards Review we assumed that the lower results of RADEON 8500 in this test (as compared with NVIDIA GeForce3 Ti500) come from less smart memory controller of RADEON 8500. It seems like today's tests confirm this assumption. The overclocked graphics memory of RADEON 8500 provided a much higher performance gain than that generated by the overclocked core.
As we saw, extreme overclocking can make the performance of RADEON 8500 increase by 20%-25% in Quake3 Arena and 3DMark 2001. We dare say that in none of the today's or even upcoming games someone will ever get a decisively greater performance gain.
Having taken NVIDIA GeForce3 as an example to follow, ATI clocked up its RADEON 8500 trying to squeeze all the juices, so its overclockability proved lower than we had expected.
The outcome of ATI's desire to bring up the clock frequencies to the maximum is that far not all the chips managed to reach new horizons. Those, which failed, were marked as "LE" or even "double LE". On the one hand, it's no good: cards of the kind are weaker than ATI RADEON 8500. But on the other hand, graphics cards like that are a lot cheaper and can be overclocked any time. For instance, we overclocked the fastest version of ATI RADEON 8500 and squeezed some 20%-25% from it.
Speaking about RADEON 8500 architecture, extreme overclocking showed that in most tests graphics memory is more efficient to overclock, i.e. in most cases the performance of RADEON 8500 was restricted by the graphics memory bandwidth. Bearing in mind that faster graphics memory chips are entering the market, ATI is free to give up synchronizing core and memory and launch even faster solutions with the graphics memory working at 700MHz or at even higher frequencies by default.
We wouldn't dare state with 100% certainty if extreme overclocking of RADEON 8500 is a worthwhile thing for everyone. It's up to you to decide whether to take the risks or not. We only tried to demonstrate the overclocking potential of these graphics cards.
- This research is just a kind of experiment and shouldn't be taken as an appeal to taking up extreme overclocking and graphics cards resoldering.
- These modifications shorten your card's service life.
- Any sort of mechanical modifications deprives the users of the warranty.
- Should the graphics card or other components be wrecked, the users bear the complete responsibility for their actions.