VisionTek GeForce3 - Overclocking
|Complexity and Clockspeed
First off, VisionTek does not recommend or support overclocking. their products. But, I did it anyway. *grin*
The drop in core clockspeed from the GeForce2 Ultra to the GeForce3 is a bit deceptive. With the addtion of 22 million transistors, the chip is able to perform more instructions per clock cycle. The addition of the Lightspeed Memory Architecture also permits for more efficient dispatch of information out to the memory chips which are running at the same speed as the Ultra. Moreover, four textures can now be applied in one cyle instead of two.
With that taken into consideration, my first assumption was that the GeForce3 would not overclock well, and even if it did, it would not help performance significantly because the added complexity would create new and unseen before bottlenecks to performance. Was this assumption correct?
For the most part, no.
Alienware Area 51:Aurora (Modified)
- AMD Thunderbird at 1.1GHz
- Abit KT7A-RAID KT133A motherboard
- 256MB PC133 Virtual Channel CAS2 SDRAM
- 15GB ATA66 7200rpm HD
- Sound Blaster Live! X-Gamer
- Linksys 10/100 NIC
- MidiLand S4 8200 5.1 Sound System
- Martian Red Case and Keyboard
|Let's Start With Quake III Arena
The usual torture test of Quake3 timedemos (1.17 demos converted to 1.27) was first on the batting order. The GeForce3 is fast. 1600x1200x32bpp is now a comfortable gaming resolution, with framerates higher than the Ultra at 1280x1024x32. I was giddy. 16x12x32 ran at around 55 frames per second with the sound on at extreme visual quality settings. Framerate seems more consistent.
I started bumping the core clock and found that as the core increased in speed, so did the framerates. I had grown so used to seeing little or no change because of memory bandwidth restrictions, I had to make sure I was moving the right slider in the driver controls. How far did I get?
How does 245MHz sound?
It is shocking to see a chip as complex as the GeForce3 already be able to sustain a 25% increase in clock speed. The increase equated to roughly a 5% gain in framerate.
For the memory, I reset the core and started increasing the memory speed. This is where I expected to see the big leap in performance. While I did see a increase in framerate with every hop in clockspeed, I did not see the rapid increase in performance we see on the GeForce2. This is good news/bad news situation. While the GeForce3 is more balanced, it appears that its efficiency has negated some of the performance gain when upping the memory bandwidth.
The VisionTek GeForce3 hit 545MHz before locks plagued gameplay.
I saw a 7% increase in overall framerate when at 545MHz. Memory overclocking is as atributable to the memory chips and board design as it is to the quality of the manufacturer's fabrication. Nice work, VisionTek.
Independent testing had run its course. Time to increase both clocks at the same time.
I am happy to report that increasing both clocks at the same time results in faster increases in performance than each clock alone. At 245MHz core and 545MHz memory, the card proved to be around 15-20% faster than default.
After Quake3, I went to 3D Mark 2001 which should stress the all components of the GeForce3, including the shading and vertex capabilities. My first attempt at the 245/545 clock settings was successful, but the second run failed. I dropped back down to my current default seting of 240MHz core and 540MHz memory.
It seems as though the additional stress of using more of the additional 22 million transistors was enough to push the chip a bit harder causing locks on clockspeeds that were safe on the "less complex" Quake3. To verify my new settings were solid, I unleashed the Dronez rolling Demo. After two hours with no locks, I was comfortable leaving the VisionTek GeForce3 at 240/540.
Tables, charts, screenshots and analysis will round out the first part of the VisionTek GeForce3 review. Keep in mind that my results should not be compared to your results, because I am an image qulaity freak that leaves the sound on for benchmarking and and doesn't shutdown and restart between runs. These numbers merely the real world increase in framerate I experienced when increasing the clockspeeds.
1024x768x32bpp is now slightly CPU limited. The more fillrate intensive settings show the 15% to 20% increase in framerate. Later in the review, I will look at ways of using this increase for additional image quality. Thought I would give you a screenshot of 1024x768x32bpp with Quincunx enabled, my current preferred setting for gaming.
3D Mark 2001...
A surprise to me was the continuation of the 15% to 20% increase into the DirectX 8 intensive sybthetic. I was sure that this benchmark would find a new bottleneck but the GeForce3 proves to be a scalable chip design. Rumors abound of a GeForce3 Ultra, and with these results, why not?
Dronez - Default
Dronez - Overclocked
Quick note, the 4.19 minimum framerate in the non-oced screenshot was a fluke. While I could not reproduce the low minimum, the average was the highest of four runs by one tenth of a frame per second. It's pretty obvious the Dronez benchmark, view settings here, benefits in several ways from the increased clockspeeds.
The fact that the GeForce3 achieves increased framerates by overclocking the core only conveys that the GeForce3 is far more in sync with its surroundings than the GeForce2. Why would the GeForce3 exhibit this behavior?
My theory is that as the clock speed increases, so does the ability to complete performance enhancing calculations such as early Z-Buffer checks. The core is able to delegate information over the crossbar to the memory more often in a given period of time. The fact that we see the GeForce3 get very close to theoretical fillrate limits in synthetic benchmarks is a sign of the chips ability to skirt potential wait-states as information is fetched from storage. Prior to the GeForce3, 3D chips behaved in a much more linear manner. The GeForce3's complexity seems to thrive on the oppurtunity to perform more calculations per second, instead of becoming frustrated, and tripping over itself trying to get information to the memory and the screen.
As for the memory clock, when given more oppurtunities per second to trade information with the memory, the GeForce3 does so willingly. Put both together and framerates hop to stratospheric levels.
Despite the young architecture and insane number of transistors, the GeForce3 is acting like a mature and refined chip. The delay for the A5 re-spin was worth it and NVIDIA's engineers should be commended for designing and refining the NV20 in time for release when faced with the extremely compressed six month product cycle.
VisionTek also deserves compliments for the level of quality of the VisionTek GeForce3's board. I have had many cards that were fresh off the line and this is by far the highest percentage memory overclock I have ever seen in a first run circuit board.
Stay tuned for more review.
Next up, I will look at how VisionTek's GeForce3 offering impacts gaming in its current state and take on the question of "Should I upgrade?" I might even work in a Q&A with VisionTek.
Back to nV News