Oblivion
Obviously, Crossfire works. Even with two high end cards such as the Radeon X1900 Crossfire Edition and the Radeon X1900XTX or (XT), you can still get an increase in performance, though not as impressive like we saw with a pair of Radeon X1600XTs. That's mostly because we're testing with a system limited setup, So, what happens when we're not system limited? Will Crossfire offer higher frame rates in the next generation of games? To answer this question, we put these two Radeon X1900 cards to the test in Oblivion.| Grass Test | Tree Test | |||
| Default | AA AF | Default | AA AF | |
| 1024 | ||||
| Min | 43.67 | 56.67 | 29 | 36 |
| Avg | 50.19 | 63.68 | 36.28 | 39.93 |
| Max | 62.33 | 80 | 44.33 | 46 |
| 1280 | ||||
| Min | 36.67 | 49 | 24.67 | 36 |
| Avg | 40.93 | 55.87 | 30.3 | 40.41 |
| Max | 52 | 73.33 | 36.67 | 46.33 |
| 1600 | ||||
| Min | 32 | 41 | 21 | 30 |
| Avg | 36.07 | 44.52 | 27.13 | 36.64 |
| Max | 43 | 55 | 34 | 43.67 |
Now, there's something strange here. It looks like at the moment, the current driver's Oblivion profile will only enable Crossfire AFR mode when the game is run with AA and AF enabled. We tried forcing AFR by renaming the executable and to some extent it worked. However, there's some display corruption that's quite annoying - flashing textures and background when you're looking at the journal or map. So, it's obvious that creating profiles for proper use of AFR is no easy task and a continuous work in progress.
Performance Limiting Factor
Well, we've seen what both Crossfire and SLI has to offer in terms of performance. However, the geek inside of us will not rest. Why does the two react so differently? After some research, we think we've found the answer - vertex shaders and / or polygon processin power. Look back to our article, SLI: Myth and Reality or any other article concerning SLI and Crossfire. When both Crossfire and SLI is enabled, the rendering load gets split equally between the card. That means only the pixel shaders units are used and the vertex shader units on the second card is idle the whole time. Look at the theoretical performance of these cards. Pay special attention to the Triangle and Fill-rate columns.| Core Clock (MHz) | Fill-rate (Mp/s) | Texture Fill-rate (Mt/s) | Triangle (Mtris/p) | Memory Clock (MHz) | Memory Bandwidth (GB/s) | Triangle/Fill Rate | |
| X1600 XT | 590 | 2360 | 2360 | 738 | 690 | 22.1 | 0.313 |
| 6600 GT | 500 | 2000 | 4000 | 375 | 500 | 16 | 0.188 |
| 6800 GT | 350 | 5600 | 5600 | 525 | 500 | 32 | 0.094 |
| 6800 GS | 425 | 3400 | 5100 | 531 | 500 | 32 | 0.156 |
| 7600 GT | 560 | 4480 | 6720 | 700 | 700 | 22.4 | 0.156 |
It's obvious which card has the most polygon pushing performance - the Radeon X1600XT and the GeForce 7600GT. However, the Radeon X1600XT is more 'asynchronous' compared to the 7600GT - the ratio of triangle and fillrate is actually double that of the GeForce 7600GT (and 6800GS). What does all of these mean? With Crossfire and SLI you effectively double the theoretical fillrate (a rough estimate - in real life it should be lower due to latency and other factors) that the system becomes vertex limited most of the time. Now, take a look at the table below to see what we mean.
| Core Clock (MHz) | Fill-rate (Mp/s) | Texture Fill-rate (Mt/s) | Triangle (Mtris/p) | Triangle/Fill Rate | |
| X1600 XT | 590 | 4720 | 4720 | 738 | 0.156 |
| 6600 GT | 500 | 4000 | 8000 | 500 | 0.125 |
| 6800 GT | 350 | 11200 | 11200 | 500 | 0.045 |
| 6800 GS | 425 | 8960 | 13440 | 531 | 0.059 |
| 7600 GT | 560 | 6800 | 10200 | 700 | 0.103 |
We simply multiply the fillrate and texture fillrate numbers by two, for simplicity's sake. Notice how the Radeon X1600XT in Crossfire not only has more fill rate than the GeForce 6600GT in SLI, but it also has roughly 50 percent more triangle processing power. So, it will deliver a higher increase both when we're fillrate bound and when we're geometry bound. Obviously, ATI is aware of this, so they beefed up the vertex / polygon processing power of their card to make them scale well with Crossfire. It's now clear this is most probably the reason why the entire GeForce 6 series don't offer much performance increase with SLI. Looking forward, it looks like NVIDIA took the lesson ATI gave them to heart and beefed up the GeForce 7600 to be less vertex limited. Indeed, it will be very interesting to test SLI on the GeForce 7600.
What about the high end parts? Let's take a look.
| Core Clock (MHz) | Fill-rate (Mp/s) | Texture Fill-rate (Mt/s) | Triangle (Mtris/p) | Memory Clock (MHz) | Memory Bandwidth (GB/s) | Triangle/Fill Rate | |
| X1900 XTX | 650 | 10400 | 10400 | 1300 | 775 | 49.6 | 0.125 |
| X1900 XT | 625 | 10000 | 10000 | 1250 | 725 | 46.4 | 0.125 |
| 7800 GTX | 430 | 6880 | 10320 | 860 | 600 | 38.4 | 0.125 |
| 7900 GTX | 650 | 20800 | 31200 | 1300 | 800 | 51.2 | 0.063 |
| Core Clock (MHz) | Fill-rate (Mp/s) | Texture Fill-rate (Mt/s) | Triangle (Mtris/p) | Triangle/Fill Rate | |
| X1900 XTX | 650 | 20800 | 20800 | 1300 | 0.06 |
| X1900 XT | 625 | 20000 | 20000 | 1250 | 0.06 |
| 7800 GTX | 430 | 13760 | 20640 | 860 | 0.06 |
| 7900 GTX | 650 | 41600 | 62400 | 1300 | 0.03 |
NVIDIA have also learned their lesson here - the GeForce 7900GTX is less vertex limited because it has a much higher clock than the GeForce 7800GTX (but retaining the number of vertex shader units). Now look back again to the table for the low end cards. While the GeForce 7600GT is 'only' about 1.5 the fillrate of the Radeon X1600XT, NVIDIA really wanted a win here. The GeForce 7900GTX has roughly twice the fillrate of the Radeon X1900XTX. Of course, that's theoretical performance and we all know those never tell the whole story.
[Previous Page]
[Go to top]
[Next Page]