The specifications
Now that this clarification is made, lets move to the cardīs concrete characteristics:
The only modification to the Radeon X1900 for calculation units is an increase in pixel shading pipelines. The rest remains unchanged compared to the X1800. We have developed three criteria, which are important for us in the comprehension of GPU performances. The first is fillrate, and the Radeon X1800īs was already higher than the 7800 GTX 512 MBīs. As the X1800 is restricted by the memory bus, itīs the same for the Radeon X1900.
The second point we evaluate is texturing power. NVIDIA already was 30% ahead, but ATI didnīt think it was necessary or found it too costly to better equip the X1900 on this level. NVIDIA keeps this advantage.
The third area is mathematic calculation power. Here, we counted 2 instructions per cycle for the GeForce as compared to one for the Radeon. It is important to point out that in reality the situation is much more complex. Nevertheless, it appears that the Radeon X1800 had an important deficit in terms of calculation power compared to the GeForce 7800 GTX and an enormous gap separated it from the 7800 GTX 512 MB. With the Radeon X1900 ATI has decided to close this gap to avoid NVIDIA having too big an advantage for complex pixel shaders which will be used in future games. This is an important point, because today we tend to think that the Radeon X1800 is comparable to the GeForce 7800 in terms of calculation power, while it actually is not. ATI had to increase the X1800īs calculation power to equal things out. Of course, NVIDIA still keeps an advantage for texturing units and ATIīs architecture isnīt perfect, but how can it be when there are so many factors?
Finally, you will see the very small difference between the Radeon X1900 XT and the XTX model, +4% calculation power and +7% in memory bandwidth. This is a small improvement and doesnīt justify the price difference. The first is announced at 549 and the second 649!
Pixel shader : calculation power
We graphed X1800 XT, X1900 XTX and 7800 GTX 512 MB calculation power and left the calculation units efficiency as the unknown parameter (instead of the pixel shading pipeline efficiency) for more meaningful data. Each of the 24 pixel shading pipelines of the GeForce 7800 can process 2 mathematic instructions or one texturing instruction + 1 mathematic instruction per cycle. For the X1800 and X1900, each pipeline, whether itīs for pixel shading or texturing, can process one instruction per cycle.
The graph represents calculation power according to the
mathematic instructions / texturing texturing ratio:
The four GeForce 7800 GTX 512 MB graph represents the number of registers used, because calculation power is dependant upon it. The upper one is the best case and the driver integrated compiler tries to stay there.
On the left, there are more texture accesses than other instructions. This happens in two different situations. There is the relatively theoretical cases that donīt interest us here, and others where filtering and texture accesses require several cycles. Overall this is the same as using several instructions. Maximum filtering, anisotropic 16x and trilinear filtering requires 32 cycles. Fortunately, complete filtering is only applied to a minimal number of pixels. When this anisotropic filtering is selected, maximum filtering could be used but the GPU is in charge of simplifying the filtering for pixels that donīt need the maximum one (sometimes more than what is required, to gain some more FPS). NVIDIA has the advantage here.
At "1" there is one texturing instruction for one mathematic instruction. In this case, X1800 and X1900 calculation power is similar, but the GeForce 7800 GTX 512 MB finishes first. Once the ratio increases, the X1900 takes the lead and reaches its maximum when there are three mathematic instructions for one texturing instruction. As you may have noticed the advantage over NVIDIA isnīt that large. What the graph says is closer to "the Radeon X1800 has a calculation power deficit" than to "the X1900 has a huge advantage".
Pixel shading performance
What are the results in practice? To begin with we show the use of complex shaders of three applications; 3DMark05, Far Cry and Tomb Raider AOD. They represent average size shaders that are found in current games and were executed on the entire screen in an external application. We used the Radeon X1900XT and not the XTX model to be able to judge the difference between the X1800 and X1900 based on GPU cores clocked at the same frequency.
For the first two shaders, the 7800 GTX dominated the X1800. Thanks to the Radeon X1900, ATI takes back the lead for one of them. For the third shader that relies more on dependant texture accesses (indirections), the Radeon X1800 was already ahead. The X1900 only provides slightly better results.
Next was the two pixel shader tests of 3DMark06. The first is complex and the second, Perlin Noise, very complex with approximately 500 instructions.
In the first test, the X1800 was slightly behind while the X1900 finishes just in front of the 7800 GTX 512 MB. In the second, the X1900 increases the performance gap and shows its affinity for very complex pixel shaders as we might have expected.
Finally, we tested two lighting shaders that represent a very good compromise between theoretical and practical calculation transfer rates:
The first thing we noticed, was the X1800īs calculation power deficit (as explained above) and that the two 7800 GTXs were clearly ahead. In the first test, the X1900īs results are 2.3 times better than the X1800, finishing 33% in front of the 7800 GTX 512 (23% when it uses FP16). In the second test, the Radeon X1900 XT is 2.6 times faster than the X1800 XT. The end result is 55% and 33% better, respectively, than the 7800 GTX 512 depending on FP32 or FP16 use.
