Benchmarks

Theoretical Rates

Before we move on to analysing the performance of our two review samples, let us consider the mechanical aspects of them. As they both use the same RV515 chip, it is the clock speeds and memory bus widths that separate them; we shall also compare them to a reference X300 SE sample - this board is from ATI's previous generation of graphics adapters in the same market sector although it does not make use of the HyperMemory feature.

 With a 33% core speed advantage over the X1300 128MB model, the PRO BRAVO Edition board is the fastest of the three - both are clocked considerably higher than the X300 SE and so should display a distinct performance delta with the older model. In terms of memory bandwidth, the PRO board comes first again and although the other two both use a 64-bit memory bus, the X300 SE's memory is clocked at nearly half that of the X1300. The latter also has the advantage of being able to use system memory as local memory so the actual speed advantage should be even larger than the above theoretical figures suggest.

Fillrate Tests

One can examine the initial impact of memory bandwidth by testing the raw fill rates of the graphics cards; for this we used nDAW Interactive's benchmark:

As expected the PRO is the fastest here but if one compares it to the other X1300 sample, the bandwidth difference of 100% is quite apparent as the PRO produces almost 50% more raw colour fill rate and nearly 90% more fill rate when alpha blending single textures. Further evidence can be seen in the comparison with the theoretical outputs, where the colour fill figures differ by roughly 10% - interestingly though, the HyperMemory model is marginally closer to its potential when performing z writes. For both cards, the alpha blend and FP texture tests show that bandwidth is impacting on their capabilities but the large disparity to their expected throughputs possibly suggests that not all of the ROPs in the RV515 are capable of blending - whether this is the case or not will be investigated later in the review. The gap of more than 50% in FP texture figures is due to the fact that the chips cannot sample 16-bit texture components in a single cycle.

Although the process of texturing is becoming less significant in modern games when compared to the number of cycles required to process a pixel shader, it is still key to a graphics card's performance in a great many games. The X1300 PRO and HM boards differ on a basis of core speed once the number of texture layers exceeds 2 but this is to be expected given that there is only one TMU per pixel pipeline and therefore multiple core cycles are needed to sample and process extra layers. However, the jump in difference as the texturing goes from zero to 1 texture highlights the importance of memory bandwidth - the gap closes back up again when a second layer is added because of the extra core cycles required.