ATI Reads Into the Future
As I mentioned on the previous page, ATI drastically increased the pixel processing power of the X1900 GPU by tripling the number of pixel pipelines from 16 to 48. However, they did not increase the number of vertex shaders (still at 8), nor did they increase the number of texture address units (still at 16) or render units (also at 16). At first glance, this might seem like a big oversight on ATI’s part, and that the R580 would be hindered technically because of it. But that doesn’t seem to be the case in our initial benchmarks and perhaps ATI did their homework after all.
ATI’s basic reasoning behind their 48 pixel pipes to 8 vertex pipes and 16 texture address units comes down to the proportion of shader operations that correspond to each. In general, shader operations can be split between two classes: they are either texture operations that fetch data from memory or they are arithmetic operations that perform mathematic manipulations of data in the form of pixels. Back when shaders were first introduced, the split between the two was fairly even.
However, as this ATI provided graph shows, the trend is shifting heavily towards arithmetic shader operations over texture operations. As such, the need for more and more pixel processor power, in the form of ALU units, is needed to play the latest games and those on the horizon. ATI has now given their flagship X1900 a 3:1 ratio of arithmetic unit to texture units (pixel pipes to texture units) and is basically making a prediction that the future of PC games will be more and more dependent on pixel processing than they will be on data consumption. This could end up being a risky venture, as NVIDIA has a much tighter ratio on their pipeline divisions and obviously see game development going in a slightly different direction.
Another point ATI brings up in their X1900 white paper is that increasing the number of texture units on a GPU can be “fruitless” with increases in the memory bandwidth and memory capacities to be utilized by the texture units. If 16 texture units are already filling up the bandwidth that the ATI ring-bus memory controller can handle, then adding 8 more really wouldn’t improve performance proportionally as you would expect. It’s the law of diminishing returns at work.
General Purpose GPU Usage
Another interesting note is the fact that the people that have really been pushing the GPGPU (general purpose GPU), such as Standford’s Folding@Home members have been pushing for more pixel shaders, as they are the units most important to parallel mathematical computing. Could GPU physics processing be far behind?
Xbox 360 Similarities
It is interesting to note, though it really gets us very few answers, that the Xbox 360’s GPU designed by ATI also uses a 48 unified pipeline similar to what we see here on the R580. The Xbox GPU’s unified shaders indicate that they are capable of doing BOTH arithmetic and data fetches and thus the comparison breaks down pretty quickly.