Preparing for PCI-Express
Looking at ATI's approach towards PCI-Express, we see that their solution is relatively simple and straightforward. Out of the box, all ATI graphics cards will have native support for PCI-Express. As a result, all VPU's will have the full features of this new standard at their disposal. As was previously outlined, this includes a full 16 lanes and an enormous 8GB/s of total bandwidth. In contrast, a generic bridged solution will likely lose a great deal of these features and functionality. Here, such a generic bridged solution introduces a great deal of complexity into the design which ultimately increases latencies and inhibits overall performance. ATI has outlined this and more in a recent press release which can be found (in full) here. For the sake of conserving space, we have chopped it up into a more manageable size by including only the bare-essentials.
Approaching the first point which is made above, we see that ATI is illustrating the fact that they are the first to offer a fully native solution for PCI-Express. However, the ensuing statement being directed towards their competitor are false given the information we have been given. Rather, this claim seems to be directed towards a generic bridge solution as the listed characteristics largely do not apply to NVIDIA's HSI chipset. The main advantage ATI has over NVIDIA's HSI solution is the ability to support a full 4GB/s upstream bandwidth which NVIDIA cannot. The following page will outline this technology in-depth and illustrate that the theoretical performances between the two solutions is a lot closer than we are led to believe. The HSI bridge has nothing to do with the physical compatibility with PCI-Express and we will show that it is not constrained to AGP speeds. Lastly, neither format will introduce any variables which would inhibit software compatibility. Per usual, this will be entirely dependent upon that vendor's driver quality alone.
Moving to the next topic, we see that ATI is claiming that a truly native solution will have better reliability due to lower complexity and fewer failure points. In the case that it is a truly native solution developed around PCI-Express, this is certainly the case. However, to make things a bit more complicated and dramatic a very reliable source has provided us with information which questions whether ATI's current solution has brought bridged complexity within the core. Supposedly, the following photos compare an AGP-based RV360 mobile core (M11) to a native PCI-Express RV380 mobile core (M24). Comparing the two images, we find it difficult to distinguish one core from the other. However, if you look at the left-most edge of each core you will see the only appreciable difference between the two. With no confirmation from ATI thus far on what the purpose of this section of logic is or even if these are the cores in question, we cannot make a judgement. If possible, we will update this article with an appropriate response from the company after the appropriate products are launched.
You can click on the following image for a larger image.
Rounding out the final two points, we come across power management. On the desktop side, this is much less an issue. However, for mobile solutions ATI's support for PCI-Express' power management functionality will likely offer a significant advantage to their product. At this time, we have not heard of any such plans from NVIDIA for their next-generation mobile product. Lastly, we have the issue of cost. Since our readers want to buy cards to go play them, we will spare the discussion concerning inventory risks and chipset ASP's. In the end, each vendor will be hitting the same price-point meaning you won't be paying anything extra regardless of what choice you make.
One aspect to consider here is the fact that the VPU must be designed to take advantage of all the features and functionality PCI-Express brings to the table. In some cases, you might find that the VPU architecture is a bottleneck long before the performance features of PCI-Express ever come into play. To illustrate this scenario a bit clearer, imagine dropping a 400bhp engine into a Ford Focus. Although there is an incredible amount of horsepower available, the car was never designed to handle that kind of power so you are left with a broken axle and an empty wallet. However, if you put that engine into a Ferrari or similar exotic car you have yourself an excellent combo with spectacular results as the car was made to deal with this significant power. With this in mind, it will be interesting to see how ATI's family of VPU's deals with native PCI-Express support. Although the benefits are assured on the higher-end models in the future, the benefits for the current generation and lower end/mainstream products remain to be seen. Then again, those Canadian engineers are crafty and are always capable of surprising us.
Although the immediate success of ATI's native implementation remains to be seen, there are some very exciting possibilities on the horizon which could take advantage of the enormous bi-directional bandwidth of PCI-Express. With this bandwidth at their disposal, developers can now begin exploring methods of offloading the CPU by bringing some tasks to the VPU. One example we were provided with is assigning the handling of physics calculations to the VPU instead of the CPU. Here, the architecture of the VPU will likely be able to process these calculations faster than the CPU would. As a result, games which were once CPU-bound now have more headroom and performance should increase accordingly. In addition, developers can also explore methods of controlling artificial intelligence with the VPU instead of the main processor. Again, this will likely add a healthy boost in performance to a taxing game engine. Despite being excellent additions to gaming, we must remember that this is all a work in progress. No titles on shelves today can take advantage of these features, so we are once again faced with a timing issue. The true question here is for the mainstream user or the hardcore gamer, will we see any immediate and noticeable gains from PCI-Express with the next wave of hardware or is this transition a bit premature?