Here I have to add that in order to get PCI Express x16 bus fully supported, the drivers for the video subsystem as well as for the overall platform need to be revised thoroughly. However, this process hasn’t yet been completed. The fully-fledged support of the PCI Express x16 is most likely to appear only with the Longhorn announcement. All this results into the fact that contemporary graphics cards with PCI Express x16 interface cannot take full advantage of the new graphics bus, so that we will really benefit from this innovation much later.
However, we can notice a few improvements of the practical bandwidth of the bus between the chipset and the graphics card even now. Even PCI Express x16 graphics cards from NVIDIA with HIS bridges and AGP 8x bus “inside” the graphics cards, still can use the higher bandwidth of the new bus. To connect the graphics chip and the HSI bridge in NVIDIA based graphics cards they use AGP interface overclocked to AGP 16x. The bandwidth of this bus is about 4GB/sec, which is exactly the bandwidth of PCI Express x16 in one direction. In other words, this solution may lead to data transfer rate drops only in case of the duplex mode, i.e. when the data is transferred along the PCI Express x16 bus in both directions simultaneously. As for the conversion of the data from PCI Express into AGP format, NVIDIA claims that there is a maximum of 3-5% latency worsening in this case.
However, it is very simple to find out if these theoretical suppositions are actually true or not. We have a special utility at our disposal, which allows measuring how fast the data from the main system memory can be written into the graphics memory. Due to this small software tool developed by Andrew Filimonov, who is also the developer of our brand name Xbitmark test, we can evaluate the efficiency of PCI Express x16 implementation in ATI and NVIDIA based graphics cards. For this test we measured the data transfer rate in the i875 based platform equipped withy an AGP 8x NVIDIA GeForce FX 5900XT, and the data transfer rate along the PCI Express x16 bus of the i925X Express based platform equipped with the ATI and NVIDIA based graphics cards supporting PCI Express x16 interface. In this case we used NVIDIA GeForce PCX 5900 and ATI RADEON X600. The results of this experiment are given on the diagrams below:
As we see, the tremendous growth of the theoretical bus bandwidth between the graphics core and the chipset doesn’t result into a tremendous practical data transfer rate boost. However, I haven’t expected anything else, to tell the truth. The raw software for the PCI Express x16 bus limits the maximum bandwidth increase provided by the new bus to 40% during data transfer towards the graphics card. As for the data transfer rate in the opposite direction, we will see a more significant increase in the practical bandwidth here. Also note that ATI’s solution without the additional bridge-chip looks more attractive. RADEON X600 transfers data along the bus much faster than GeForce PCX 5900.
Here I would like to point out the following issue. Despite the fully-fledged duplex mode supported by the PCI Express x16 architecture, we see that the data transfer rate from the graphics card is considerably lower than the data transfer rate to the graphics memory by both: ATI solution with the native support of this bus as well as by NVIDIA solution with the converter bridge-chip. But this peculiarity is a distinguishing feature of the AGP bus, and it actually shouldn’t occur in PCI Express x16 solution. Therefore this strange observation may give you food for thought especially as far as the “fairness” of the native PCI Express x16 implementation in ATI chips is concerned. However, the most likely explanation in this case is either the issue with the PCI Express x16 implementation or the problem with the drivers.
However, it would still be a way too early to expect that the graphics cards supporting new bus interface would ensure any increase in the gaming performance, even if the PCI Express x16 had been properly implemented from the very beginning. AGP bus is pretty slow compared with the local graphics memory. Besides, it has been out there long enough for the game developers to arrive at an undeclared conclusion that they should avoid transferring data along this slow bus at any rate. Therefore, the graphics accelerators store as much data they might need for building the frame as possible in the local graphics memory. That is why the effect provided by the higher bandwidth of the bus between the graphics accelerator and the chipset will be minimal today. On the other hand the launching of the new graphics bus with high bandwidth and low latency has every chance to ruin the stereotypes so that in the near future software developers will no longer avoid transferring the data along the new PCI Express x16 bus. Then we will probably get a better chance to evaluate the efficiency of the new PCI Express x16 technology.
One more indirect advantage of the new PCI Express x16 bus is a more powerful voltage regulator circuitry implemented in it. It can even accept 12V power lines, and the maximum load it can bear reaches 75W. Due to this outstanding fact, many graphics cards, which have been requiring an additional power supply connector, have every chance to give it up once and for all. For example, the NVIDIA GeForce PCX 5900 and ATI RADEON X600 we tested this time didn’t require any additional power supply.
Having introduced the new PCI Express x16 bus with its new i925 and i915 chipsets, Intel gave up backward compatibility. The new chipsets do not support AGP 8x that is why most mainboards based on these sets of core logic will have no AGP 8x slot onboard and will require new graphics cards. However, some mainboard manufacturers are still going to release solutions based around i925/i915, which will be equipped not only with the new PCI Express x16 but also with the old AGP 8x slots. In this case, you should keep in mind that the AGP slot on these boards is implemented via the PCI bus, which limits its performance dramatically and affects negatively the graphics accelerator performance.