AnandTech: ASRock Core 2 Duo: AGP/PCI Express Graphics Performance, Part Tres

AGP and PCI Express Performance

In our first article we compared the performance of DDR versus DDR2 on the ASRock 775Dual-VSTA motherboard and found there was very little difference between DDR-400 and DDR2-533 on this platform. In our second article we compared the performance of DDR against DDR2 on several different platforms that included the VIA PT880 Pro, Intel 865, Intel 945P, Intel P965, and Intel 975X chipsets. Our results showed that while there were differences in memory performance between each chipset and speed setting, it mattered little in the overall performance of our system. This was mainly due to our selection of mid-range components that likely would be used when upgrading to our motherboard and processor choice.

Our article today will look at the performance differences between AGP and PCI Express graphics cards on the ASRock 775Dual-VSTA motherboard with Intel's E6300 Core 2 Duo. We will state up front that our article today is not a video card review. Instead we are verifying if AGP performance on our test motherboards is acceptable when comparing it to PCI Express performance on the VIA PT880 Pro chipset used on our ASRock test platform. Our tests today will be utilizing the EVGA 7600GS and 6800 Ultra series of video cards in both PCI Express and AGP configurations. Both series of cards offer decent performance that is well suited for 17" or 19" LCD monitors running at resolutions up to 1280x1024. While both cards will struggle with current games such as Oblivion -- a game that will bring most systems to their knees -- they still offer a fair amount of performance for games released the past couple of years and can certainly handle any normal desktop application work with ease.

The typical user who will purchase a motherboard of this type or the AGP only ASRock 775i65G are primarily concerned about extending their current component investments while upgrading to the latest Core 2 Duo processor series from Intel. These component investments usually include AGP graphics, DDR memory, along their current power supply, storage, and optical drives. Based upon this profile our tests will utilize DDR memory only and video cards that represent typical performance in the mid to lower range of graphics capability at this time. Additional graphics performance results when utilizing DDR2 memory combined with other chipsets can be found in our last article.

The ASRock 775Dual-VSTA motherboard offers AGP 8X/4X capability along with PCI Express X4 graphics performance. However, our PCI Express video cards will theoretically be at a disadvantage due to the bandwidth differences between X4 mode and our AGP 8X capability. We would like to see if the difference in bandwidth affects performance, though realistically we're only measuring the performance of one specific design as opposed to AGP versus PCI Express; we cannot say for certain that the PCI Express implementation of the PT880 chipset is fully competitive with other PCI Express implementations. Still, it is possible that the X4 slot will saturate the PCI-E bus, which can certainly occur in some cases based upon graphics settings and applications.

AGP 8X has up to 2.1GB/s of shared bandwidth. The typical GPU can make use of nearly all the bandwidth, but the upstream bandwidth isn't as important. Various requirements for upstream transfers end up limiting the maximum upstream throughput to around 266 MB/s, and switching back and fortch between reads and writes can incur a further performance penalty. Typical desktop applications tend to utilize the downstream bandwidth (read from system) the greatest amount of time while upstream bandwidth (write to system) is seldom used and is rarely an issue. The design of AGP took this into account, which is why the realizeable upstream bandwidth is so much lower than the downstream bandwidth.

PCI Express X16 operation has 8GB/s of theoretical bandwidth that is segregated for upstream and downstream paths due to its serial bus design. Technically, the PCI-E bus transmits at 2.5 Gbps on each link, so 2.5 Gbps x 16 = 40 Gbps. Converting to bytes that gives 5 GB/s, but like most serial buses there is a 20% transaction overhead that reduces the useable bandwidth to about 4 GB/s. Thus we arrive at the result of 4 GB/s of read and write speed for each direction at maximum bus capacity. Once again, the majority of application usage is spent on the read side with the write side capacity being wasted for the most part in current PCI Express graphics systems, but there is the potential for it to become more important in the future, and SLI/CrossFire implementations using the PCI-E bus to transmit data certainly benefit.

Since our test board is operating in X4 mode we have one fourth of the bandwidth available which equates to about 1000MB/s upstream and 1000MB/s downstream. This is slower than AGP 8X since the available read bandwidth is limited and can be saturated by certain applications, making AGP 8X potentially more effective. PCI Express X4 offers about half the available read bandwidth of AGP 8X, and the write performance advantage of PCI-E goes largely unused in most applications.

Let's see if this theory holds true in our test results -- again, recognizing that we are only comparing performance on one specific chipset.