Archive 1997–2024

Intel Developer Forum Spring 2004 - Wrapup

More on PCI Express and Graphics

There is going to be a lot of fighting over the next few months about the performance of PCI Express based graphics solutions. Unfortunately, we won't really know what's what until we have hardware in our hands that we can play with to test real world performance. In order to try to straighten out some of the madness that will surely be going on, we will try to present a better picture of what things are looking like with PCI Express graphics cards at the moment. First we will attempt to explain why the contenders chose their paths.

ATI went with a native PCI Express interface because it gives them a full 4GB/s upstream and downstream bandwidth at the same time. This will allow for some massive amounts of data to move between the GPU and the CPU/main memory in both directions. They also have the advantage of not needing an extra component on the graphics card itself.

NVIDIA chose to use a bridged solution (which they like to call their High Speed Interconnect or HSI solution) which gives them the ability to only produce one GPU for both AGP and PCI Express based solutions while the transition is being made to the new platform. This gives them the advantage of being more flexible to demand for AGP and PCI Express based products, and they won't have to forecast just how many of which type of GPU they will sell in any given silicon run.

The main point of contention between the two camps is bandwidth. We know what ATI is capable of; now let's take a look at NVIDIA. Because the bridge is in such close proximity to the AGP interface of the GPU, NVIDIA is able to run the AGP side of their bridge at 4GB/s (which is twice the bandwidth of AGP 8x). As NVIDIA is bridging PCI Express's 4GB/s up and 4GB/s down to an AGP 16x bus, they will not be able to sustain the full bandwidth of the x16 PCI Express interface in both directions at the same time. If data is moving in only one direction, there is no bandwidth loss. If data needs to move up and down at the same time, one data stream will have to sacrifice some bandwidth to the other.

So, what kind of impact will this have? We can't really say until we have hardware. Of course, DDR400 will only give us 3.2GB/s of bandwidth (or 6.4GB/s in dual channel mode), so transfers from memory to the GPU will actually be memory speed limited on systems that don't have 8GB/s of memory bandwidth. As for games, we can take a look at the history of the impact of increasing AGP bandwidths. It is possible that future games (and possibly games ported by lazy console developers) may want to use the CPU and main memory a great deal and therefore benefit from PCI Express, but again, we won't know until the hardware and the games are out there.

NVIDIA's NV5x line is slated to be native PCI Express, but they also plan on continuing AGP support through that line by flipping their bridge chip upside down. In fact, we were told that if demand persists into the future, it is very possible that NVIDIA will bridge its NV6x GPUs back to AGP as well. This bus transition is a really tough one to call. ISA took a long time to disappear, and PCI graphics cards are still selling (generally to those who want two graphics cards, but the point is that they are still out there). The difference here is that Intel is pushing really hard for this transition, and AGP is a much more targeted solution than either the ISA or PCI buses.

From a business standpoint, NVIDIA has made a safer choice than ATI. Of course, if performance ends up suffering, or if ATI shows effective, affordable applications for the PCI Express bus that NVIDIA can't take advantage of, NVIDIA will suffer. In this business performance is everything. Obviously having the full bandwidth of PCI Express is a desirable thing. Eventually everyone will be making native PCI Express GPUs. But when is eventually? And when will that available bandwidth be tapped, let alone necessary? Only time will tell, but hopefully we've filled in some of the gaps until that time comes.

Originally published at https://www.anandtech.com/show/1248 (Wayback Machine).