As a follow on to our March, 2004 article on PCI-Express, we have decided to do a series of articles better describing the differences and advantages that PCI-E has to offer. This months article will deal with the graphics portion of PCI-E technology and improvements that have been made over the previous generation graphics bus.

PCI Express Basics

As you may well know, PCI Express is a 3rd generation interconnect technology. What this essentially means is that PCI-E will replace PCI, AGP, PCI-X, and all other bus systems on the mainboards that you see today. PCI-E establishes a point-to-point connection, (or several point-to-point connections if necessary) from chip to chip, or chip to adapter on your motherboard. This full-duplex, dedicated, scalable lane increase bandwidth, decreases latency, and allows your devices to perform as they were meant to. A single lane (X1) of PCI-E has a bandwidth of 250MB/sec. By running multiple lanes, (up to X32), the bandwidth increases arithmetically.

PCI Express Graphics

PCI-E Graphics will generally run on a X16 PCI-E bus which will give it a whopping 4GB/sec of available bandwidth. When comparing this to our current standard of AGP 8x which is limited to 2.1GB/sec, you can see that this is going to be a phenomenal advancement in your graphics cards ability to receive and transmit data across your system. While all of the top level graphics cards vendors know that PCI-E is the future of the market, some are taking different paths to get there. The top tier players, NVIDIA and ATI have very different approaches to the same problem, how to get their cards out to market first, and what is the best solution for their customers.

ATI and PCI Express

ATI is taking PCI Express at it's purest form. They are designing their next generation of cards to support PCI-E natively and are busy re-working their current line of graphics cards to be compatible with the new standard as well. At launch, ATI's PCI-Express cards will mirror it's current line of high end AGP 8x Graphics cards. The only difference between the current cards and their launch offerings will be the on-chip interconnect interface. ATI describes their PCI Express cards as the "true" PCI Express solutions. While this is true, that their native implementation is strictly PCI Express, their claims that other cards which use bridged interfaces are more prone to failure, and will have performance bottlenecks and incompatibilities with software applications have yet to be proven.

Theoretically, the performance of a bridged PCI-E solution could suffer as it will be limited to a total of 4.2GB/sec bandwidth as compared to the 4GB/sec of bandwidth both in the upstream and the downstream directions for a combined total bandwidth of 8GB/sec, but this is unlikely due to the fact that the signals traveling in one direction will almost match on the two interfaces and that is all that really matters. That that being said, current applications do not even fully saturate the AGP 8x interface that is in use today.

As for their claims of their implementation being more cost efficient, I'm not sure that the cost of an additional chip on a card will be outweighed by ATI's need to completely redesign their interconnect portion of their own GPU's to make the transition to PCI-E. In addition, maintaining teams and facilities to design and construct both AGP and PCI-E cards, as ATI will need to do, can be very expensive. This can be balanced out by ATI's partner companies not needing to design around the bridge chip, thus leading to simpler (and less expensive) implementations of native PCI-E solutions.

NVIDIA and PCI Express

NVIDIA's strategy to introduce this new technology is to initially use a bridge chip. This chip will be completely transparent to the software which runs across the PCI Express bus. While the standard thinking is that if you add a bridge chip, you will lose performance due to the latency increase in the signal path, this may not be true with the NVIDIA cards. Initial testing has shown that NVIDIA's latency tolerance is higher than the latency of the chipset and the interconnect, thus preventing any lag from the addition of the bridge chip.

This is not unheard of, even recently with the introduction of SATA hard drives into the market, the majority of them are bridged technology... only one or two manufacturers are actually using a full native SATA implementations, and we aren't hearing any complaints about speed or ability of the non-native drives at this time. But, getting back to NVIDIA... by assembling their cards with their own bridge chip, the HSI (High-Speed Interconnect), NVIDIA is saving themselves a lot of time, trouble, and money. Reports have estimated that by using the HSI Chip at initial launch, NVIDIA will save itself over $20M this year alone. Not having to re-design all of their current GPU's to work with PCI-Express allows them to work faster and to concentrate on their upcoming cards.

The HSI chip allows AGP functions to be translated into PCI Express functions through an overclocked AGP function. When NVIDIA designed it's GeForce FX GPU's, they were made to sustain a minimum of AGP 12x interface speeds. The high-end cards, such as the GeForce PCX 5950 will use GPU's running at AGP 16x speeds. Although the NVIDIA cards will be initially utilizing the HSI bridge chip, they will still fully support the PCI Express advanced power management features that are built into native PCI Express cards. An advantage that NVidia will have, that ATI won't, is that, the HSI chip is reversible. This will allow NVIDIA to port it's future cards over to run on the AGP bus with hardly any effort or expense. This is a huge advantage when you think of how long it is going to take the public to switch over to the new PCI Express compatible motherboards. NVIDIA can continue to market their new cards to the non-upgraded crowd, while ATI will have to either design entirely new cards for them, design their own crossover chip, or forfeit the market segment to NVIDIA without a fight. We'll have to wait to see how it all works out.

The Big Winner

Who will come out on top in the PCI Express graphics wars? My prediction is that the consumer will end up being the ultimate winner. Healthy competition drives the market in both technological advances and in price. While NVIDIA will probably own the market for the next several months due to their brand new 6800 chip and their implementation of Shader 3.0 technology, ATI is not too far behind and we are eagerly awaiting the release of their next chipset to one-up the current leader.