The move, the second of its kind in recent months, indicates the seriousness with which the chip industry is responding to plunging PC prices. It further suggests that the new market — in which "free" seems to be the only final offer — may ruthlessly punish those who have made performance their mandate.

Specifically, S3 and Via plan to collaborate on integrated products that will combine the graphics controller, frame buffer and north-bridge functions of a PC on one chip. The expected product line will support microprocessors from both Intel Corp. and Advanced Micro Devices Inc., and the first chips could be available by the second half.

The move is both unprecedented and ironic for S3, which hasn't involved itself in core logic since a decade ago, when it began life as a multiprocessing-core-logic vendor. But it also marks a major shift in business strategy for serving the low end of the PC market.

"We used to develop high-end products and then let them trickle down through the various price segments as they got older," said Andrew Wolfe, chief technology officer at S3. "Now we have to develop unique products for every segment."

S3 isn't the only company walking that path. "Integrating chip sets and graphics for value computers is a definite trend, and it can be seen all over," said Nathan Brookwood, president of market-research firm Insight 64 (Saratoga, Calif.). He pointed out that the graphics accelerator and the microprocessor both require fast access to main memory, which makes the north bridge a logical component to merge with graphics because it also houses memory controllers. "The performance is not as good, but most PCs offer more performance than most users require, so this seems like a reasonable sacrifice."

Intel will follow suit later this quarter when it releases the 810 chip set. Code-named Whitney, the component will include a 3-D graphics engine and will target the low-cost computer market.

But the first vendor to seize the new initiative was S3 competitor Trident Microsystems Inc. (Mountain View, Calif.), which is already shipping an integrated core-logic/graphics device, the result of its own partnership with Via. The two companies have been working together for more than a year, and the first products were released last fall at Comdex. The devices merged the graphics controller and the north bridge and eliminated the need for a frame buffer, according to Abbas Razavi, Trident's director of marketing for sub-$1,000 PC initiatives.

Trident's version of the integrated device is designed for the notebook segment; Via released a desktop version. Both versions support the AMD Socket 7 architecture, and devices for Intel-based Slot 1 systems are due out later this quarter.

"In the sub-$1,000-PC market, performance has a lower priority, and cost is the most important factor," Razavi said. With PC prices diving below $500, saving even a few dollars on the bill of materials can have a much more significant impact today than a few years ago, when the PC price standard was in the $2,500 range.

"Integrating these functions and getting rid of the frame buffer allows the OEM to eliminate $20 in memory-chip costs and saves on-board real estate and power consumption as well," he said.

Razavi said discrete graphics components will always offer better performance and will remain popular in the hard-core gaming and enthusiast markets. But he predicted that integration will become more common in midrange systems as smaller geometries bring better imaging capabilities to even integrated graphics and core-logic devices. "We expect to see a plethora of integrated products coming out from different vendors by the end of this year," he said.

There is agreement among most vendors that the integrated parts — particularly if they also eliminate a dedicated frame buffer — can save money. But the question of their impact on performance is a complex one. One reason is that the new designs eliminate the frame buffer by summoning up one of the hoariest monsters to stalk the PC graphics business: the unified memory architecture.

In the UM architecture, the graphics controller uses a portion of main memory for its frame buffer. Thus, both drawing operations and frame-buffer reads to refresh the screen must go through main memory, along with cache fills and empties and PCI-bus activity.

No one would suggest that the unified memory approach is the route to high-performance gaming 3-D. Separate graphics chips on add-in boards with large amounts of dedicated memory will continue to dominate that market. But it would also be unwise to dismiss the performance that can be had with the UM architecture.

By placing the DRAM controller, the memory-bus-arbitration logic and the graphics controller on the same chip, integrated north-bridge designs are able to give the graphics engine far better access to main memory — both in low latency and in high burst rates — than would be possible with a separate graphics chip. They can also reorganize and shape memory accesses coming from the graphics controller and the peripheral bus, reducing the number of time-eating page misses caused by mixing graphics and cache traffic on the same bus.

Designers are calculating that those advantages will make up some of the raw bandwidth given up by switching from dedicated, wide synchronous-graphics-RAM frame buffers to shared main memory. And they have an ace up their sleeve as well. That is a careful study of just how normal applications use memory bandwidth.

"In most systems, it turns out that memory isn't the bottleneck," said Dean Hays, director of marketing at Via. "As long as you have enough overall memory bandwidth, a frame buffer in main memory can be OK. Yes, you can tell the difference if you devise a test where you have everything blasting at the same time — a DVD decoder running, video going through the frame buffer, and intense calculations — but how realistic a scenario is that?"

To make sure the bandwidth is there, Via, in its design for Trident, took one added precaution: Engineers designed their DRAM controller to exploit NEC's Virtual Channel DRAM. "It really makes a big difference," Hays said. "Virtual Channel DRAM isn't necessary to make the system run well, but it does make a difference."

Peter Glaskowsky, senior analyst for MicroDesign Resources (Sebastopol, Calif.), said an integrated approach could also work for the notebook market, where power consumption and board space are critical factors. Putting core logic on the same die as graphics, Glaskowsky said, is "fundamentally a better deal in most technical respects because it provides a better match of the real-estate value for the two subsystems."

A trick that probably won't get tried anytime soon is an integrated north bridge with on-chip embedded DRAM for a frame buffer, or even a video cache. "One of our big advantages has always been that you can build system logic chips, and even very good graphics chips, in processes that are one generation behind the leading-edge memory and CPU guys," Via's Hays said. "We enjoy the benefits of great yields and depreciated fabs. But if you start playing with embedded DRAM, you are in a complex, expensive process again, and you lose that advantage."

Potentially, the integrated chips could meet the real needs of many entry-level computer buyers. One factor that may limit the trend, though, is upgradability.

Niles Burbank, desktop-components product manager for ATI Technologies Inc. (Thornhill, Ontario), said his company's strategy is to focus on developing discrete chips for the high end of the PC marketplace while recycling proven components in the lower PC tiers. By continuing to shrink the components, today's low-end accelerators are often faster than their state-of-the-art predecessors.

Burbank said the company chooses to avoid integrating its components with core logic because that would limit flexibility and upgradability. "At the very low end of the market it makes sense to integrate, but that doesn't offer the flexibility of a discrete component," he said.

Glaskowsky agreed, saying, "End users don't care about integration per se, but the lack of flexibility and upgradability may reduce the perceived value of integrated graphics products, and they will remain stuck at the low end of the market."

That might not be that bad a place to be stuck, however. The median price in the PC market is spiraling down like a struck duck, concentrating volumes — and the chances for growth — at the low end. Even corporate buyers are showing increasing interest in the low-end machines — which, ironically, often meet the needs of corporate users at least as well as they meet the needs of first-time home users.

Should that trend continue, it may be the systems with integrated north-bridge chips that constitute the market's center, as expensive, standalone 3-D chips with superb rendering performance are relegated to the add-in card market.

That, in turn, could upset the always tentative balance of power in the PC-graphics business. Vendors such as S3 and Trident have not kept up with the performance leaders in gaming 3-D. That means they have not devoted huge amounts of engineering and lavish amounts of silicon real estate to top-end 3-D rendering. Hence they are in a better position to integrate their graphics cores into an inexpensive north bridge.

For the 3-D performance leaders, such an integration may well prove much more difficult. The result could be yet another switching of leadership positions as volumes shift toward the low end.

-Additional reporting by Ron Wilson