Product profusion portends a return to past form

Based on inaccurate information from the vendor, I recently incorrectly stated that ATI Technologies was building its Radeon 9800 graphics accelerator on 0.13-micron technology (see "Graphics power plants expand their integration grasp," EDN, April 3, 2003, pg 16). ATI in fact bases Radeon 9800 on the same 0.15-micron process it used on the earlier generation Radeon 9700. The company derived the 9700-to-9800 performance gains exclusively from circuit and layout optimizations, not from a lithography reduction.

That said, even though the Radeon 9800 Pro employs a comparatively trailing-edge manufacturing technology, numerous published benchmarks give it the image quality and performance nod over Nvidia's 0.13-micron-based GeForce FX 5800 Ultra (see "Graphics flurry may leave you blurry," EDN, March 20, 2003, pg 16). Nvidia has been busy doing its own fine-tuning, though, and the company believes that, with its GeForce FX 5900 Ultra, it'll recapture the crown (Picture). For how long, though, is unclear; ATI is undoubtedly working on a 0.13-micron version of the Radeon 9800 architecture, now that the 0.13-micron-based Radeon 9600 proof-of-concept is in production.

Whereas GeForce FX 5800 Ultra, which has yet to—and may never—enter high-volume production, ran at 500-MHz core- and memory-clock speeds, GeForce FX 5900 Ultra employs yield-friendlier, 425-MHz core and 450-MHz memory clocks. What, then, is the source of Nvidia's optimism regarding its latest progeny? For one thing, GeForce FX 5900, like Radeon 9700 and Radeon 9800, moves from its predecessor's 128-bit local-memory bus to a 256-bit-wide frame-buffer interface. And, as Intel demonstrated with its Pentium M CPU (formerly known by its Banias code name), raw clock speed is only one factor in overall performance; equally important is how much work the processor can do in each clock cycle. GeForce FX 5900's floating-point performance is, according to vendor estimates, twice that of the GeForce FX 5800, and improved compression and caching boosts memory-bandwidth efficiency.

Nvidia has enhanced the on-chip anisotropic-filtering algorithm so that its results are equally eye-pleasing for high-motion and static—those all-important screenshots—scenes. And, working with its industry collaborators, the company has redesigned and therefore significantly muted the large fan that cools the chip, whose din on the GeForce FX 5800 more than one reviewer equated to a vacuum cleaner (see "Graphics tit for tat turns topsy-turvy," EDN, Dec 12, 2002, pg 20). The company bases its 128-Mbyte boards on the conventional GeForce FX 5900, whose core- and memory-clock speeds weren't finalized at press time. They will have an estimated price of $399, equivalent to previously published prices for similarly outfitted GeForce FX 5800 Ultra boards. The 256-Mbyte boards with onboard GeForce FX 5900 Ultras will cost approximately $499. Both board variants should be on store shelves next month.

What CPU and core-logic chip set should you pair with these graphics powerhouses? As I recently predicted, AMD has increased its Athlon CPUs' front-side-bus speeds to 400 MHz (see "Which way?" EDN, April 24, 2003, pg 38). Nvidia's companion core-logic chip sets are up to the challenge; the company has been shipping 400-MHz-compatible parts since February. The nForce2 Ultra 400 "north-bridge" controller, employing dual 64-bit channels of DDR400-1 memory, can stably run at up to 500-MHz speeds; it's an overclocker's dream. The more economical nForce2 400, with which Nvidia and its partners hope to hit the $70-motherboard-price point, includes a single 64-bit memory controller. Pinout-compatible with the nForce2 Ultra 400, this chip can time-multiplex the controller between two external memory channels, enabling motherboard manufacturers to design one board that will accept either device.

What if one of AMD's latest 64-bit Opteron processors has caught your eye? Nvidia's prepared for the ramp of this market, too, with its nForce3 Professional line. Because Opteron contains a DRAM controller, Nvidia has consolidated the remainder of the required logic into a single chip. Now available, the nForce3 Pro 150 offers an AGP8X interface; a 10/100 Fast Ethernet controller; three Ultra ATA-133 channels with RAID 0, 1, and 0+1 capability; PCI and six-port USB controllers; and support for AC'97 Version 2.1 audio.

Due out this fall, the nForce 3 Pro 250 replaces one of the Ultra ATA-133 channels with four Serial ATA ports. Notably, nForce 3 Professional eliminates the audio DSPs that the nForce and nForce2 MCP "south-bridge" chips have. (They're comparatively unimportant in Opteron's initial server and workstation target markets.) None of the new nForce2 and nForce3 chips contain an embedded graphics core, either. Optimizing the flow of graphics data traffic from DRAM through Opteron, to the HyperTransport link and from there to nForce3, and finally over AGP to the graphics card, will be critical to maximizing perceived performance. Unlike in earlier chip-set generations, the core logic in this version doesn't itself manage the DRAM subsystem.

Nvidia, 1-408-486-2000, www.nvidia.com.