Graphics tit for tat turns topsy-turvy
By Brian Dipert -- EDN, December 12, 2002
Graphics powerhouses Nvidia and ATI Technologies are at it again, trading marketing pot shots and unveiling contending desktop graphics products. This time, however, Nvidia finds itself chasing its Canadian competitor, versus the other way around, a situation Nvidia hasn't experienced for several years now. ATI (www.ati.com) launched its Radeon 9700 chip several months ago; boards are now in high-volume production, just in time for holiday shoppers. Nvidia, however, is now only beginning to ship samples of its GeForce FX (formerly known via its NV30 code name), and boards based on the chip are not scheduled to appear on store shelves until at least February (Picture). What happened? Off the record, company contacts grumble that part of the blame lies with delays in Microsoft's (www.microsoft.com) finalization of its DirectX 9 API, and Nvidia's unwillingness to roll out silicon based on a still-subject-to-change beta API specification.
Foundry partner TSMC's (www.tsmc.com) production-qualification delays didn't help, either. Unlike ATI, which built the first-generation Radeon 9700 on a 0.15-micron process, Nvidia decided to take a risk with the more advanced, but also unproven, 0.13-micron, copper-inclusive process (see "Graphics rivals ready for the holidays," EDN, Aug 22, 2002, pg 18). Short-term, Nvidia's decision led to an approximately four-month slip from the original target availability date. Long term, however, Nvidia hopes that the combination of higher process performance capability and more usable die per wafer, translating to lower cost per die, will give it a competitive edge. And, with the bugs finally ironed out of TSMC's process, Nvidia also aspires to return to its historical every-six-month new-product-roll-out cycles.
At first glance, the 125 million-transistor GeForce FX and its Radeon 9700 adversary look a lot alike, both offering 8-pixel pipelines, DirectX9 v2.0-compliant pixel and vertex shaders, and AGP8X support. Peer closely, though, and, Nvidia claims, the differences will become more apparent. The company hopes to hit 500-MHz core and memory speeds, and the memory delivers 1-Gbps-per-pin peak transfer rates, courtesy of DDR-II SDRAMs. GeForce FX supports both 64- and 128-bit precision floating-point internal data formats, and its shaders expand on the base DirectX-required feature sets with such features as long programs, conditional execution and branching (vertex only), subroutine calls and returns, and a unified instruction set.
Nvidia believes that the combination of twice as many pixel pipelines and higher internal speeds, along with a redesigned vertex shader architecture, will deliver on average three times the frame rates and three times the vertex-processing capability of GeForce4 (see "Graphics evolution 'foursees' increasingly rich images," EDN, Feb 7, 2002, pg 24). The vertex-shader redesign in this product generation stands in contrast to the simpler doubling of shaders that marked the GeForce3-to-GeForce4 transition. The company estimates that GeForce FX boards with 128-Mbyte frame buffers will retail for approximately $399.
As it strives to reassert its desktop dominance, Nvidia is following in ATI's footsteps on the mobile front, too, rolling out a notebook variant of its previous-generation flagship chip (see "Proliferations pad graphics portfolio," EDN, Oct 3, 2002, pg 16). Whereas past GeForce4 Go products were, for all intents and purposes, power-optimized variants of the GeForce4 MX (itself heavily influenced by the now-ancient GeForce2), the new GeForce4 4200 Go, as its name implies, is a low-power proliferation of the AGP 8X-inclusive GeForce4 Ti 4200. The large die, coupled with dual address buses of the Lightspeed memory architecture, prohibit any multidie frame-buffer-inclusive packaging options for GeForce4 4200 Go. Per-formance gains also come with a proportionally large price tag: Nvidia is targeting GeForce4 4200 Go at systems selling for $2500 and above.
Nvidia, 1-408-486-2000, www.nvidia.com.
