Graphics advancements span PCs to cell phones
By Brian Dipert -- EDN, June 24, 2004
Nvidia’s travails of the past two years are a reminder of the lesson of Aesop’s classic fable The Tortoise and the Hare. Nvidia attempted to outleap its primary competitor, ATI Technologies (www.ati.com), to 0.13-micron technology with the GeForce FX 5800. But the initial revision of the process and, therefore, the chip based on it was excessively power-hungry. Nvidia’s foundry partner TSMC (www.tsmc.com) also experienced lengthy production delays, causing Nvidia to miss the critical 2002 holiday-shopping season. ATI, in contrast, targeted the more mature 0.15-micron process for its Radeon 9700 graphics flagship, which smoothly ramped into high-volume production. Because ATI had the market to itself for many months, any potential lithography-driven cost disadvantage versus Nvidia went unrealized. And, again due to ATI’s market dominance at the time, Microsoft (www.microsoft.com) settled on the Radeon 9700’s 24-bit, floating-point shader architecture for initial revisions of the DirectX Version 9 API, leaving Nvidia’s 32-bit approach in the 5xxx series at a performance disadvantage on DirectX Version 9-based applications.
Fortunately, Nvidia, unlike many other semiconductor companies, had sufficient money in the bank to help it ride out the rough times. It was also able to convince sufficient numbers of game developers to support 16-bit, floating-point partial precision, which Nvidia could speedily handle in hardware, in their DirectX 9 programs. With its latest graphics accelerator, Nvidia asserts that its past woes are behind it, due in part to its healthy relationship with a second 0.13-micron foundry partner, IBM (www.ibm.com). Nvidia squeezes 222 million transistors, some used to construct 16 parallel pixel pipelines, each with a matching texture unit, into the GeForce 6800. (The GeForce FX 5800, in contrast, contained four pixel pipelines, each with two corresponding texture units.) The GeForce 6800 also includes six vertex units, each capable of simultaneously processing one vector and one scalar operation, and supporting vertex fetches and branches and, therefore, displacement mapping, along with vertex instancing for high-efficiency model batching.
Nvidia hasn’t backed down from its support for 32-bit, floating-point shader instructions, and the upcoming DirectX Version 9.0c finally supports them at full precision. Although the GeForce 6800 significantly raises the industry bar on graphics performance, its video capabilities are also equally impressive. Nvidia moves away from the hard-wired MPEG-2 decoding pipeline in past chips to a more flexible video-processor approach that’s currently capable of hardware-accelerating MPEG-2 decoding and encoding, along with performance-boosting Windows Media Video 9 decoding at high resolutions and implementing advanced deinterlacing techniques. Support for WMV9 encode, along with other video codecs, is ongoing.
The GeForce 6800 does have the potential for memory-bus bottlenecks. As a historical analogy, the GeForce FX 5800’s narrow 128-bit external frame-buffer bus hampered its performance; the 256-bit interface in the follow-on GeForce FX 5900 relieved this restriction. Similarly, a 1.1-Gbps-per-pin, 256-bit DDR-3 SDRAM bus may still be insufficient to keep up with the GeForce 6800’s 16-pipeline pace. Time will tell; $499 Ultra boards running at a 400-MHz core frequency and with 256-Mbyte frame buffers are due on store shelves in June, along with Nvidia’s first iteration of its Release 60 drivers. A less-than-$300 non-Ultra variant of the GeForce 6800 with only 12 functional pixel pipelines and 128 Mbytes of DDR-1 memory is scheduled to appear one month earlier; Nvidia has not yet finalized that device’s core and memory speeds.
Although PCs may provide most of Nvidia’s near-term revenues, the company also remains focused on cell-phone and PDA opportunities that last year’s acquisition of MediaQ enabled it to pursue. Nvidia has, at least for the moment, shelved MediaQ’s ARM core-inclusive chips but is continuing the development of multimedia coprocessors. The latest products include the GoForce 3000, targeting $149 cell phones, and the higher end GoForce 4000 for fuller featured $199 phones. GoForce 3000 embeds 320 kbytes of SRAM, interfaces to 2 million-pixel camera modules, hardware-accelerates 2- and 3-D graphics functions, and decodes CIF-resolution MPEG-4 Simple Profile and H.263 video streams at 30 frames/sec. GoForce 4000 doubles the amount of embedded memory over its lower end sibling and adds CIF-resolution, 30-frame/sec video-encoding capabilities, along with 3 million-pixel still-camera support. Nvidia has its work cut out for it; the market for handheld multimedia accelerators is quickly becoming crowded. Intel’s (www.intel.com) 2700G, previously code-named Marathon, is a formidable foe, especially when the company bundles it with the latest generation PXA270 XScale processors.
Nvidia, 1-408-486-2000, www.nvidia.com.
