Graphics-memory-management scheme resurrects, advances past initiative
By Brian Dipert -- EDN, January 20, 2005
When Intel began, in the late 1990s, championing the AGP (accelerated graphics port), one commonly touted capability was AGP Aperture mode—that is, the ability to employ system memory as part or all of the graphics card’s frame buffer. Intel’s i740 graphics accelerator supported AGP Aperture mode, as did a series of 3Dlabs chips whose Virtual Texturing feature the company also implemented over PCI (see “Make way for the graphics Goliath,” EDN, March 13, 1998, pg 12). And some integrated graphics cores within core-logic chip sets from Intel and others, which connect to the remainder of the core-logic circuitry over an internal AGP bus, also implement AGP Aperture mode.
However, aside from the earlier cited Intel and 3Dlabs examples, discrete graphics chips have not widely adopted AGP Aperture mode. Part of the reason is memory performance: Cost-focused main memory has slower peak and sustained transfer rates than graphics-tuned DRAM. Part of the reason is system-performance overhead: An intermediary AGP-plus-core-logic bottleneck between the graphics chip and main memory increases latency, and AGP 1.0 implementations offered only 266-Mbyte/sec downstream transfer rates from main memory to the graphics subsystem. (The latest generation AGP 8x delivers 2.133-Gbyte/sec peak downstream speeds.) Also, all AGP variants support only PCI-derived, 266-Mbyte/sec peak upstream speeds from the graphics subsystem to main memory, thereby explaining why vendors have limited AGP Aperture mode use to mostly one-way, downstream applications, such as texture caching. And part of the reason for the limited AGP Aperature mode adoption is cost; although graphics-tuned memory is more expensive than main memory on a cost-per-megabyte basis, it has still benefited from the downward DRAM price spiral of recent years. Why hassle with the implementation issues of AGP Aperture mode when, for a minor increment in cost, you can simply increase the size of the local frame buffer on the graphics board?
That said, a minor increment in cost is still something more than “no cost,” and profit-squeezed chip and board manufacturers are looking to slash expenses from the bill-of-materials equation whenever and wherever they can. So it is, then, that Nvidia has revisited the concept of main memory as graphics memory with its GeForce 6200 TurboCache products, as competitor ATI Technologies will also do with its upcoming HyperMemory line. What’s changed to make the concept more feasible this time around? In two words: PCI Express. Each serial PCI Express Version 1 link is capable of full-duplex, 250-Mbyte/sec peak transfer speeds. A 16-bit PCI Express bus between core logic and the graphics accelerator can theoretically, therefore, simultaneously receive and transmit 4 Gbytes/sec of instructions and data. First generation Intel PCI Express implementations in the i915 and i925 core-logic chip sets provide only 3-Gbyte/sec downstream and 1-Gbyte/sec upstream speeds, but subsequently introduced chip sets from Intel should improve in this area, as do chip sets from Nvidia and other suppliers. Keep in mind, too, that “PCI Express Aperture Mode” implementations are more complicated with AMD CPUs than Intel microprocessors because AMD Athlon 64, Opteron, and Sempron processors integrate the DRAM controller, thereby potentially creating incremental overhead between main memory and graphics subsystem versus in the Intel-based case.
Nvidia plans at least three desktop PC-targeted TurboCache variants of Nvidia’s GeForce 6200 graphics processor, a four-pixel-pipeline spin-off of the GeForce 6800 (see “Graphics advancements span PCs to cell phones,” EDN, June 24, 2004, pg 14). A $79 board employs the 32-bit local-memory-interface version of the 350-MHz GeForce 6200; contains a 16-Mbyte, 350-MHz DDR frame buffer; and addresses as much as 128 Mbytes of memory, including system memory (Picture). For $99, you get a 64-bit interface to 32 Mbytes of local frame buffer, and, for $129, you get 64 Mbytes of local frame buffer—again over the 64-bit memory-bus option—and the ability to access as much as 256 Mbytes of graphics memory. Will TurboCache “fundamentally redefine the price/performance of entry-level PCs,” as Nvidia’s marketers tout, or will it be nothing more than “TurboCash,” a means by which Nvidia can improve its chip profit margins at a given retail price point by removing graphics memory from the board? Time will tell; competitive price pressure and the comparable pricing of previous generation Nvidia boards with larger local frame buffers will both play a part in TurboCache’s success, along with the degree to which consumers long trained in a bigger-is-better frame-buffer mantra embrace the revised TurboCache pitch. Nvidia also plans power-optimized GeForce Go 6200 TurboCache parts for notebook PCs; in this application, the TurboCache-memory-budget concept has a greater likelihood of success.
Nvidia, 1-408-486-2000, www.nvidia.com.
