by Anton Shilov
04/04/2008 | 06:30 AM
Although both leading developers of discrete graphics processing units (GPUs) recently introduced their new dual-chip flagship offerings, there is stagnation on the market of single-processor graphics cards. Radical performance increase of standalone chips may be expected along with the launch of next-gen solutions, but nobody knows when this happens. One of the guesses is that both GPU suppliers wait for GDDR5 to arrive.
Nvidia Corp., currently the leading supplier of discrete desktop GPUs, unveiled its GeForce 9800 GTX just a few days ago without making loud announcements or claims. The reason for this quiet launch is simple: the new GeForce 9800 GTX is not faster than the GeForce 8800 GTX released almost one and a half years ago and offers almost no new capabilities. A similar thing can be said about ATI, graphics product group of Advanced Micro Devices, who launched its Radeon HD 3870 back in November with fanfares, but without any performance improvements over ATI Radeon HD 2900 XT, which was meant to be a product due in 2006, but slipped into May, 2007.
Given that ATI Radeon HD 2900 XT and Nvidia GeForce 8800 GTX graphics boards supported multi-GPU capabilities, two of each graphics boards could easily provide performance of flagship ATI Radeon HD 3870 X2 and Nvidia GeForce 9800 GX2 a long time ago. Of course, at pretty high price-points.
The new generations of discrete graphics adapters – ATI Radeon HD 4000 and Nvidia GeForce 10 – are projected to deliver the new definition of high-performance, but unfortunately neither of the GPU developers wants to pinpoint the exact timeframe when the next-generation products are available. Perhaps, because nobody is exactly sure when they can deliver them?
It is not a secret that modern graphics chips are extremely complex in terms of transistor count, which makes them hard to develop and produce. Limiting transistor count of the chip allows to make its creation and ramp up easier, but also reduces potential performance because of lower amount of execution units. Another thing that limits performance is memory bandwidth.
There are two ways to boost memory bandwidth: to push memory chips clock-speeds up and to widen memory bus. Both ATI and Nvidia experimented with 512-bit and 384-bit memory buses in the past, but the latest graphics chips by two companies utilize 256-bit memory bus, partly because the latest ATI RV670 and Nvidia G92 chips are aimed at market segments from about $150 and upwards.
Existing GDDR3 memory chips may run at 2.0GHz – 2.2GHz, which provides up to 140.8GB/s memory bandwidth in case of 512-bit bus, though, such chips are pretty expensive, just like print-circuit board with 512-bit memory bus. For ATI the use of high-speed GDDR4 is an option, whereas Nvidia GeForce 8 and 9 GPUs do not support GDDR4. Therefore, the company either needs expensive GDDR3 in conjunction with wide memory bus, or GDDR5 with its extreme clock-speed potential of up to 6GHz amid wide availability from Samsung, Hynix and Qimonda.
Even though neither ATI/AMD or Nvidia has publicly indicated huge interest towards GDDR5, which is due later this year, it is likely that the new memory type will play a significant role in the next round of graphics card war.