AMD plans an absolute GPGPU monster

WE HAVE SEEN a lot of interesting R600 docs and one of the prototype boards and we are fascinated by the level of flexibility that ATI will offer to the world and its partners when the technology finally hits the market.

We are not referring here to the number of 3Dmarks achieved, but rather to the professional range of products.

While Nvidia is preparing the G80GL, a Quadro version of G80, with not many changes beyond lowering the clocks and increasing the amount of video memory (drivers are the biggest worry, especially the mandatory 64-bit optimisation), the R600 marchitecture and the PCB designs are so flexible that in the next six months we will be inundated with various products based on R600, R610 and R630 designs.

We have already disclosed an unrivalled amount of information when it comes to these products, but for the very first time, we'll be mentioning the professional range of products. Most notably, the most powerful one.

ATI is preparing something really, really special. If it manages to pull it off, it will be a breakthrough even Captain Hook couldn't have dreamt of. We are talking about a GPGPU product, the FireStream/Stream Processor board with no less than four (4) Gigabytes of local video memory.

The biggest issue we can think of is the probable requirement for 1024 Mbit GDDR-2/3/4 memory chips, and neither Samsung nor Qimonda is manufacturing them. However, design documents specifically mention 4GB of memory, which made my head dizzy. If they cannot make it, they will release a version with "only" 2GB, but our sources are talking about the ultimate monster board.

The planned design of the board with maximum of four Gigabytes isn't some trick like putting a GPU on an ATX form factor mobo and calling it a "graphics board", the design is a variant of its upcoming top-end 12-inch design.

This board will feature only one DVI connector, but this board is not targeting the world of multiple monitors, rather one where monitors are irrelevant. It will be interesting to see those External CrossFire boxes stacked into an array of GPUs whose sheer power is unrivalled in the computing world.

Did we mention real purpose of GPGPU? - annihilating the importance of CPU and server CPU margins, which cannot compare with R600 or G80 in terms of pure processing power. Of course, this comparison is valid only in GPGPU-friendly case scenarios, so we're talking about streamlined computing only. Engineers at PeakStream and Stanford Uni are already having wet-dreams about the possibilities that a single-GPU configuration will do, yet alone multi-GPU one.

We wonder who will use this board for searching algorithms first, Microsoft Live or Google/YouTube? µ

L'INQ
GPGPU.org