AMD Puts 5 TFlops In Your PC


Sunnyvale (CA) - We are at the beginning of a week with plenty of supercomputing announcements, most of which target high-end and enterprise computing, but some in fact have mainstream appeal. AMD’s ATI division, for example, unveiled its Firestream 9250 card, based on the upcoming RV770 GPU, which will be part of the 4800 series of graphics cards. According to AMD, the new GPU can deliver out more than 1 TFlops - a performance that would have required 5000 Pentium Pro processors 12 years ago.

AMD kicks off this week, with the announcement of the 9250 card, without providing lots of details and not even an availability date. But besides the name of this card, it is interesting to note that the card is rated at more than 1 TFlops of performance and a power consumption of about 170 watts. Compared to the current 9170 model power consumption is up from 150 watts, but the performance has more than doubled from 500 GFlops.

Since the 9250 has a RV770 core, future owners of 4800-series cards will also be able to squeeze that performance also out of their graphics boards. The ATI team declined to say how much more than 1 TFlops the card can hit, but we were told that a 4x Crossfire X configuration is good for almost 5 TFlops. So, we would be tempted to assume that ATI is playing with a number of about a theoretically possible 1.2 TFlops per RV770 GPU.

And yes, the R700 - the dual-GPU 4870 X2 card - is good for almost twice that performance: According to AMD, the R700 will deliver 2 TFlops per board.

"Teraflops" does not mean a lot to most of us and floating pointing performance has not (yet) become a critical performance description for mainstream computing. But to put that number into perspective, consider the fact that 5000 Pentium Pro processors delivered 1 TFlops back in 1996. To match the 1.2 TFlops, you would need 6000 of these CPUs and to match the (32-bit) floating point performance of a quad-Crossfire system you would need 24,000 of these processors.

Though these are highly theoretical numbers and the actual performance always depends on a specific application, there is no doubt that today’s graphics cards are sleeping performance monsters. We wonder who will wake them up.