Jump to content
The Inquirer-Home

New FireStream 9170 confirms GPGPU attack

320 Double-precision FP units power RV670 chip
Thursday, 8 November 2007, 16:11

IN THE LIGHT of next week's supercomputing conference in Reno (SC07), AMD decided to pre-announce the FireStream 9170, company's third GPGPU card.

After low-volume trials with FireStream board based on X1900 card, and R600 prototype board with 2GB of memory, AMD is now coming to market with a real volume product, the FireStream 9170.

As the name tells, this part is loosely based on Radeon HD 3870, but it sports 2GB of GDDR3 memory. Once again, AMD opted for cheaper solution, even though using GDDR4 would significantly drop the power consumption. It seems to us that AMD is only playing the power card when they like to. FireStream 9170 will consume around 150W, which is still brilliant amount of power for getting 0.5 TFLOPS.

Since RV670 has double-precision floating-point units, one of key milestones has been completed, and now we have a race between G100 and R700 to get the first part capable of doing 1TFLOPS with double-precision FP64, a prime requirement for getting some nice government contracts.

According to AMD's own data sheet, this part should give 497 GFLOPS if the GPGPU core is running at 775 MHz. If AMD decides to ups the specs a bit (10-15 MHz more), it will have 0.5 TFLOPS per chip. Upcoming R680, dual-RV670 board in its GPGPU form will be able to host 4GB of local memory, and should be able to pass the magical 1TFLOPS barrier. µ

Share this:

Comments
Missed Opportunities

With AMD going to GPCPU architectures and dual video cards already having 4GB of RAM, it makes one wonder why they aren't slapping 8GB of RAM on the Hypertransport bus and sharing it with the CPU?

Wouldn't putting CPU, GPU, and Memory all on Hypertransport finally unlock the technological advantage of the on-board memory controller in Barcelona allowing a huge performance boost overall?

Well, it SOUNDS like a good theory...

posted by : Bill, 08 November 2007 Complain about this comment
Marketing blah

This is all blah the 0.5 TFlops is all theoretical and will never be reached in a real world application. For example branching performance is still horrible even on the most sophisticated GPUs out there nowadays.

posted by : DrBalthar, 14 November 2007 Complain about this comment
PwN4g3

Th15 1s tH3 PwN4g3!!!!!!

posted by : iphone , 22 November 2007 Complain about this comment
Advertisement
Subscribe to the INQ Newsletter
Sign-up for the INQBot weekly newsletter
Click here to sign up Existing user
Advertisement
INQ Poll

Browsers

Who will win the next round of browser wars?