The stream computing race is on - AMD FireStream 9270 is OpenCL-ready, faster than Tesla C1060
Posted by Tony DeYoung on February 09, 2009
GPU Cafe notes that the FireStream 9270 will be released Q1 2009. What makes this interesting to me, beside the performance, is the listed support for OpenCL. AMD is embracing this parallel processing standard in a big way. I wonder how long it will be before mainstream CAD and 3D Viz software vendors can determine how to use OpenCL to accelerate their computations.
If you are into the specs and numbers race, here's some of the relevant info:
- Currently the highest performing HPC processor outperforming Nvidia's Tesla C1060 by 28% (single-precision at 1.2 TFLOPS) and 207% (double-precision at 240 GFLOPS) at peak compute rates (but in real life, the API will make a big difference here).
- 2GB GDDR5 RAM
- Peak bandwidth 108.8 GB/s
- 160 watts typical, <220 watts peak
For more details, check out the PDF press release.
Comments
Hi,
is it possible to get some example code? I will run the code on my own ... please let me know or send me an email.
Best,
Maxwell
By maxwell on 2009-04-02
FireStream doesn’t support CUDA, which is becoming THE industry standard. Although, the FireStream may be technically faster, Tesla is faster in realworld applications such as in simple tasks like Rendering.
By Peter on 2009-07-05
Hmmm - my take is that Cuda is pretty niche.
OpenCL will be big, as will Direct X compute shaders.
Cuda has gotten a small start in scientific market, and a little in rendering. But it is proprietary (no matter how Nvidia likes to spin it). OpenCL is an open standard like OpenGL. That is where I expect to see all the momentum.
And with DirectX 11, Microsoft is going to turn some heads and make some in roads.
What is great about both cuda and Stream is that they have demonstrated the need for open standards that address the GPU computing market.
By Hadley on 2009-07-10
Add your comment
Note: All comments are moderated for spambots so there will be a posting delay.
Your email address will not be published.