So why was the opening scene of Casino Royale so macroblocky? | Main | Lessons in Bad Dialog Design - From Apple!

July 22, 2008

I can spraypaint over the Nvidia logo too...

Please note, this blog has been archived and now lives at www.discretecosine.com

This press release from Pixellexis/Ambric Silicon caught my attention today. Pixellexis are the folks who tried to raise a buzz at NAB this year with their RedBox external processing unit. Every time I went by their booth, it was a bit empty. As Mike said, "I'm not sure I'd start a company based on the idea that computers will never get faster."

In any case, the press release talks about the new Ambric Silicon chips they're using, which have 300+ cores in a single die. Which sounds suspiciously like a commodity GPU. With OpenCL in the near future (and far cooler than the public info would have you believe), it's hard to imagine dropping serious cash on a proprietary secondary processing unit.

Posted by at July 22, 2008 10:59 AM | Misc

Comments

Hi Colin - Just a bit of background to put the GPU - Ambric comparison in perspective both with respect to current capabilities, and Moore's law trends:

1. For the same video compute-capability, comparable high-end GPUs consume sixteen (16) times more electricity (and 16X heat output) than Ambric chips. This energy-efficiency disparity for respective next-gen chips will grow even larger for silicon architecture reasons - a whole 'nother topic beyond the scope of a quick post.

Pixellexis products are aimed at rack-mount, production environments where energy-efficiency is a very big deal. There are other hardware advantages, but just the scalability and computational-density enabled by energy-efficiency is a huge differentiator for video pros.

If a 2RU box could have as many top-end GPUs as Ambric silicon enables, the box would need well over 2,000 watts of power supplies, A/C, shrouds, etc... not very feasible, (vs about 150W.)

GPUs are indeed better for certain floating point apps such as seismic-analysis and 3D medical CT voxel data, but not for most types of video processing, where computational-density matters, or where PCs lack an extra SLI slot. And GPUs are totally ill-suited for embedded-systems.

2. Ease of massively-parallel programming. It's beyond the scope of a quick post to explain this and MIMD vs SIMD etc., but there is quite a bit of info on the website www.ambric.com

Posted by: Leigh Anderson at July 23, 2008 6:07 PM

I was probably being too harsh. I just found Pixellexis' marketing runup to NAB obnoxious (I found RED's to be obnoxious too, and Pixellexis just stole their formula) and the company always has that 'vapor' feel...

Posted by: colin McFadden at July 23, 2008 6:30 PM