[FFmpeg-devel] transcoding on nvidia tesla
Fri Feb 1 01:09:22 CET 2008
Regarding HD encoding, I was illustrating the end-result benefits of the
overall solution, not a particular use-case, such as HD encoding.
I am not an expert in video transcoding technology, but I am aware of
overall benefits of doing something like this.
The Tesla GPU 1RU system, which can be clustered for performance and fault
tolerance, can have eight 128-processor gpu's. It's only 550W, has 76.GB/sec
peak per cpu memory access. In addition, just the deskside version of the
product, can bring up to 8 teraflops to that one pc. A high-density server
could have even more.
The performance enhancements are 100% real. For example,
http://www.o0o.it/pro/, shows that a Mac Pro with two dual core xeon
woodcrests at 2.66ghz crank out no more than 60gigaflops. So a tesla at a
minimum of 8 teraflops per pc is 130 times more powerful.
I'm not claiming this is my unique idea, but I'm not the only one who sees
the benefits in doing something like this.
I'm merely suggesting looking into this, and if there are individuals who
can do this for a financial and overall IP contribution to the project, it
would be very beneficial to everyone.
If it [tesla] wasn't that much higher in performance, I'm sure nvidia
wouldn't have invested tens of millions or more into developing such a
system, if it wasn't powerful. Also, ATI made an announcement to release
something similar to tesla in 09.
Additionally, everyone knows that GPU's have higher flops than regular
processors, so it's a natural fit for video transcoding, hence offloading to
gpu's that some people tried doing.
From: ffmpeg-devel-bounces at mplayerhq.hu
[mailto:ffmpeg-devel-bounces at mplayerhq.hu] On Behalf Of Mike Melanson
Sent: Thursday, January 31, 2008 3:22 PM
To: FFmpeg development discussions and patches
Subject: Re: [FFmpeg-devel] transcoding on nvidia tesla
Dan Khasis wrote:
> I can make a significant contribution to get this going, and purchase a
> or two for development testing.
> When it comes to converting thousands of videos per second/minute on one
> two teslas, versus 10,000 servers across 10 datacenters, using kilowatts
> power, the costs associated with several of these servers become
> I guess the real question is if it's possible, how long it would take, and
> how much I would have to contribute :)
> Based on my understanding of how well the developers of ffmpeg, have made
> ffmpeg the de facto video conversion application, I think it's safe to say
> they can do this too.
> With a tesla or two, you could be transcoding thousands of HD streams
You make a compelling case. However, if you're doing HD transcoding, I
suspect that refers to H.264 encoding? Perhaps it would be best to focus
on libx264 for such an effort?
Alternately, maybe you could encode HD SVQ1 in real time. :)
ffmpeg-devel mailing list
ffmpeg-devel at mplayerhq.hu
More information about the ffmpeg-devel