[FFmpeg-user] Best CPU for ffmpeg

[Henk D. Schoneveld]

On 28 Jan 2014, at 12:13, Ed Torbett <ed.torbett at simulation-systems.co.uk> wrote:

>> I don't know that's why I asked.
>> (But I am quite sure that "high quality" - of course
>> depending on the definition - makes it more cpu-bound than
>> I/O bound.)
> 
> For single files, certainly. But in the case of encoding multiple files simultaneously, the cost of non cached random disk page accesses would rise significantly. As to how significant that is, I admit I have no idea. I base my knowledge of disk I/O performance on working with enterprise-grade databases, which I admit is a tiny bit different…
I’m under the impression that putting up a simple test-set of > 10 files and encode with #cores single threaded ffmpeg instances to look and measure what’s happening is beyond what can be done. What I mostly read are the words think feel believe, I almost never read the words measure and calculate. How come ? 
On  http://www.cpubenchmark.net/high_end_cpus.html one can see that on the page with the fastest CPUs there is a E3-1280 with a PassMark of 10110
The same CPU is the fastest single threaded CPU with a Passmark of 2317 it’s a 4 core CPU

I do have a Celeron G1610 where these numbers are 2557 and 1399 respectively.

I also happen to haven a i5-2500K. With the passmark and with my encoding results with both CPUs I know that the Passmark is quite accurate for encoding-speed prediction. 
i5-2500K passmark numbers are 6486 and 1891
the 2Core G1610 has ~ 40 % encoding speed relative to the i5-2500K if used in single threaded encodings. 2 x 1399 vs 4 x 1891 = 2798/7564 = 37%
Multithreaded the relation is 2557/6486 = 39.4% according to the Passmark. But in real live multithreaded encoding the result is 20.2/51 = 39.5%

The G1610 does 13fps single threaded and 20fps multithreaded, 2 threads.
The 2500K does 19fps single threaded 31fps 2threads and 51fps 4threads
Let’s assume a 4CPU 2500K which is 16Cores at 19fps which is 304fps in a 25fps file of 2300Kb/s results in 304/25*2300Kb/s = 27968Kb/s = 3496KB/s = 3.5MB/s
The source files will be read at 304/25*18000Kb/s resulting in ~ 27MB/s
Reading from one disk and writing to another or on a single SSD won’t be any problem on this non-existent 16Core super machine, with a theoretical Passmark of 25000.

I hope I didn’t mess up my calculations. 

> 
> 
> _______________________________________________
> ffmpeg-user mailing list
> ffmpeg-user at ffmpeg.org
> http://ffmpeg.org/mailman/listinfo/ffmpeg-user