Portable Field Transcoding with Small-Form-Factor Computers
What if you need to transcode files when you're in the field, in between production shoots or as a way to push digital dailies out to local clients? Recognizing the need for portability in such a scenario, we ran a number of tests to determine the ideal balance of size and speed, and we share those results here.
So now that we know the test files, the software, the multiple hardware choices, and the final output formats. what were the results?
We were surprised at our results, to say the least, so we re-ran the tests several times. But each time we got the same results, so we’re comfortable sharing what we found.
Since we were transcoding a movie-length file, we knew we would see significant transcoding times, and we threw three transcodes at AME at the same time, with the understanding that AME would either choke and run each transcode sequentially or rise to the occasion and leverage a single decode for all three transcodes.
What we didn’t expect, though, was to see a single transcode take a significantly shorter period of time—even when extrapolated to correspond to the three-transcodes workflow—than the per-transcode time it took to simultaneously transcode all three outputs.
Here’s what we mean.
On the GA-H87TN, using the i3-4130 CPU with integrated GPU HD4600 graphics, the total time for all three transcodes was the lowest overall at 8 hours, 13 minutes, and 32 seconds (8:13:32). Compare that to the MacBook Air, which took over twice the amount of time (17:20:09 total) or even the NUC D54250WYK (12:40:02 total).
Open CL transcoding minutes for three simultaneous transcodes
So far so good, and we attribute the significantly shorter time for the 1RU rackmounted Inteset case unit, in large part, to the i3-4130 being the only one of the three test beds that had appears to be Open CL-certified GPU for the Mercury Playback Engine.
When we tried to use the Mercury option for the NUC and MBA test devices, a warning appeared saying that the GPUs were not certified. We plowed ahead with the GPU-based testing, though, on the assumption that a lack of certification might not necessarily translate in to a lack of performance.
We were partially right, as our NUC tests confirmed: from the 12:40:02 total transcode time for all three output files, utilizing the GPU, we software-only version took longer, but only slightly. Final outputs were 12:53:12, or essentially 13 minutes longer in an almost 13-hour process.
NUC D54250WYK(H) transcoding minutes for three simultaneous transcodes
If that weren’t perplexing enough, when we ran the software-only tests for the GA-H87TN/i3-4130 combination, times rose from 8:13:32 to 10:03:35 total time. That’s in line with expected results, as it means that the GPU was providing somewhat considerable benefit to the whole transcoding process.
GA-H87TN/i3-4130 transcoding minutes for three simultaneous transcodes
Yet when we ran the software-only version of the test for one file output—the 5Mpbs version, which was the largest transcoded file size—the result was 2:47:02 for one file. Extrapolate that out to three files, and it brings the total to 8:21:06 for three files. That’s a total of 6 minutes difference in a slightly more than 8-hour transcoding time. In other words, the software version for files done in sequence came very close to the hardware-assisted simultaneous transcodes.
All that said, what we can clearly discern from our tests is twofold: First, Adobe has some work to do on Adobe Media Encoder CC 2014 and the certification process. While this is the first year in about the last five that we won’t have representation at the Adobe MAX conference that takes place in early October, we suggest readers be on the lookout for whether Adobe gets its act together in Open CL across the Mercury Playback Engine landscape.
Second, we can clearly say that the smallest box—small enough, in this case, to fit on the VESA mounts on the back of a monitor—wasn’t the best box. Nor was the laptop we had on hand in our field testing. The best option was a middle-of-the-road unit that balanced integrated GPU and a tight-fitting data drive in a 1RU package, combining the best benefits of a small desktop machine with a form factor at home in any rackmounted production environment.