Transcoding Quality: APP Versus CUDA Versus Quick Sync
Now let's switch from our Blu-ray source to the trailers we transcoded on the previous page. Since we are providing download links to our source content and our transcoded results, we are unable to use the original Blu-ray content due to copyright law.
First off, special thanks to Elemental Technologies for letting us have an exclusive first look at Badaboom 2.0. The company is basically in an alpha stage, so everything is still very preliminary. The reason we wanted to include even this early version, though, was that it came most highly recommended by Nvidia after discussing the results of our Brazos coverage. So, even though it is an early build, it should answer our quality-related questions. But because we are dealing with pre-beta build, we will not be releasing benchmark figures.
AMD's APP
The transcoded video from the Radeon HD 6970 looks vastly better in MediaEspresso than it does in MediaConverter. Everything we output from MediaConverter looks like it has been put through some sort of heat filter. It looks as if you are always looking at the image through a mirage effect. Chris was sitting next to me and remarked, "the whole scene is shimmering."
Nvidia's CUDA
There are two anomalies here. As you might notice, the light on Doug's collar is slightly brighter in MediaConverter, and the frame is off in Badaboom (it's not the same frame at all). Both programs generate a file that doesn't properly track, which means during the reassemble process, something is off-kilter. This screen capture is not the result of human error, as we used the batch function to track and output specific frames. This is something specific to using CUDA in MediaConverter and Badaboom. Given that Elemental is still in alpha, we readily expected some errors, and entropy-related issues that affect the reassemble process are usually a quick fix. The issue in MediaConverter is more puzzling, though, since we used the retail edition to get CUDA functionality (the latest beta inexplicably doesn't recognize our GeForce GTX 580).
Aside from tracking issues, this time we are looking at poor quality output from MediaEspresso. It is particularly pronounced in some smooth panning scenes, and when there is a slow fade into another scene. Overall, CUDA does look better in MediaConverter than it does in MediaEspresso, but Badaboom's output looks the best. This would indicate that there is something problematic with the implementation of CUDA in the first two apps. With that said, only the CUDA videos from Badaboom and MediaConverter demonstrate poor motion prediction in areas we haven't see with MediaEspresso.
Intel's Quick Sync
The output from MediaEspresso is probably the best of the three tested programs. It is harder to nit pick when things look so similar...and we're not complaining. Aside from another tracking issue, output from Badaboom also spits out a file that is a tad grainier compared to the other three. Even the Performance and Quality settings for MediaEspresso show less difference than one might think. No doubt, the Quality setting outputs a larger file size, but you only notice the differences in high-motion areas, and where fine detail is showcased (like hair, for example).
CPU-Based Encode / Decode
Of all the output settings, the full software transcode pathway is consistently the best. We don't have any tracking errors, save for Badaboom's early build. The detail is consistently good, no matter how many times we transcode the file, and you can almost always which file was transcoded on the CPU, comparing output files from each program to each other.