Real world testing

With such a stellar success on the HQV Benchmark, we brought in some real-world tests to challenge the ATI Radeon X1800 XL. The unusual cadence detection was responsible for the substantial improvement in performance, so I brought in some of my own clips with unusual cadences.

Real-World Cadence Challenge #1: MegaZone 23 Trailer

First up was the trailer for Megazone 23. For the non-anime fan out there, Megazone 23 is one of the classics of 1980's anime and was the basis for Robotech: The Movie. Although this is actually a 3:2 sequence, and the X1800 was unable to detect the appropriate cadence in our real-world test. We saw plenty of jaggies and feathering artifacts and our findings were confirmed by ATI. What's interesting is that NVIDIA's PureVideo, which only scored 51 points on the HQV Benchmark actually does well here (although it occasionally drops the resolution). NVIDIA GPUs are able to detect this cadence appropriately and for this clip, NVIDIA actually locks onto the cadence faster than even HQV solutions. The XGI Volari 8300 was not tested.

According to ATI, in turns out that one of the challenges with this clip is that there's a lot of noise in the clip which causes the GPU to assume that the fields are different when they are not. Their detection thresholds are apparently conservative in order to detect bad-edits appropriately. Fortunately, the feathering artifacts are due to a bug in the driver which is correctable and should make its way into a future driver.

Real-World Cadence Challenge #2 and 3: More Anime

With the experience of the Megazone 23 trailer, I decided to challenge the video processors with newer anime. In retrospect, using anime as an "unusual cadence" test makes a lot of sense because there are a lot of fast cuts and edits which makes it very challenging for video processors to lock onto the right cadence. On that same DVD were trailers for Magical Shopping Arcade Abenobashi (trailer) and RahXephon (trailer). In these tests, I found that the HQV processor handled the cadence detection well whereas the ATI X1800 had substantially more difficulty resulting in feathering artifacts. NVIDIA had no feathering artifacts, but frequently dropped to half-resolution when it was unable to appropriately detect the cadence. The bug fix ATI mentioned for a future driver release would take care of the feathering artifacts although it's unclear if tweaks to their cadence detection mechanism can be done.

Interpretation? Although ATI Catalyst 5.13 has raised the bar when it comes to PC video, it doesn't do nearly as well with the unusual cadences in some of the real-world challenges that we threw at it. I know what some cynical readers might be thinking, but this is really a reflection of how challenging video processing is. There's no brute-force solution.