DELIVER



AV1 - the next standard-codec?

Pros and cons of the new AV1 codec.

How is the new AV1 codec performing in practice?

Only recently it has become possible for codec and streaming specialists to directly compare the currently or perspectively relevant codecs using the same encoder.

Since the time of H.264 seems to come to an end and its successor, the H.265/ HEVC, is having a hot battle with the free codecs of the VPx and AVx series, it is the right moment to have the opportunity of such a direct comparison.

Admittedly, the HEVC has been on the market for some time, whereas the AV1 is still in a relatively early stage of development.

Nevertheless, the ambitions of all developers and providers on the one hand and the requirements of customers on the other are enormously high.

Let's take a look at the various parameters that have been tested so far and

the results.

Despite very high image quality at low data rates, AV1's advantages and disadvantages are still very close to each other. The reason is that in the currently available versions of AV1, the encoding and decoding speed is not yet competitive. Also, in the playback of AV1 there have been display problems without additional GPU and implemented hardware acceleration. This is partly due to the fact that the AV1 decoder can currently only use one CPU core. Of course, there is still a lot to be done, but in this article we will take a look at the current status.

Encoding speed



Figure 1 shows the encoding times in seconds as well as the corresponding real-time performance.

AV1 encoding lasted 62 hours and 48 minutes, which was 45,216 times longer than real-time. Compared to less than 5 minutes for HEVC and VP9 and 58x and 45x real time. Although subsequent versions of the AV1 are faster, the gap is still huge. By the way, the test was performed with an HP ZBook notebook equipped with a 2.8 GHz Intel Xeon E3-1505M V5 CPU with NVIDIA Quadra M1000M graphics chipset and the HD Graphics 530 CPU embedded in the CPU.

This results in a basic problem when using the AV1. Since time (in this case encoding time) always means money, AV1 is currently not affordable for most users. The significantly longer coding times could only be compensated by savings in bit rate and in combination with very high view numbers.

Nevertheless, the image quality of the AV1 is convincing. Figure 2 shows the average VMAF values. VMAF stands short for Video Multi-Method Assessment Fusion and is a method for objective evaluation of image quality that was not introduced until 2016. The measurement algorithm combines metrological and human sensory evaluations and merges them into a single score.



As you can see, AV1 performed best, followed by x265, VP9 and finally H.264.

In general, a VMAF score of 93 means that a clip is perceived without annoying visual artifacts. HEVC and VP9 reach this lower limit at about 2 Mbps, H.264 at just over 4 Mbps. AV1 is considerably higher at all measured data rates.

At very complex settings where the requirements to the a codec are particularly high, AV1 reaches the 93 VMAF limit at about 2 Mbps, HEVC at about 2.7 Mbps, VP9 at slightly above 3 Mbps. Again, H.264 is far behind. It only reaches the limit value at approx. 4 Mbit/s.

As a result, AV1 can reduce the data rate by about 24% compared to HEVC, by about 33% to VP9 and by 49% to H.264 for the same quality. These values refer to complex encoding-intensive settings. With average complex settings, the savings are about 10% higher for HEVC, H.264 and 4% for VP9.

Basically, AV1 can save considerable bandwidth when used for HEVC, VP9 and especially H.264. However, the actual perceivable difference in quality is not that significant for the same bandwidth.

What do these tests mean in everyday life? 

H.264 should produce almost perfect quality at 8 Mbps. HEVC and VP9 achieve this quality level at 3.5 Mbps to 4 Mbps. AV1 and future codecs should perform this in the range of 2 Mbps to 3.5 Mbps. Thus, while the overall quality difference for easy-to-code videos is only slight, AV1 can significantly reduce the overall data rate for complex content with no quality difference.

Decoding

The decoding speed is shown in Figure 5.

AV1 decoded in 0.66x real-time, HEVC in 8x real-time, VP9 in 10.5x and H.264 in 14x real-time.
You can also see that at a CPU utilisation of about 20%, only one of the system's four available processor cores was used. If AO-Media can recode the player to address more than a single core, it can play 1080p video in real time on the ZBook. However, it should be noted that CPU utilisation is much higher than with the other formats, which has a disadvantage in terms of battery life or CPU temperature, for example.

H.264, HEVC and VP9 all played in real-time with minimal impact on the CPU. Using  AV1, playing a 1080p file was not possible at all. Without a more efficient decoder this does not speak for the playback performance of AV1 on devices without GPU or any other form of hardware acceleration.
Summary.

The tests show an amazing quality compared to existing codecs. However, current coding costs are far above of what the vast majority of video publishers can afford. It is also questionable how many devices can play AV1 without additional hardware acceleration. For most mobile devices, this hardware playback support should be available from 2020.

However, AV1 is under development and it can be expected that the encoding and decoding performance will improve rather quickly. However, AV1 may not be a realistic alternative for most manufacturers for the next 12 to 18 months.

Related Articles: