Your browser is not supported. Please update it.

18 September 2014

Strong tide for H.265 and software encoding at IBC

Attendees at IBC 2014 seeking clarity over the direction of video encoding will have come away this week with some guidance over two key points. Firstly the emerging generation of codecs will be dominated by HEVC, now known as H.265, with Google’s alternative VP9 falling by the wayside, despite being royalty free and of course used for YouTube. The second point is that the future of transcoding lies largely with software running on generic silicon for all but the most computationally intensive encoding tasks, such as live camera feeds. A side plot on this second theme is that Intel’s latest i7 processor based on its Haswell technology is not yet deemed fit for purpose by at least some software encoding vendors.

It is possible some attendees might have been stifling a yawn in the sense that we have been here before. The software versus hardware debate has been raging for a decade at least, as have concerns over individual processors, while Google was also battling with the ITU-T Video Coding Experts Group with its preceding VP8 technology, in that case against H.264.

But history never quite repeats itself and the difference this time around lies in the advancement of silicon technology, which has rather distilled the traditional hardware versus software debate, blurring the lines somewhat. At the same time concern over power consumption has risen up the agenda from two perspectives, firstly to cut costs and allay environmental concerns during encoding at the headend, but equally to avoid drain on mobile device batteries during decoding, or even for encoding at the client end to upload video from the field.

An underlying trend is that the big silicon companies, especially Intel, have become more dominant in terms of process technology in their big ticket generic processors, leaving purpose built ASICs and FPGAs (Field Programmable Gate Arrays) lagging behind. The result is that the performance advantages offered by ASICs, especially through having logic dedicated to a particular application such as video encoding, have come to be mitigated by lagging behind in process technology.

Various factors are at play here, a major one being the rising cost of production masks for leading-edge technologies that are more prohibitive in fabricating dedicated silicon for a particular task, while Intel enjoys greater economies of scale. At the same time, production costs for the older process technologies at larger dimensions have been falling rapidly. Intel is on 32nm process size for the chips such as the E5 based on Sandy Bridge technology and used by Envivio for example in its software encoders, and down to 22nm for the latest i7. ASICs tend to be on 40 nm, 45 nm or even larger. These measures are the half the distance between identical features on a chip, so smaller dimensions mean higher performance because the electrical signals have less far to travel in the execution of a given task.

Increasing process density has gone hand in hand with the trend towards incorporating more dedicated functions inside generic chips, which began over 35 years ago with the inclusion of floating point arithmetic units in early chips, admittedly usually then as co-processors. Now whole GPUs (Graphical Processing Units) are being incorporated and in future security functions will be embedded within chips for set tops, gateways and almost certainly mobile devices. It also looks like decoding will be embedded into mobile phone chips to improve performance and above all reduce drain on the battery. It is here that the software versus hardware debate meets the conflict between H.265 and VP9, for Google believes that it can at last score in the emerging market for embedded decoding in mobile devices, and also possibly encoding as well. The point here is that as mobile devices become capable of shooting HD and even 4K video, then there will be growing demand for efficient encoding for storing and uploading content.

Currently this looks like a nonstarter for any codec given that encoding is much more compute intensive than decoding and would take an unacceptable toll on battery life. But for this reason Google has been working hard on encoding efficiency and believes that when embedded properly in a mobile chip, performance will be acceptable, claiming a significant edge over H.265 in this regard.

So far VP9 has been given a rather bloody nose by H.265, just as VP8 was by H.264. Yet Google started with high hopes, believing that its royalty free approach would carry the day given increasing irritation with the licensing structure that has encumbered the MPEG group of standards. There is little to choose in performance between H.265 and VP9, which both roughly halve the bit rate for a given video quality compared with H.264, with potential to go further, especially for 4K content. But there are some subtle differences that make one or the other better for given situations.

Both work by taking macro blocks of pixels and rebuilding them to exploit redundancy through intra-frame predictions, reducing the amount of data needed. In both cases the maximum macro block size is 64×64, compared with just 16 x16 for H.264, providing greater scope for compression. It also confers greater flexibility, so that for example smaller blocks can be used to home in on areas of detail such as facial features, while using full size 64 x 64 blocks for larger uniform areas with little variation such as sky or perhaps grass. Here VP9 scores by allowing not just square blocks like H.265 but also rectangular blocks, such as 8 x 4, to correspond with target shapes in the picture, which increases coding efficiency. But this is balanced out by H.265 having 35 prediction modes to rebuild frames compared with 10 for VP9, which gives greater scope for exploiting redundancy on a larger scale.

One advantage Google has is sole control of the algorithms, giving it the potential to move forward more quickly with technical advances. It now aims to exploit this advantage by moving ahead quickly with a sequel to VP9, VP10, and then delivering a new generation codec every 18 months, more in line with the succession rate of web products rather than the traditional leisurely rate of broadcasting, where it took a decade for H.264, otherwise known as MPEG4, to supplant MPEG2, which itself is still widely deployed.

Of course these codecs have been optimized as they go and that will be the case for H.265, but Google reckons it can progress faster by bringing out completely new codecs more often, perhaps increasing the maximum macro block size further, as well as the range of prediction options, to exploit ongoing improvements in processing power. At the same time Google believes there will be a big market for codecs optimized for embedding in chips for mobile phones, both for encoding and decoding. It becomes a moot point then whether this is called hardware or software decoding, with Google preferring to call it the former, given that the processing will then after all take place within dedicated circuitry.

Meanwhile at IBC we found Oslo based Appear TV continuing to bang the drum for traditional hardware encoding, arguing that software running in generic silicon lacked the performance needed for live stream encoding. The company’s business development director Stephen Luff agreed that software encoders were faster to develop and that therefore were making the running for H.265 at present. His argument was that Appear TV would take its time developing a highly efficient implementation of H.265 in hardware, which would be ripe for market just when operators were wanting to deploy the codec in earnest.

This approach was ridiculed by software encoder vendors such as Elemental and Envivio, which contend that the market for hardware will be confined to live camera feeds and a few other niches. Appear TV concedes that the hardware approach has to be more flexible than in the past and so is moving towards FPGAs, so that codecs can be upgraded in the field, sacrificing a bit of performance and energy efficiency in the process.

At the same time the software encoder vendors are demanding greater efficiencies of generic silicon, so that Envivio for example is persisting with Intel’s 32 nm E5 rather than the latest 22nm i7, which on the face of it should be more efficient. Envivio’s Boris Felts, vice president of products, was critical of Intel’s focus on consumer applications for the i7 which meant it performed poorly for compute intensive applications, compared to the E5 which was designed more for embedded applications. However Felts agreed that in time the i7 would improve and become the processor of choice for a range of encoding products.