The argument went back and forward ‘ ‘If you really can send a HD video signal at 3 Mbps, then that changes everything, but you need to convince me that you really can.’
This was more or less the gist of the first 30 minutes of discussion this week between Faultline and the CEO of Broadcast International, Rod Tiede, a small 45 employee ‘yet another codec company’ that would have very little credibility in making this claim but for two things ‘ standing beside him is a very big brother, in the form of a senior IBMer, and the fact that the system has already been shown working to the surprise of the audience of the IBC broadcasting show in Amsterdam in September.
To say that this would be game changing if it was entirely and completely true, is an understatement.
Tiede explains, ‘Our software, called CodecSys, has been ported to the Cell chip, the supercomputer on a chip designed by IBM to drive the PlayStation 3 Sony games console. We use more than one codec, currently three or four and we will use more, but we use them in a single encoder. Each codec is good at something in particular, like dark or bright scenes, or slow-motion or fast motion scenes. The system takes a look at the outputs and selects the one that uses least bandwidth and then recombines the outputs. It then checks the final scenes to see how a human eye would perceive it and makes sure the video is of high enough quality.
‘We have tested this against the 19.4 Mbps MPEG2 video that is being used for HD at EchoStar, and our output is equivalent or higher quality, in 3 Mbps,’ he said. The company uses an H.264 codec from Vanguard Software Solutions plus other specialist codecs it has designed itself.
We remind him that many encoder manufacturers including Harmonic and Tandberg, are claiming 6 Mbps and falling, for H.264 encoded HD, but he comes back with the completely bulk standard nature of the IBM blades that will drive this, reminding us that the company is adding more codecs all the time, and that the process is entirely software upgradeable.
That’s true also for encoder manufacturers that are using Texas Instruments or other DSPs, and which use specialist decoder chips such as Ambarella (which Harmonic uses) though not where Asics are used. FPGA’s are also used for encoders, and while these can be software upgraded Tiede makes the point that this is only possible when they have the processing headroom to do more than they are currently doing.
The Cell chip, as we have explained before is made up of 8 slave processors which use an MIMD (Multiple Instruction Multiple Data) architecture to allow significant parallelism on the chip. These all offer to carry out work for a central Power chip, and swap workloads over an extremely fast onboard bus, with various chunks of cache and memory on board as well.
It is an architecture which is becoming mature and the four year old SDK (Software Development Kits) have been refined so that virtually any software engineer can program for the device, without an understanding of parallelism, all on a version of the Linux operating system.
It processes millions of small calculations exceptionally fast, and is designed to render full motion video comfortably in video realistic games. It has been shown rendering no less than 48 separate TV screens concurrently and IBM is using it in the creation of one of the world’s largest supercomputers.
The chips plug together easily in twos, and each of these IBM blades will use two Cell chips. They can also plug together more loosely in far higher numbers, which leads to the supercomputer style approach. The chip was based on an original design by IBM, and improved with the help of Sony and Toshiba, and will move to 45 Nanometers in the next generation from 65 nm now. It is roughly 100 times faster than the fastest Intel chips, when used on repetitive, CPU constrained calculations.
The CodecSys encoder uses one two processor blade to handle 4 channels in and 4 channels out in HD and although both companies were very careful not to give a price indication, the Cell chips won’t cost much more than $100 to $200 apiece (the PS3 is being sold for $399 with a Blu-ray drive in it) and the Xbox processor costs around $106, and given that this is off the shelf hardware, with no special modifications, it is likely to mean that encoder manufacturers are going to be undercut by this hardware software combination, as well as being outperformed by it. This will either lead to more companies taking on the Cell chip, or radical new architectures being used by their chip suppliers in an attempt to keep up.
One of the problems in HD encoding is that most systems need to break up the screen and string together 8 or 10 DSPs in order to handle each HD stream. The work is divided and then put back together again, and that provides something of an overhead, but it’s the only way the chips can solve the decoding problems in a parallel configuration. But given that DSPs are reasonably cheap these days, the chip costs can be brought down in line with the IBM blade. But if one blade can handle 4 HD streams, then if a DSP design requires 8 chips per stream, in order to compete it will require 32 DSPs for the same price.
That’s IF they can compete. The next phase of the Broadcast International plan is to make the encoding process two phased, allowing multiple passes at encoding the video, which will all happen in real time because the PS3 is so fast. ‘We will be able to take this all the way down to 1.5 Mbps,’ when we do that,’ claimed Tiede.
It is generally understood in the encoding community that a gradual fall in the output stream size of HD will come as encoder makers get more used to the H.264 codec standard and begin to use some of its more powerful tools.
For instance the frames which are calculated from the reference I Frames can be made to refer not only to the immediately preceding I Frame, but to multiple past I Frames plus the next delivered I Frame. That has the benefit of leading to far lower bit rates, but causes issues with rendering video soon after a channel change. Also there are all sorts of movement prediction calculations that are possible which are processor intensive, which can be optimized depending upon the type of video that is being encoded.
Quality MPEG 2 streams took around a decade to fall from 7.5 Mbps to under 3 Mbps, with one or two encoder makers pushing to under 2 Mbps before adopting the next generation MPEG 4 Part 10 or H.264 technology, which has driven this down to around 1 Mbps for minimal quality SD streams.
But this step from Broadcast International and IBM takes us there in one single step, with more bandwidth reduction to come and if it can genuinely be delivered, it changes all the cost calculations for IPTV, satellite, cable, mobile TV and TV over the web ‘ overnight.
Right now IPTV suppliers like AT&T have had legendary problems with getting the quality it needs for multiple HD streams per home, when using copper technologies like bonded ADSL2+. There is quite simply too much data to get across the network in order to provide sufficient quality HDTV against the cable operators.
There are many ways of attacking the bandwidth problem ‘ and each one comes at a price ‘ The network can have a strict tier of Quality of Service implemented across it for devices which carry the video multicast, but such routers and switches are rarely perfect and when IP packets are lost, this doesn’t prevent an I-Frame from being damaged with a resulting ‘visual artifact’ or glitch on the screen. The more data that a network carries, the more dropped packets it has, the more video quality problems it has.
Another way around this is to place appliance sensors which carry out deep video packet inspection all across the network, reporting back to a network management system, to isolate sources of unacceptable jitter and re-align the network so these go away. A third approach is to use a Forward Error Correction protocol, which requires at least a 10% data overhead, so that lost packets can be re-calculated. Another approach is to have a regional packet resend utility, built into local routers ‘ an approach that Cisco is adopting. Some networks operators believe they must upgrade their access networks so that there is a lower signal to noise ratio (get rid of the copper for fiber or fast Ethernet), so that there are less lost packets.
It may be that over time a combination of all these approaches built in a hierarchy will be the only way to build IP networks which genuinely deal with all of the issues of high data volumes ‘ but there’s a surefire way of putting that day off or mitigating it somewhat ‘ simply send half or a quarter as much data and that’s what this encoder approach promises.
If Broadcast International is true to its word and drives HD down to 1.5 Mbps, then it will mean almost being able to send real time video files across the public internet. Add a little buffer at the far end (DVR or PC hard disk) and even live TV will be possible to send across the internet. Already multiple pass encoding techniques, using far less power than the Cell chip, and issuing I Frames infrequently (every few hundred frames instead of every 12) can be used to encode non-active, non-real time, video. But the vagaries of the wider internet usually mean that each piece of video has to be encoded multiple times, perhaps at 20 different data rates, in order to offer the right streaming speed for any given end to end data rate.
Most approaches to this involve storing multiple formats on a large server, and encoding them sequentially, then switching between them as requests for video files come in, something that’s very hard to do for real time video feeds. The Cell approach may well come into its own here, encoding each stream on the fly and issuing multiple resolutions, varying them as the conditions on an internet connection change second to second.
IBM and Broadcast International initially announced this co-operation back in March and followed it up with demonstrations at IBC in September.
The two companies have established a joint development center so that BI can collaborate with IBM engineers on integrating the CodecSys system onto the IBM BladeCenter QS20.
Broadcast International reckons that before it began using the Cell Blades it required something like 70 Intel x86 servers to run CodecSys making it more expensive that existing market options. Now it can fit into a single blade, anyone using this in their head end is going to find they suddenly have tons of physical space and spare electrical power ‘ always a premium in network centers.
The two say that they expect first orders in the next 90 to 120 days and we felt that the biggest issue was whether or not IBM would really push someone else’s technology. Daniel Greenberg IBM program director responded by reminding us that the technology sits on an IBM processor describing a sale as a ‘double win,’ as it will do the integration and project work too, and that IBM doesn’t have anything similar in terms of encoders. ‘We’re active in the media and entertainment sector, for instance with our Media Hub process management and DRM product lines, so we have access to some key clients here. We don’t expect to ship the types of numbers of Cell chips that we are shipping in the games console market, but if IBM gets the systems business associated with this product there are significant amounts to be chasing.’
We’ll let you know if that 3 Mbps is feasible in real world situations once someone takes a risk on the new encoders and goes ahead and installs them.