Your browser is not supported. Please update it.

10 April 2014

Ericsson, Envivio and Cisco flesh out encoding strategies at NAB

Ericsson’s recently announced Virtualized Encoding software was the elephant in the room at NAB 2014 this week, directly or indirectly shaping announcements by many other vendors. In a separate story we report how Elemental Technologies announced that its encoding has been fully integrated into the Ericsson Virtualized Encoding, keen to ensure that it will benefit from some of the anticipated deployments.

Ericsson Virtualized Encoding is a software package designed to arbitrate between encoding resources, whether these are in-house or cloud based, and select the right platform for a particular video output on the basis of priorities set by the operator, relating to quality, cost and performance.

Ericsson raised some eyebrows at the Virtualized Encoding launch by insisting that far from being dead, hardware based encoding would feature heavily and continue to be the best solution for large scale high performance applications.

But this debate is becoming rather a distraction, since it is clear from recent trends, not just in encoding, but silicon development itself, that the distinction between hardware and software is increasingly meaningless. The definition of hardware based in this context means ‘dedicated to a specific task’, as opposed to being capable of executing any given instructions.

This definition first became slightly fuzzy with the arrival of Field Programmable Gate Arrays (FPGAs), which are still dedicated to a given application, but with the ability to be configured after manufacturing. FPGAs emerged with the development of Hardware Description Languages (HDLs) for customizing chips, replacing the use of direct circuit diagrams. FPGAs made it possible to upgrade the logic to take account of changes to an encoding standard for example, although with limited scope for major enhancements because performance is constrained by the chip design.

Since then there has been increased use of custom logic in generic processors from Intel, AMD and others. Another key factor is that generic processors tend to be first to migrate to smaller process design and so in that respect benefit from the resulting improvement in performance before ASICs (Application Specific Integrated Circuits) or FPGAs, offsetting the advantages of having totally dedicated logic.

There are three factors determining a chip’s performance. The most fundamental is the process design, defined as the chip’s ‘half pitch’, meaning half the distance between two corresponding points of its memory cells. It is therefore a measure of the feature sizes and circuit distances and as it comes down the processor speed increases correspondingly. The current state of the art is 14 nm half pitch, which is about 700 times smaller than the first silicon chips just over 40 years ago in the early 1970s. It is the reduction in process size that has primarily sustained Moore’s law, which predicts that the number of transistors on integrated circuits doubles approximately every 18 months (or two years depending on who you listen too).

The law has since been applied to other aspects of computer hardware such as storage density and network bandwidth, but they all depend ultimately on continuing improvements in process design, which most experts believe cannot continue below 5nm, a point likely to be reached around 2020. This does not mean that Moore’s Law will be refuted then, but does mean that a fundamental shift in integrated circuit technology would be required, possibly using silicon photonics, which Intel and IBM in particular have been working on.

The second determinant of chip performance is its microarchitecture, which is the actual layout specifying where its components such as gates, registers and also large entities such as arithmetic logic units (ALUs) are located, the connections between them, and how they interoperate. The micro-architecture has scope for optimization and effectively connects the underlying process design with the third determinant of performance, the logic or instruction set architecture (ISA). A given ISA can be implemented on a particular process design via different micro-architectures.

A challenge for ASIC and FPGA makers lies in keeping up with the rate of process evolution as the half pitch size continues to come down. This has become harder over the last few years as Intel has upped the ante, driven by its ambition to gain a greater slice of the mobile chip market dominated by RISC (Reduced Instruction Set Computing) processors based on the ARM architecture.

From about 2008 Intel started to exploit its one major competitive advantage apart from its strong position in the desktop and server arena, of being the only significant IDM (Integrated Device Manufacturer) left in the microprocessor industry. All the others, even big players like Samsung and IBM that still do a lot of product manufacturing, have teamed up with the likes of GlobalFoundries for joint R&D. Remaining big silicon players such as Qualcomm, Nvidia, Toshiba, Broadcom and Texas Instruments, are fabless, outsourcing their manufacturing to just a few big specialists like GlobalFoundries, TSMC, or UMC.

Having end to end control from design to fabrication has enabled Intel to open up a lead in process, starting volume production of 22nm processors as early as October 2011, while most of the industry is still on 32nm, the preceding process round. Then just this month Intel unveiled a 14 nm processor dubbed “Braswell” at its Developer Conference in China, aimed at Chromebooks with a 64 bit Android kernel on board.

ASIC and FPGA producers have become acutely aware that it is imperative they stay as close to the process curve as possible in order to remain competitive and meet the challenge of generic chips. Many ASIC designs were stuck on 65nm until quite recently and have only just progressed to 45nm, the round before 32nm. However STMicro has invested heavily in process design, partly through partnership with IBM, and in 2010 announced a platform to support design and development of ASICs for networking and video processing at 32nm. The driver was the growing complexity and levels of integration required on single ASICs that could not be accommodated even on 45nm chips. Putting more functions on one chip makes it easier to handle connections between these processes than if data has be routed between physically separate silicon components. Since then STMicro has progressed towards 22nm development.

The game could soon change radically as Intel exploits its process lead to reach partnerships with fabless companies, offering its design in return perhaps for access to intellectual property that could lead to incorporation of more dedicated cores or regions in its own chips. This would diffuse the hardware versus software debate further, as chips support a growing number of dedicated processes as well as instruction sets. By luring fabless companies firstly to its 22nm and then 14nm processes, Intel would be well placed to cherry pick technologies it wants to license and integrate into its own chips.

The proliferation of cores within single chips is also diffusing this debate. Moore’s Law has long been a self-fulfilling prophecy in that it has actually been adopted to drive processor innovation, instead of merely being a consequence of it. In order to obey Moore’s Law, chip makers used to keep on increasing clock rates in line with decreasing feature size through process improvement. But this became unsustainable over a decade ago, because it became increasingly hard to keep devices cool. The rough rule as stated by the Institute of Electrical Engineers at the time was that for every 400 MHz increase in clock speed, power consumption rose 60%. So the industry embarked on a new journey towards increasing numbers of cores, essentially multiple chips on the same die and socket benefiting from much faster inter-chip communications than if they were separately attached to a circuit board. The main challenge lay in efficient parallel processing of tasks requiring more than one core, although of course the approach scaled perfectly for applications executing in only a single core.

But as the number of cores proliferated, overheating became a problem once again. Several strategies have been adopted to mitigate this, one being throttling, meaning dynamically scaling of a core’s clock frequency to reduce power and therefore heat production when it has less work to do. But this is not particularly efficient, reducing the amount of computational work done per watt of power, so in terms of the energy budget is counterproductive.

A more innovative and promising approach has emerged from the University of San Diego with its Greendroid project, which replaces frequently executed program with hard-coded circuits. The effect is to populate the silicon with cool and highly efficient ‘helper cores’ instead of duplicating ever more cores that can never be fully utilized. This is surely the return of ASICs via the backdoor.

Coming back to encoding we see such trends at the higher application level. Envivio has just announced at NAB a revamped approach much more closely aligned with Intel’s multicore architecture in order to provide the power needed to execute 4K Ultra HD HEVC (H.265) compression on commodity servers. HEVC encoding is distributed across multiple server nodes, while also exploiting hardware features on individual cores to accelerate performance. Envivio’s Muse Ultra HD can encode either one 4Kp60 (progressive video at 60 frames per second) or two simultaneous 4Kp30 (30 frames per second) video signals in HEVC format on a single 2-RU Intel-based server.

With this development Envivio claims that by being able to exploit the latest Intel process technology it at least matches or even exceeds the performance of comparable ASIC based systems and at higher density through making use of compact commodity servers without need for dedicated circuitry. The one concession Envivio still makes is that generic processors consume more power for a given task than the ASIC equivalent. It is unlikely that power gap can ever be closed entirely given than an ASIC’s circuitry can be optimized just for the one task, minimizing the total electrical path length on the device. A shorter path means less power is dissipated as heat. But as we have seen perhaps ASICs can effectively be deployed as cores on generic processors.

Meanwhile Harmonic has taken a different tack, with its admission that software encoding alone cannot quite cope with the heavy computational demands of HEVC and so decided it needed a little help from a dedicated engine. Accordingly at NAB Harmonic announced it had chosen Altera Corporation’s new 4Kp60-capable H.265 enhanced motion estimation engine (EME) based on FPGAs as part of a server co-processing package also featuring its own HEVC software. Harmonic claimed this will enable the system to evolve, while also offloading the computational intensive aspects of HEVC encoding to the dedicated FPGAs.

Other encoding developments at NAB seemed to emulate Ericsson, with Envivio itself having integrated core software encoders and network media processors into the OpenStack cloud software, an open source project with more than 200 companies as contributors. Envivio is also partnering with OTT streaming platform, 1 Mainstream, to deliver a turnkey end-to-end cloud-based video service. As well as encoding, these services include packaging, DRM, billing, ad mediation, content delivery network (CDN) and apps. Video delivery is currently available on platforms such as Apple TV, XBox, Roku, Amazon Fire TV, iOS/Android tablets, and Samsung Smart TVs.

Envivio also announced its encoding has been fully virtualized and commercially deployed at large scale by a leading unnamed US cable operator. This is almost certainly Comcast, which alone accounted for 22% of Envivio’s Q4 2013 revenue having only come on board early that year. The only other possibility is Time Warner Cable, Envivio’s other major US MSO customer and being taken over by Comcast anyway. Elemental took away a lot of the Envivio business at Comcast, so we suspect this is Time Warner.

Cisco is also moving cloudwards, announcing plans to virtualize and cloud-enable the video processing elements of its Videoscape TV service delivery platform at NAB. This included enhancements to its Videoscape AnyRes encoding solution to support full-frame rate 4K/Ultra High Definition HEVC encoding, but sticking entirely to Intel blade servers without any hardware assistance.