Production switchers

It is interesting to note that video switchers are not sold as digital switchers anymore. Today, it's assumed that all switchers are digital, but of course there are several factors driving that change in production technology.

The difference is the software

Years ago, when Abekas was still selling innovative digital production switchers, I commented on how building complex digital system products must be expensive. An Abekas executive admitted that, in reality, it was easier to build a digital switcher than an analog one. He said that the final test of a digital switcher was a matter of burn-in to get rid of infant mortality and then a verification that the hardware worked. The result was a significant reduction in labor cost, though perhaps an increase in R&D cost. Only a cursory system test was needed after each board was checked in a test jig, which exercised all of the software.

Analog switchers required complex and detailed manual setup of a myriad of adjustments, but digital switchers either worked flawlessly or they didn't. Or perhaps more accurately, the software worked repeatedly, or it didn't.

My point is that digital switchers are just as complex as analog switchers, but the adjustments to the operation of the switcher are made at the time the software is written and tested. There is no shortness of flaws, or bugs, in any software system. When parts of the system are written in DSP code, as is the case in most modern switchers, the manufacturer must test extensively for the quality and technical accuracy of the computations done on a large amount of data, every second.

Think about the amount of data that must be processed. Let's say a switcher has four sources on-screen simultaneously, and let's assume that only 40 percent of each picture contains active pixels on the composited screen at any one time (taking into account keys and picture manipulations). The output picture would contain less than 2Gb/s of content and be made from input streams totaling less than 5Gb/s. If each output pixel requires five calculations total, which is easily the case with layered effects, the total calculation capability of the system must be in the neighborhood of 6 billion operations per second. In reality, the number of individual operations done is far greater, with deserializing inputs, processing data in parallel and formatting of the output stream each requiring many operations.

Of course, with the capabilities of modern processors, this should be a piece of cake. Put enough processing power in the loop, and you can jam almost any number of inputs into a production switcher and perform an arbitrarily large number of calculations needed to manipulate each source in the output stream of bits.

However, there is a fact that must be taken into account. Production switchers are isochronous devices (i.e. real time), and the latency of the system must be constrained to low values to make practical use of live sources in a production. Most modern switchers process the full signal path in one video frame, about 33ms. One input or 10 on the screen cannot change this latency, or the system will not work in a real-world implementation.

In practical terms, this would not be possible without specialized hardware processors. Each input is deserialized in real time with low latency, and each process must equally be done in a short time period. Manipulations that change the picture size or geometry, often outboard processes called digital video effects, are now inboard processes done sometimes on boards with that function alone.

Keys and other pixel-level manipulations are done in DSP, operating at truly astounding speed. Even more amazing to me is that many processes must happen in a series, stacking up calculations and shortening the available time period to produce the final output pixels.

Supporting software

Behind all of this must be two other supporting software systems. One real-time system moves the data from processing element to processing element, and must do so with tight tolerances, considering that each pixel is less than 1ns long. To provide a reference, light moves about 8in per nanosecond.

The other software system can virtually loaf in comparison, for it controls the operator interface and coordinates communication between the control panel and the processing engine(s). That system is fast enough if it can reliably move commands from the real-time control panel to the processor fast enough to make a cut happen on the next frame after the button is pushed. The shortest time period for this synchronization of operator interface and processing power is between 1ms and 16ms, during which time perhaps a million pixels or more could be processed to the output of the switcher.

Today's capabilities

The best news is that this technology has been available for a generation and has gotten steadily better. One of the first all-digital production switchers, the Thomson Grass Valley Kadenza, was introduced 20 years ago. It required parallel digital connections (SMPTE 125M) but offered some innovative approaches, such as either a layer approach or a more conventional mix/effects orientation. Over the past 20 years, the cost has come down, and the capability has gone up. The Kadenza was 525/625 only, in no small measure because digital interfaces for HD signals weren't available. Current products offer HD or SD capabilities, and more than one manufacturer offers processing of (essentially) any input format, SD or HD, in the same output stream.

I can clearly see a trend toward picture format agnostic processing in production switchers. This capability comes quite naturally as a result of the special purpose scaling engines that are available in so many products today, from HD/SD format converters to aspect ratio converters and up/downconverters. Scaling engines are modest cost options in cameras, VTRs, frame synchronizers and other devices. Modular products occupying one or two slots in a card frame can now perform scaling and format conversion that once took many boards in expensive special purpose hardware.

The Holy Grail

The logical conclusion is that product differentiation will lead to increasingly flexible capacities in switchers and long-term reductions in cost. At some point in the future, the Holy Grail of a blade server, which performs the functions of a production switcher, will be achieved, but for now, it seems out of reach for practical and affordable baseband real-time production switchers.

John Luff is a broadcast technology consultant.

Send questions and comments to: