These days, a number of companies are ferociously burning venture capital in order to develop new programmable architectures that offer a better tradeoff in the performance/power/cost space than existing devices such as traditional von Neumann processors or FPGAs. Typically, these architectures attempt to parallelize the processing problem with cleverly connected redundant hardware. In order to obtain their venture capital, these companies produce sophisticated and complex PowerPoint propaganda describing in detail (or so it seems) how their path to parallelism outperforms existing architectures by factors ranging from two to two thousand.
Savvy venture capitalists perform their due diligence, bringing in digital design divas of their own to review the radical claims of the would-be fundees. Often, these claims triumph in the test of the third-party expert with convincing theoretical wins against the status quo. Generally, however, the focus is on the hardware architecture itself. When it comes to the question of “How do you program it?”, the PowerPoint prognostications become somewhat pixilated. This question, however, is the Achilles’ heel of the parallel processing revolution.
Customers see the promise of unbelievable performance at a price and power that they’ve never imagined, only to have their smiles fade when the discussion turns to the programming model. Performance never comes for free, it seems, and in most of these cases the real cost is extracted from the software developers. New programming “paradigms”, “revolutionary” languages, and “extensions” are all among the euphemisms offered to wary software developers instead of the more realistic “You’ll have to re-learn your craft from scratch.” Once real-world problems such as legacy software and legacy programming skills enter the picture, however, the market extracts a heavy toll on these pretenders to the parallel-processing throne.
Stretch Inc. is making a series of announcements this week that show a strategy for spanning that critical software development gap and pushing a new, radical hardware architecture through to commercial success. Like many of these companies, Stretch spent years extolling the benefits of their silicon strategy – roughly based on an array of Tensilica Xtensa processors with FPGA-like programmable hardware extensions and interconnect. Also, like many companies in this arena, they experienced only marginal market success with their reassurances that software developers could “create application-specific instructions from their C/C++ program using easy and familiar C-based tools.”
Stretch’s next step was also similar to that of many others – they set about focusing on particular vertical markets with some level of explanation of how the architecture and tools could be put to work solving recognized, high-value problems. When marketing folks try to match up massive processing performance demands with even more massive market opportunities, many of them end up with problems like H.264 encoding rolling off their tongues. Stretch was no exception, and they produced an H.264 “video development platform” that built a partial bridge between the performance-starved development community and the lifeline of parallel processing silicon. This strategy worked better but still didn’t produce the massive market uptake you need to support a high-growth fabless semiconductor business.
Now, Stretch has built the rest of the bridge. Instead of continuing to work at convincing developers that their architecture is easy to program and integrate, they’ve buckled down and done the work themselves. The result is a series of soup-to-nuts solutions for the surveillance video market, estimated at over 30 million cameras shipping this year, with explosive growth over the coming years. In their latest announcement, stretch is showcasing a 16-channel H.264 digital video recorder in a PCI Express add-in card and a high-definition multi-standard intelligent IP camera reference design kit.
The bulk of the security industry still uses analog cameras, but the picture is changing rapidly. The cost advantages of doing encoding at the camera so that data IP standards like Ethernet can be used for transmission are enormous, as commodity Ethernet (or wireless broadband connections) are far more economical than stringing analog A/V cabling over a large facility. In order to support both the analog and IP camera camps, Stretch integrated their devices into both a server-side encoder (the 16-channel PCIe card) and a camera-side encoder (the IP camera reference design kit). Both solutions include boards with Stretch processors and all required peripherals, reference designs, and tools for customizing the solution for specific end-product requirements.
Stretch’s solutions go far beyond the usual encoder-only functionality. That can be had by commodity ASSPs at a relatively low cost. Because of the processing power and flexibility of the Stretch processor, pre-processing intelligence that takes effect prior to encoding can be built into the device. In order to save bandwidth and server storage space while capturing and preserving the maximum amount of useful information, analytics can be built into the system to detect scene changes, blind cameras, motion in the frame, etc. When the scene has no interesting activity, low frame rates and low resolutions can be captured. Then, when something interesting happens – motion in a certain area for example, the frame rates and capture resolution can be increased dramatically.
Of course, the algorithms for such analytics are highly application dependent. This is where the Stretch solution shines. The programming interfaces and tools that previously were required for development of the entire application can now be razor-honed to serve only the high-value analytics portion. OEMs can buy Stretch’s turnkey solution – including boards that could be taken to production, and spend engineering resources only on their value-added analytics portions. The power of the parallel processing elements works quietly in the background, enabling the basic cost/performance ratios that make the solution appealing.
Stretch’s approach to marketing high-performance programmable silicon should be a case study for the rest of the industry. If Stretch’s play works, it is easy to see how additional turnkey solutions could be created serving other adjacent markets with exactly the same silicon platform. It is also easy to imagine OEMs taking Stretch’s processor platform and creating their own solutions to serve specific, high-growth markets.
While both traditional processors and FPGAs enjoy the legacy of a strong ecosystem of tools, IP, infrastructure, and engineering expertise around their architectures, new and innovative silicon platforms like Stretch’s have to find a way to survive and profit in the market without that benefit. This turnkey board plus tools plus IP approach promises to punch through the ecosystem barrier in much the fashion envisioned by Geoffrey Moore’s “bowling pin” approach to market penetration. If Stretch’s security cameras knock down the first pin, you can look for a lot of programmable platform vendors to follow suit.