| Moore's law is increasingly irrelevant. While Moore's observation about technology progress in the semiconductor industry still holds, fewer and fewer chip designers find leading edge technology affordable or practical. |
Today, while high-volume standard products such as microprocessors and field programmable gate arrays (FPGAs) continue to take advantage of Moore's law scaling, an increasing number of chip designers find the application of leading-edge technology out of reach. Escalating mask costs, long design cycles and expensive electronic design automation (EDA) tools are creating serious obstacles to the use of leading-edge technology.
The far reaching consequences of this fact include the decline of application specific integrated circuits (ASICs) as a viable design alternative for systems companies. Handel Jones, of International Business Strategies, forecasts that ASIC design starts will decline 44 percent between 2002 and 2007.
Soon, some application specific standard products (ASSP) developers and fabless semiconductor companies will be hard pressed to find markets large enough to justify developing single-purpose chips.
Is this the end of the semiconductor industry? Hardly. These new conditions will usher in an era of innovation in which designers use architectural innovation to solve problems that in the past were frequently solved by moving to the next-generation process technology. Programmability will feature strongly among possible solutions.
Programmability permits fabless semiconductor companies to create single chips that can address multiple markets, thus amortizing research and development over higher volumes. Programmability has the potential to provide end users a cheaper, predictable design flow in a shorter design time as compared to traditional hardware development.
General-purpose processors will continue to be widely used, but new types of application specific processors will proliferate because of their improved efficiency. Compared to ASICs, general purpose processors use more power and offer lower performance than parallel hardware implementations. However, adding some domain-specific hardware to a processor goes a long way. Witness digital signal processors (DSPs). Adding multiply-accumulate hardware, new addressing modes and modifying memory hierarchies produced an efficient vehicle for signal processing. Programmable DSPs are now a $7.5 billion market, according to Forward Concepts.
Network processors and security processors are two more recent additions to the growing list of processors aimed at specific domain applications. New classes of application oriented processors, dubbed Application Specific Instruction Processors (ASIPs) by the University of California at Berkeley's Kurt Keutzer, are going to proliferate.
The proliferation of ASIPs will depend on the end users' ability to productively program them, yet achieve high performance allowing them to differentiate themselves from other ASIP users. Effective software development for ASIPs requires a new programming paradigm to harness the performance of these devices. For high programmer productivity, the ideal environment begins with a powerful programming language for designing applications in the target domain (such as packet processing, audio, video.
Some languages of this type already exist. For example, The MathWorks' Matlab environment and language is the de facto standard for DSP algorithm development. New languages will emerge as well. Adoption of early network processors was hindered by unproductive programming environments for their intricate architectures. New developments like MIT's Click language are aimed at providing an efficient vehicle for network-domain applications development.
While embodying high levels of abstraction, the domain-specific programming language must also be converted to an executable form that effectively exploits the inherent parallelism in the processor.
This conversion from application domain-level input to architecture-specific output is akin to synthesis in the hardware domain. Ideally, this synthesis process will effectively shield the programmer from the more complicated details of the architecture in order to maintain a high level of productivity yet maximize use of hardware resources. Companies such as Catalytic are addressing this challenge in the DSP domain, and others will emerge in the future. Innovation isn't over in the semiconductor industry — it's merely taking on new forms to meet new realities.
Andy Haines is vice president of business operations at startup Catalytic Inc.