MicroUnity's mediaprocessor

The identity of Amiga Inc's "Superchip" remains a mystery. However, Chris Hanretty reckons he may well have found out which chip it might be.

Contents

The Future of Computing and Communications

Microprocessors have evolved over the last quarter century as self-contained devices for calculating and controlling things. Growth in electronics is now shifting towards interconnected devices whose primary function is to communicate. The goal of delivering the content and services of the entire global network with an ease and affordability like TV, radio, and telephone rather than PCs will impact processor evolution enough to merit a distinct category: the mediaprocessor.

How do communications algorithms differ from classical embedded and desktop applications? Classical applications typically perform arithmetic, boolean, and shift operations on a few different sizes of data (e.g. 32-bit integers; 64-bit floating point). Communications processes, on the other hand, operate on a much wider range of data widths and mathematical domains (e.g. Galois field). Encryption and error-correction codes require bit-level and Galois processing; video, RF, and modem processing need 2 to 12 bits to represent their samples; 8 to 24 bits is common for audio, and thousands of bits for packet protocols. A single sample may require hundreds or thousands of operations in the course of filtering, compression, encryption, modulation, transmission, equalization, demodulation, error correction, etc. Such high broadband rates strain both computational throughput and bandwidth of the memory system. On the other hand, the total memory required is often small, typically dominated by a few megabytes of frame storage.

Legacy microprocessor architectures have attempted to respond to these needs. Most have defined multimedia extensions that improve support for audio, video, and graphics. They are also incorporating interfaces for faster memory and real-time I/O. Backward compatibility with legacy code and interfaces, however, adds complexity and cost to these designs.

New mediaprocessor technologies aim to reduce this cost. One area of innovation is execution units that systematically and efficiently implement sub-instruction level parallelism such as SIGD (Single Instruction on Groups of Data) over all multiprecision data types. Another is programming models that eliminate redundant register files, condition codes, and mode bits to simplify code generation and streamline interlocks, bypass, and exceptions in pipelined and instruction-level parallel machines (e.g. VLIW, superscaler, and decoupled access execution designs). A third is efficient protection and synchronization mechanisms for the sharing of memory and data path resources among many user-level and secure-kernel threads of execution, enabling thread-level parallelism. A fourth is packet-oriented interfaces compatible with multiple streams of broadband traffic across few packages at low pin count.

These innovations harness parallelism inherent in communications. The winning approach is modest use of each of these bandwidth-enhancing mechanisms in a mathematically pure and concise architecture. For example, a mediaprocessor with 128-bit SIGD 4-operand instructions, 4-way issue and five threads would achieve about 10,000 bits of operand throughput per cycle, compatible with very low voltage or small driver operation needed to save power and silicon area. Efficient use of these degrees of parallelism, moreover, are within reach of current vectorization and instruction and thread-scheduling software technology.

Ultimately the dominant cost in broadband evolution will be the development and maintenance of an enormous body of software. Current microprocessor hardware and software is stretching to accomplish the audio, video, graphics, and GUI processing needed at the presentation and application layers of the communications protocol stack. Far greater challenges remain at the lower transport to physical layers where algorithms are evolving to enable broadband and wireless links in the network. The high standards of code robustness needed for these lower layers are inspiring new CASE methodologies, such as symbolic verification. The greatest economy in mediaprocessing will derive from amassing rich software development tools around a general and unified programming model that supports the entire communications protocol suite.