For Embedded Applications with Data-level Parallelism, a Vector Processor Offers High Performance at Low Power Consumption and Low Design Complexity. unlike Superscalar and Vliw Designs, a Vector Processor Is Scalable and Can Optimally Match Specific

Designers of embedded processors have typically optimized for low power consumption and low design complexity to minimize cost. Performance was a secondary consideration. Nowadays, many embedded systems (set-top boxes, game consoles, personal digital assistants, and cell phones) commonly perform computation-intensive media tasks such as video processing, speech transcoding, graphics, and high-bandwidth telecommunications. Consequently, modern embedded processors must provide high performance in addition to low cost. They must also be easy to scale and customize to meet the rigorous time-to-market requirements for consumer electronic products. The conventional wisdom for high-performance embedded processors is to use the superscalar or very large instruction word (VLIW) paradigms developed for desktop computing. Both approaches exploit instruction-level parallelism (ILP) in applications in order to execute in parallel a few operations per cycle. Superscalar processors detect ILP dynamically with hardware, which leads to increased power consumption and complexity. VLIW processors rely on the compiler to detect ILP, which leads to increased code size. Both approaches are difficult to scale because they require either significant hardware redesign (superscalar) or instruction-set redefinition (VLIW). Furthermore, scaling up either of the two exacerbates their initial disadvantages. This article advocates an alternative approach to embedded processors that provides high performance for critical tasks without sacrificing power efficiency or design simplicity. The key observation is that multimedia and telecommunications tasks contain large amounts of data-level parallelism (DLP). Hence, it’s not surprising that we revisit vector architectures, the paradigm developed for high performance with the large-scale DLP available in scientific computations. Just as superscalar and VLIW processors for desktop systems adjusted to accommodate embedded designs, we can revise vector architectures for supercomputers to serve in embedded applications. To demonstrate that vector architectures meet the requirements of embedded media processing, we evaluate the Vector IRAM, or VIRAM (pronounced “V-IRAM”), architecture developed at UC Berkeley, using benchmarks from the Embedded Microprocessor Christoforos E. Kozyrakis