A quantitative analysis framework to evaluate the performance and costs of computer architectural alternatives

Computer architects are currently confronted with the task of evaluating the quality of architectural designs not just in terms of the single familiar criterion of performance but also in terms of design criteria such as die area, power consumption and energy consumption. In this dissertation, I propose and evaluate a quantitative analysis framework, implemented in the Solomon software tool, that can precisely characterize how the best architecture design in a study changes across the full spectrum of architectural contexts which describe the importance of each criterion to the analysis. Just as computer architects currently use average performance to compare architectures when multiple benchmarks are important, the analysis framework I propose uses an architectural cost index value to compare architectures by aggregating multiple criteria in accordance with their importance to the architectural context. The architectural cost index for each architecture is then graphed across all architectural contexts to portray how the best architecture changes in different contexts. To compute an architecture's cost index, I propose a new aggregation function called the trade-off sum that represents an architectural context as a linear trade-off such as: an N% area increase is expected to improve performance by at least 0.5N%. To evaluate the analysis framework, I conduct a major architectural study to determine how effectively 21 vector architectures use additional area to improve media program performance. To compute the area for these architectures, I develop a functional area modeling methodology that is accurate to within 14% of the measured area of circuits in commercial microprocessors for eight major processor components and within 5% for five of these eight components. Architects can use the Functional Area Modeling Tool for Architects (FAMTA) to automatically compute the area of processor components from high level descriptions, much like a compiler allows developers to write software in a high-level language rather than machine language. Using the new analysis framework, I demonstrate that current media hardware designs are best in contexts where an N% area increase is expected to improve performance by 0.98N%–1.39N%. When the improvement can be less than 0.98N% (performance is more important), then vector media architectures, which incorporate vector processing technology, use area more effectively to improve performance than current designs.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.