Algorithm/architecture codesign of low power and high performance linear algebra compute fabrics

In the past, we could rely on technology scaling and new micro-architectural techniques to improve the performance of processors. Nowadays, both of these methods are reaching their limits. The primary concern in future architectures with billions of transistors on a chip and limited power budgets is power/energy efficiency. Full-custom design of application-specific cores can yield up to two orders of magnitude better power efficiency over conventional general-purpose cores. However, a tremendous design effort is required in integrating a new accelerator for each new application. In this dissertation, we present the design of specialized compute fabrics that maintain the efficiency of full custom hardware while providing enough flexibility to execute a whole class of coarse-grain operations. The broad vision is to develop integrated and specialized hardware/software solutions that are co-optimized and co-designed across all layers ranging from the basic hardware foundations all the way to the application programming support through standard linear algebra libraries. We try to address these issues specifically in the context of dense linear algebra applications. In the process, we pursue the main questions that architects will face while designing such accelerators. How broad is this class of applications that the accelerator can support? What are the limiting factors that prevent utilization of these accelerators on the chip? What is the maximum achievable performance/efficiency? Answering these questions requires expertise and careful codesign of the algorithms and the architecture to select the best possible components, datapaths, and data movement patterns resulting in a more efficient hardware-software codesign. In some cases, codesign reduces complexities that are imposed on the algorithm side due to the initial limitations in the architectures. We design a specialized Linear Algebra Processor (LAP) architecture and discuss the details of mapping of matrix-matrix multiplication onto it. We further verify the flexibility of our design for computing a broad class of linear algebra kernels. We conclude that this architecture can perform a broad range of matrix-matrix operations as complex as matrix factorizations, and even Fast Fourier Transforms (FFTs), while maintaining its ASIC level efficiency. We present a power-performance model that compares state-of-the-art CPUs and GPUs with our design. Our power-performance model reveals sources of inefficiencies in CPUs and GPUs. We demonstrate how to overcome such inefficiencies in the process of designing our LAP. As we progress through this dissertation, we introduce modifications of the original matrix-matrix multiplication engine to facilitate the mapping of more complex operations. We observe the resulting performance and efficiencies on the modified engine using our power estimation methodology. When compared to other conventional architectures for linear algebra applications and FFT, our LAP is over an order of magnitude better in terms of power efficiency. Based on our estimations, up to 55 and 25 GFLOPS/W single- and double-precision efficiencies are achievable on a single chip in standard 45nm technology.