580 California St., Suite 400
San Francisco, CA, 94104
Key research themes
1. How do hardware architectures including DSPs, FPGAs, and processors optimize real-time image and multimedia processing?
This research area investigates the design and comparative capabilities of Digital Signal Processors (DSPs), Field-Programmable Gate Arrays (FPGAs), and processors with specialized architectures in accelerating image and multimedia processing tasks. It is critical because real-time applications require high-throughput, low-latency processing, which general-purpose CPUs often cannot meet. The research explores architectural trade-offs in computation speed, memory bandwidth, programmability, and power consumption across hardware platforms tailored for such workloads.
Key finding: The paper comprehensively reviews modern DSPs, FPGAs, and GPUs, highlighting technical differences and performance considerations pivotal to computer vision tasks. It identifies that DSPs offer ease of programming and energy... Read more
Key finding: Introduces the Imagine architecture optimized for media applications, employing a stream programming model and hierarchical bandwidth management to efficiently harness the data access patterns and parallelism inherent in... Read more
Key finding: Proposes a novel 2D array parallel architecture combining FPGA and DSP processors connected via FPGA-based video-addressing units enabling video-rate inter-processor communication. Designed to handle multiple simultaneous... Read more
Key finding: Develops an Enhanced Direct Memory Access (EDMA) controller integrated within AI-based DSP System-on-Chips (SoCs) that alleviates processor load by enabling high-throughput, real-time bulk memory transfers without CPU... Read more
2. How can compiler and architecture co-design enhance FPGA-based acceleration of image processing pipelines?
This theme examines methods for efficiently compiling image processing algorithms onto FPGAs, balancing programmability and performance. Domain-specific languages (DSLs) and high-level synthesis (HLS) frameworks are leveraged to generate optimized hardware pipelines that exploit data locality and parallelism within stencil and pipelined computation. The research highlights strategies for abstracting complex hardware details to streamline development while achieving near-peak resource utilization and throughput on modern FPGAs.
Key finding: Presents an FPGA backend for the PolyMage DSL, enabling the automated compilation of image processing pipelines into pipelined FPGA implementations via integration with Xilinx Vivado HLS. The approach utilizes domain-specific... Read more
Key finding: Introduces an optimization algorithm to analyze and parameterize auto-increment/decrement address arithmetic support in DSP architectures with multiple address registers and varying increment ranges. The method quantifies... Read more
Key finding: Describes the design and fabrication of DVINO, an open-source RISC-V vector processor integrating a vector processing unit capable of handling large vectors with multiple element widths, and supporting Linux boot. By... Read more
3. What architectural techniques and pipeline designs improve high performance and low power operation in DSP processors?
Research under this theme focuses on microarchitectural innovations such as pipelining, clock gating, specialized computational blocks, and addressing modes aimed at enhancing throughput and energy efficiency in DSP processors. This includes studies on pipeline organization, multiplier/accumulator optimization, bus and register file architectures, and asynchronous pipeline microarchitecture designs, all to meet the real-time processing demands of signal-intensive applications while minimizing power and silicon area.
Key finding: Develops a five-stage pipelined 16-bit RISC processor implemented on a Cyclone IV FPGA, demonstrating that instruction pipelining significantly enhances throughput and supports multiple simultaneous instructions, thereby... Read more
Key finding: Presents an asynchronous implementation of the MIPS R3000 microprocessor achieving high throughput (~280 MIPS) with low power (7W) by employing quasi delay-insensitive circuit design and fine-grained pipelining. This design... Read more
Key finding: Describes synthesis and analysis of a high-performance 5-stage pipelined MIPS-32 processor across multiple semiconductor technologies, demonstrating that critical path delay can be optimized to achieve clock frequencies up to... Read more
Key finding: Investigates low-power design methodologies based on Gate Diffusion Input (GDI) logic for key computational DSP blocks including Kogge-Stone adders, Braun multipliers, and barrel shifters. Experimental results in 45 nm... Read more
Key finding: Details the PM48dx programmable DSP chip that achieves an internal clock rate of 133 MHz with eight parallel 32-bit datapaths controlled by unified instruction streams. By integrating high-memory bandwidth synchronous SRAM... Read more