Research into memory in streaming dataflow systems
Bachelor’s Thesis / Master’s Thesis / Student Research Project
Abstract
With the prevalence of research into domain-specific accelerators for different machine learning (ML) techniques, integrating many different heterogeneous components is becoming an increasingly impolrtant consideration in platform design. Efficient data movement and buffering present a key challenge in the design of such platforms.
Two major areas of interest in this regard are the design of memory subsystems and custom on-chip buffering for the accelerators themselves, as well as the direct, streaming communication between accelerators, I/O, memory and other components. Against this backdrop, we have developed GOURD, a framework for dataflow-driven hardware design of streaming platforms (see also this thesis topic.
Some research topics that are relevant to this issue and might form the basis for a thesis or project:
- Use of decoupled access-execute architectures for data buffering in streaming contexts
- Extension of the GOURD framework with descriptions of memory interactions and accelerator-private memories
- Extension of the GOURD FPGA execution framework for memory behaviour analysis
References
- Paper: “GOURD: Tensorizing Streaming Applications to Generate Multi-Instance Compute Platforms”
- Paper: “Buffets: An Efficient and Composable Storage Idiom for Explicit Decoupled Data Orchestration”
- MLIR
- SystemVerilog
Requirements (some of, depending on the topic)
- Linux and Git (generally necessary)
- C++/Python (for the extension of GOURD)
- SystemVerilog (for any hardware implementation work)
- Understanding of computer architectures and memory systems
- Successfully attended the lecture “Grundlagen der Rechnerarchitektur” (recommended)
- Successfully attended the lecture “Advanced Computer Architecture” (recommended)
- Successfully attended the lecture “Modellierung und Analyse von Eingebetteten Systemen” (recommended)
- Successfully attended the lecture “Digital Design and Synthesis of Embedded Systems” (recommended)