Massive and energy-efficient computation capability, coupled with the enhanced programmability, makes Graphics Processing Units (GPUs) a desirable platform to accelerate a variety of data-intensive hard real-time applications such as medical data processing, autonomous auto navigation, and vision-based aircraft controls. The computer industry has already delivered several GPUs for embedded computing applications. However, none of these GPUs can guarantee time predictability, which is crucial for hard real-time systems. The integration of the CPU and the GPU on the same die makes it even more challenging to attain time predictability. The goal of this project is to fill this need, though investigation of the design of time-predictable integrated CPU-GPU architectures that can also offer high performance and energy efficiency.
The project addresses this problem by: (i) systematic analysis and evaluation of time-unpredictability of GPUs and integrated CPU-GPUs; (ii) exploration of time-predictable and heterogeneous last-level cache (LLC) architectures for integrated CPU-GPUs, including private, partitioned, and locked LLCs; (iii) design of time-predictable and heterogeneous on-chip interconnection networks that can efficiently support communications within and across the CPUs and GPU; (iv) design of time-predictable memory controllers to enable scalable, safe, and tight WCET (Worst-Case Execution Time) estimation; (v) development of the first WCET analyzer for the proposed integrated CPU-GPU processor to support hard real-time computing. Other project activities include efforts to transition the results of the research to industry, integration of topics and projects related to the research in courses, and open-source distribution of software.
