Self-Sketching Domain Specific Accelerators: Build Hardware from Software

虽然它并不总是像开源软件和应用创新那样明显，但计算硬件一直是机器智能和人工智能的关键推动者，自1987年以来，它的性能提高了5000倍。专家认为，对于机器智能的进步，更快的计算硬件几乎和神经网络一样重要。计算性能的快速增长可以推动虚拟现实、自动驾驶汽车和神经网络的创新，帮助解决21世纪的一些重大挑战。

While it is not always as apparent as open-source software and application innovation, computing hardware has been a key enabler for machine intelligence and AI, delivering 5000$\times$ performance improvement since 1987. Experts rank faster computing hardware as being nearly as important as neural networks for advancements in machine intelligence. Even more rapid growth in computing performance can fuel innovation in virtual reality, autonomous vehicles, and neural nets helping solve some of the grand challenges of the 21st century. Unfortunately, the hardware industry faces a stiff challenge, today while we get more transistors on a single chip, chips themselves are more expensive (\$/$mm^2$) and transistors are energy inefficient; these limit performance. The lack of a clear forecast for technology and the rise of AI has entangled companies in the messy task of creating their own custom accelerator chips to deliver the requisite performance. Unfortunately, it requires significant effort (multiple years) and money (10s of million dollars) for developing the accelerator chip and software bring-up, inaccessible to anyone else but to the largest vendors. It is unclear how accelerator development can track the rapid pace of software and application evolution. My long-term goal is, how to develop systems to make up for the lack of technology scaling. Our goal is to accelerate system architecture innovation and make it sufficiently open and inexpensive that anyone (even small vendors) can build hardware anywhere. To achieve this goal and navigate the uncertain hardware customization landscape we propose to "build hardware from software". We will be developing open-source tools and compilers, which similar to generating an CPU binary, will now generate the specification for the hardware accelerator. The key novelty of our work is the observation that hardware design should separate application domain insights from the low-level implementation details, and that the domain insights can be captured from the data types in the program. This will enable our tools to generate hardware that is reusable, shareable across algorithms, optimizable using a tool, and can be tuned by a feedback-driven process. This work will fundamentally change how systems research is done. When CPUs and GPUs dominated the hardware landscape, software has assumed hardware to be a fixed design that cannot be changed. Our work leads application experts and software developers to question the status-quo and accelerates the movement towards customization.