The historical increases in computing performance and reductions in power consumption, size, weight and cost of computing devices are ingrained in the fabric of the research community. These improvements have fueled innovation across scientific domains, health, industry, and government. But clock speeds have been relatively flat for over a decade and current transistor scaling will be reaching limits that are both technological and economic over the next decade. While alternative models of computation are being explored, including quantum and neuro-inspired, digital computing will remain the dominant computational technology for the foreseeable future.

To address the continued demands for more computing, specialization is one of the few levers left at the architectural level, as is evident from the recent development of processors for convolutional neural nets. Open source hardware models, new tools to aid in chip design and testing, and emerging chiplet models for fabrication all support the growth in specialization, but how to build robust and performant software ecosystems in this environment remains a challenge. Our hypothesis is that effective specialization will benefit from input from experts in applications that we seek to accelerate, as well as experts in programming languages, compilers, and architecture. This workshop will mix all of this expertise to seek to generalize previous point successes like accelerating deep neural networks with the TensorFlow API and the Tensor Processing Unit hardware. Following introductory presentations of trends and emerging ideas in hardware, software and applications, the workshop members will break into vertically integrated teams and develop a set of case studies for specialization.  

Participants will identify computer science challenges and possible approaches to the problem of building correct, maintainable and efficient software for future digital architectures using a set of case studies to explore these ideas. Increased parallelism at finer scales and locality will be major cross-cutting themes. The case study topics may shift somewhat based on workshop participant interests, but are likely to cover: machine learning beyond convolutional neural nets; streaming image analytics; high performance scientific simulation and analysis; and bioinformatics / medicine.

The workshop will explore the research questions in each area in depth and produce a report on the workshop’s findings:

• What type of architectural model would accelerate an important application or set of applications in the area?
• What approaches could be used to program those applications?
• How would code generation be optimized for the hardware?
• Is there a new notion of portability?
• Are there particular correctness, privacy or security concerns to be addressed, and if so, how?
• How broad is the class of applications that could benefit from this approach?