A report put together by the European Network on High Performance and Embedded Architecture and Compilation (HiPEAC) argues computing is at an architectural turning point that will lead to the widespread use of dedicated accelerators as a stopgap effort to close the gap between user demand and the ability to supply.
“We’ve been headed here for a while, but it’s becoming increasingly clear that we can’t keep shrinking components while increasing performance,” said Marc Duranton, editor-in-chief of the HiPEAC Vision report. “Acceleration for specific applications is the short-term route to performance gains, while we investigate new paradigms such as neuromorphic and quantum computing, which will however complement, rather than replace, silicon semiconductor technology.
“Computing technologies now form a continuum, from the devices at the edge we interact with daily, such as virtual assistants, smartphones and medical wearables, to the largest data centres and high-performance computers,” Duranton added. “Computers are also increasingly making themselves smarter through automated accelerator development and intelligent software design.”
Software and energy
Although much will depend on the development of novel hardware, the report authors argued software development should consider energy as “a first-class property” and that applications should be aware of their energy consumption at runtime. Not doing this will continue to lead to tremendous inefficiencies, they added.
The team said systems and software developers should also consider obsolescence and the maintenance of their applications as they gradually turn into legacy systems. In an environment where systems such as autonomous vehicles and IoT sensor and actuator nodes will be in use for a decade or more, the problems of keeping them running will force changes in software architecture.
The biggest change they envision is much greater levels of automation in software production. This may be the only way to create efficient code for heterogeneous systems that combine general-purpose processors with task-specific accelerators. Another approach is to make use of open-source implementations that allow for easy adaptation.
The increased use of machine learning and specialty systems such as quantum computers will also lead to a demand for engineers to acquire better interdisciplinary skills. In quantum computing, for example, the software behavior is bound tightly to physical behavior at the microscopic level. Similarly, neuromorphic computing is increasingly borrowing from concepts in cell and brain biology.
For European organizations specifically, the HiPEAC report was tuned to their commercial realities, particularly their reticence to invest the large sums of money needed to keep pace with leading-edge CMOS processes. This leads to a situation where investment in mature nodes and multihcip packaging for edge devices rather than servers is, in the authors’ view, likely to be more successful. This dovetails with a culture that is much more sensitive to issues of privacy and safety.
“Therefore, it should build on its strengths and become a leader in intelligence at the edge, allowing better control both of the devices and of privacy,” the authors wrote. “In those fields, the most advanced CMOS technology is not always the best choice, because of its cost and lack of long-term reliability.”