This paper presents the Aegis Tensor Processing Unit (A-TPU), an example of a Domain-Specific Architecture (DSA). Such an architecture overcomes the physical and security limitations of modern general-purpose computing. At present, general-purpose processors encounter two major bottlenecks: thermal power constraints caused by ‘Dark Silicon’ and hardware-level vulnerabilities due to speculative execution, for instance, ‘Meltdown’. As A-TPU avoids the von Neumann fetch-execute cycle and speculative execution, it offers immunity at the hardware level against transient side-channel attacks. The design employs a deterministic systolic array to maximize operations per cycle and to eliminate thermal power limitations. The design delivers better performance per watt for machine learning inference workloads and guarantees physical memory isolation in multi-tenant cloud environments.
Alan Jethro Ecuacion (Sun,) studied this question.