Applicability of a specialised deep-learning processor for scientific computing

Can special AI/deep-learning processors also be used for scientific computing workloads? This question will be investigated by the master project.

Specialised AI/deep-learning hardware is the latest addition to the cutting-edge computing family. Although such hardware is spearheaded for the deep-learning workloads, several of the architectural features should also be beneficial for the conventional scientific computing workloads. These beneficial architectural features include ultra-fast connections between the processing units and a more straightforward (and thus more efficient) memory system. On the other hands, there are "limitations" with the specialised deep-learning hardware, such as support for only half- and single-precision calculations, as well as the possible requirement of an unconventional programming paradigm.


The newly established national infracture eX3---an experimental platform for exploring exascale computing---is expected to include during its second phase a specialised hardware component for deep learning/AI. This provides a unique testbed for answering the above question: Can a specialised deep-learning processor also be used for scientific computing workloads? The master candidate is expected to adapt the specialised programming paradigm and/or software libraries with respect to typical workloads of scientific computing. Effort is also needed to study the obtainable computational performance, as well as identifying the bottlenecks.

Learning outcome

A successful candidate will establish her/himself as one of very first people in the whole world who have gained experience and expertise for this new type of unconventional but increasingly popular hardware for deep learning. It also bears a great chance that the candidate may provide valuable contributions to the scientific computing community. In addition, the candidate will gain substantial knowledge of parallel programming and high-performance computing. All these knowledge and skills are deemed as an important component of the expertise needed by the future workforce of scientific/technical computing.


The candidate is expected to be skilful in technical programming (experience with parallel programming is not required, but preferred). Very important: The candidate must be hard-working and bear a pioneering spirit for learning, testing, and adapting drastically new skills and knowledge.


  • Xing Cai
  • Are Magnus Bruaset


Contact person