Biresh Kumar Joardar

The availability of different core architectures (CPUs, GPUs, NVMs, FPGAs, etc.) and interconnection technologies (e.g., TSV-based stacking, M3D, Photonics, Wireless etc.) has revolutionized high-performance hardware design. However, the resulting diversity in the choice of hardware has made the design, evaluation, and testing of new architectures an increasingly challenging problem. Each computation/communication element has its unique set of requirements that need to be satisfied simultaneously for overall power, performance and area benefits. Existing heuristic-based solutions are not scalable and often lead to sub-optimal outcomes. ML techniques can be used here to solve this problem. By learning the design space of all possible solutions, ML can lead to better results much faster than traditional methods. This will reduce design time and lead to better architectures in future.
Relevant publications:

ML has become ubiquitious in real life with applications in healthcare, recommendation systems, self-driving cars, etc. However, ML algorithms (particularly Deep Learning) is computationally challenging from the perspective of hardware. General purpose cores like CPU and GPU are not optimized for these applications, leading to sub-optimal performance. Conventional ReRAM-based PIM architectures are promising but have several shortcomings, such as lack of normalization and high precision support. We can address these challenges using heterogeneous architectures such as AccuReD and ReGraphX (cited below).
Relevant publications:

New and emerging technologies like Processing-in-memory, 3D integration and ReRAMs promise significant power-performance benefits for ML applications. However, due to immature fabrication technologies, they are prone to failures and other non-ideal effects. This makes these architectures highly unreliable despite their advantages. Hence, it is important to develop error-tolerant ML applications that can deliver equal prediction accuracy even when the underlying hardware is faulty/unreliable. This will not only promote the widespread adoption of these new and emerging technologies but also lead to faster ML applications.
Relevant publications: