Deep neural networks (DNNs) have achieved remarkable success across a wide range of applications. However, their massive computational requirements have motivated the exploration of more compute-efficient alternative operations. Many of these emerging...
Deep neural networks (DNNs) have achieved remarkable success across a wide range of applications. However, their massive computational requirements have motivated the exploration of more compute-efficient alternative operations. Many of these emerging operations, while reducing arithmetic workload, exhibit low arithmetic intensity and thus tend to be memory-bound on modern hardware. As the attainable memory bandwidth lags far behind the available compute throughput, these algorithms increasingly suffer from the memory wall, which significantly hinders their performance.
This thesis presents acceleration techniques for two emerging memory-bound operations in DNN workloads, both leveraging kernel fusion. The first target is the tensor-product operation used in graph neural networks for predicting atomic energies in molecular dynamics simulations. To accelerate this memory-bound computation, we employ a fusion strategy that combines multiple kernels while exploiting computational sparsity and input reuse. The second target is the state space model-based convolution operation, an emerging alternative to self-attention in large language models. For this case, we design a custom accelerator architecture that supports real-time operand decompression, thereby reducing the required on-chip SRAM capacity for the fused kernel and improving the overall performance and energy efficiency.