AEStream: Accelerated event-based processing with coroutines2023/04/11
We published a paper on the technologies behind the event-based processing library, AEStream. Specifically, we used coroutines to demonstrate that we can achieve at least 2x speedups on conventional CPUs and GPUs, compared to parallelization approaches that does not use coroutines.
The work contributes
- A plug-and-play library that streams events from files, event cameras, network to files, CPUs/GPUs, network, and Python machine learning frameworks such as Numpy, Tensorflow, PyTorch, and Jax
- 3x throughput for events on CPUs on average
- 5x throughput for events on GPUs on average
Please find the paper, presentation slideshow, and website along with steps to reproduce our findings below