Speaker
Description
Satellite based spectroscopic observations produce massive volumes of data at very high velocities. Analysis of such observation via deep learning algorithms, e.g. for the precise estimation of the redshift associated with individual galaxies, requires massive floating point operations, generally performed on Graphics Processor Units (GPU). This form of processing is fast but requires substantial energy for the computation, and thus necessitates the transmission of the acquired measurements to ground stations for processing in large data centers. Such data transmissions are increasingly becoming a bottleneck, as transmission speed improvements do not keep up with the pace of on board data generation. An alternative technology to GPUs is Field Programmable Gate Arrays (FPGAs), which often require substantially less energy per computation compared to GPUs, but are considered too slow for deep learning based inference. In this work, through the collaboration of two EU Horizon 2020 projects, namely EuroExa and DEDALE, we demonstrate experimentally that using (i) appropriate data structures to reduce memory bandwidth, (ii) compressed fixed point indices to clustered floating point weights and (iii) massive pipelining, FPGA-based computing can yield extremely high (in the order of 99%) classification accuracy vs. GPUS in the context of top-one classification, at an order-of-magnitude less energy. For this work we considered optimized Tensorflow codes running on various GPUs vs. our proposed FPGA-based architecture for galaxy redshift estimation from extended wavelength range spectroscopic measurements using simulated measurements which are in-line will be publicly available specification of the upcoming ESA Euclid deep space mission. We show on actual runs in hardware that the EuroExa-developed Quad FPGA Daughter Board (QFDB) offers substantially lower latency vs. a similar-technology Nvidia P1000 GPU for batches of any size, it offers better throughput for batches up to 30 images (which can scale out to any batch size), and it offers roughly an order of magnitude better energy consumption for the same computations, thus becoming an interesting technology for on-satellite deep learning classification. An important aspect of this work is that the FPGA model used in this work has an equivalent rad-hard part qualified for space applications.
| Paper submission | Yes |
|---|
