The Role of GPUs in Modern SDR
summary
Argues that the best SDR architecture pairs a small FPGA for I/O with a larger GPU for DSP and ML, demonstrates an end-to-end FM→demod→ASR pipeline running on a DeepWave AIR-T with an NVIDIA Jetson, uses NeMo/QuartzNet for ASR, and shows real-time operation with about 7% GPU core use and rescoring over 24 beams. Uses Python-first GPU tooling like cuSignal and zero-copy ideas (e.g., GPUDirect RDMA) to avoid host copies; demo works live but shows occasional transcription errors.
AI-written summary · about
chapters (9)
- 0:00Start
- 0:42GPUs in Software Defined Radio
- 2:08Example End-to-End Workflow for Useful RF/ML Problem
- 17:36Radio Automatic Speech Recognition (ASR)
- 18:26NVIDIA NeMo
- 19:19GNU Radio Implementation
- 21:00Radio-ASR
- 24:28Artificial Intelligence Radio Transceiver
- 25:22Intercept and Automatic Speech Recognition (ASR)
related talks
13mARK: GPU-driven Code Execution for Distributed Deep Learning
18mThe Next Step in Low Energy Edge DSP and AI
01h 08mGPU Data Science with Rapids
13mHardware/Software Co-Programmable Framework for Computational SSDs to Accelerate Deep...
34mGPU as a Service Over K8s: Drive Productivity and Increase Utilization
23mSparse Computing Drives Energy-Efficient Computation for Artificial Intelligence
41mLuxuries, necessities, and the challenges that remain
57mBuilding GPU-accelerated multi-agent apps