Novel Technologies for Radio-Astronomical and Medical Signal and Image Processing

Novel Technologies for RadioAstronomical and Medical Signal and Image Processing John W. Romein From Cells to Galaxies September 21st-23rd, 2022 2 Outline •joint Dutch radio-astronomical & medical imaging research •tensor-core correlator 3 Roles of the parties •ASTRON: –radio astronomy •Erasmus MC: –functional ultrasound •NLeSC: funding agency + “eScience Research Engineers” 4 Joint Dutch RA/MI research •RECRUIT (Reducing Energy Consumption in Radio-astronomical and Ultrasound Imaging Tools) –explore common interests radio astronomy & medical imaging –CfP: energy-efficient computing –granted by NLeSC 5 Areas of common interests 3) (imaging) algorithms 2) common GPU/FPGA libraries 1) GPU/FPGA technology exploration 0) tools 6 1) Technology exploration – ACAP •data acquisition –Xilinx ACAP (= CPU + FPGA + DSP + AI engines + NoC) vector processors → signal processing 7 1) Technology exploration – tensor cores •hardware matrix multiplication units –limited precision input data –much faster than regular GPU cores •accelerates deep learning •NVIDIA GPUs since 2017 •see slides 12–19 8 2) Common GPU & FPGA libraries •beam forming •correlations •FFT generator –FFT invoked by GPU kernel –superseded by cuFFTDx •1-bit matrix multiplications correct phase/amplcuFFTFIR filter FIR filter + in-place FFT + correct phase/ampl 9 3) Algorithm development •common imaging techniques •??? 16 Implementation challenges 1) complex numbers 2) triangular output format 3) fast data fetching •all hidden from the user not supported by tensor cores 17 •NVIDIA A100 •sustained performance (no short-time turbos) •PCIe transfers not measured •benchmarked against xGPU Performance measurements 18 Tensor-Core Correlator: performance NVIDIA A100 19 Tensor-Core Correlator: energy efficiency NVIDIA A100 20 Next challenges •data: FPGA → GPU using 400 GbE –essential for wide-band instruments –requires new techniques –RDMA (RoCE v2) –user-space NIC access (DPDK) •next-gen tensor cores –DMA 21 Conclusions •RECRUIT –Dutch radio-astronical/medical imaging research project –explore common interests (techology, software, algorithms, energy efficiency) •tensor-core correlator: –GPU library –disruptive technology: 5–10x performance, energy efficiency 22 Acknowledgements •This work was funded by –Netherlands eScience Center (RECRUIT) –NWO Netherlands Foundation for Scientific Research (DAS-6) •thanks to NVIDIA for providing an A100 23 Backup slides