Array Processing on Steroids for the SKA Radio-Telescope
summary
Covers the SKA radio telescope’s data-processing needs, quantifying a requirement of ~200 PB/s memory bandwidth for converting antenna signals to images and detecting transients. Describes the end-to-end processing pipeline and the algorithm/system design process used to meet that scale, highlights CS-related challenges in array processing, and surveys recent technologies — including memory advances, ML accelerators, and new floating-point formats — that could help.
AI-written summary · about
chapters (32)
- 0:00Intro
- 2:08Grand Challenges in HPC
- 3:35SKA telescope
- 7:18Science Headlines
- 8:59Epoch of Re-lonisation
- 10:11Pulsar Timing Array
- 11:11SKA dataflow schematic
- 13:22Requirements & Tradeoffs
- 15:56Imaging Problem
- 17:21Standard interferometer
- 19:59Computing in radio astronomy - 101
- 23:41Interferometry radio telescope
- 24:33Outline of imaging algorithm
- 28:04Data in the computation
- 28:50SDP "performance engineering" approach
- 31:07Relative kernel cost (from parametric model)
- 31:45Supercomputer parameters
- 34:27This lecture could now go in multiple directions
- 35:36Conclusions.
- 39:34SKA to enable non-experts
- 41:17Example of near disasters ...
- 42:15Success stories in architecture specification
- 42:48LLNL Proxy App - success story
- 43:55Going forward
- 45:31SKA pipeline study
- 46:30Developments in floating point
- 48:21Domain Specific Languages
- 49:36TensorFlow and MLIR
- 50:49Combining multiple libraries & DSLS
- 52:43Scheduling work
- 53:50Double buffer pitfall
- 55:01Cache hierarchies - stitching and decomposing
related talks
10mPaco Nathan (Liber 118) Interview - OSCON 2014
20mAI in the physical world: from observation to discovery
51mWhen the Universe Calls
20mMining Large Data Sets of Biomedical Omics Da...
36mAn almost perfect match: cloud native and scientific computing
13mBuilding High-Performance AI Literature Processing
46mPath of Data & AI Challenges
28mAre We Alone? How GNU Radio Can Help Us Find ET