Sep 5 – 9, 2023
ASU Memorial Union (2nd Floor)
US/Arizona timezone
GRCon23 will be running from Tuesday Sept 5 to Saturday Sept 9 this year.

BLADE: The Allen Telescope Array CUDA-accelerated Real-Time DSP Library

Sep 6, 2023, 1:00 PM
30m
Arizona (MU 221) (ASU Memorial Union 2nd Floor)

Arizona (MU 221)

ASU Memorial Union 2nd Floor

Talk Radio Astronomy Main Track

Speaker

Luigi Cruz (SETI Institute)

Description

The Allen Telescope Array (ATA) is a radio interferometer currently composed of 42 antennas. The array is made up of 6.1-meter diameter offset Gregorian telescope elements distributed randomly with a maximum baseline of 300 meters. The recently upgraded cryo-cooled log-periodic antenna feed (known as Antonio Feed) is sensitive to a wide and continuous range of frequencies ranging from 900 MHz to 12 GHz.

BLADE (Breakthrough Listen Accelerated DSP Engine) is a C++20 GPU-based computer software developed in-house to process data produced by the array. It is being used in production at the Allen Telescope Array to combine signals received by individual antennas to synthesize the aperture of a large antenna, a technique known as “beamforming”. Moreover, BLADE is also capable of post-channelize the beam-formed data into high-resolution (<1.0 Hz/bin) spectrogram in real-time. Currently, a twenty-antenna multi-beam observation routine produces a 60 GHz of complex 8-bit integer stream of aggregated data that is processed (beamformed and channelized) in real-time.

In this talk, I'm going to discuss how a software-defined telescope with a GPU-enabled processing backend can enable new scientific capabilities. As well as how to overcome hardware bottlenecks with software and important technical considerations involved in building a distributed pipeline capable of processing a stream of ~1 Tbps of data faster than real-time. The focus will be given on how the modular interface containing the Digital Signal Processing code implements optimization techniques in the background without developer input. Examples of techniques automatically applied to the module without added complexity are CUDA Graphs, a just-in-time compilation of CUDA kernels, smart parallel batch processing, and smart heterogeneous memory management.

Talk Length 30 Minutes
Link to Open Source Code https://github.com/luigifcruz/blade

Primary authors

Luigi Cruz (SETI Institute) Dr Wael Farah

Presentation materials