INVITED I501: Streaming Signal Processing on GPUs

Wednesday plenary 1: invited talk

When

9 – 9:30 a.m., Nov. 8, 2023

Where

Theme: GPU implementations for core astronomical libraries

pretalx

In this presentation, I will discuss the latest developments in radio-astronomical signal processing on GPUs. I will present the Tensor-Core Correlator, a GPU library that combines antenna data at unprecedented speed and energy efficiency. The library is rapidly adopted by radio telescopes worldwide. We currently develop similar libraries for beam forming and filtering.

As I/O is our next bottleneck, we explore new methods (DPDK and RDMA) to stream digitized antenna data directly from the network into a GPU. I will show how the GPU handles 200 Gb/s Ethernet packets at line speed.

Finally, I will show how a proper codesign of the digitizer FPGA firmware, the network, and the GPU correlator leads to a highly cost- and energy-efficient instrument design.

Contacts

John Romein, ASTRON