Speaker
Description
Within the RADIOBLOCKS EU project, JIVE, in collaboration with partners, is investigating state-of the art accelerator technologies for implementing a (VLBI) correlator. In particular we are developing GPU kernels that implement the same algorithm as the SFXC CPU correlator. This development will provide re-usable VLBI delay and Windowed Overlap (WOLA) filter modules (which we call Radio Blocks) that can be combined with the Tensor Core Correlator Radio Block developed by our project partner ASTRON to implement a VLBI correlator as well as additional modules to implement coherent de-dispersion and multiple field center correlation. We also developed a wrapper library for UCX that allows efficient transfer of sampled baseband data directly into GPU memory. The result is a correlator that uses almost no CPU cycles. Initial power consumption estimates indicate that with these technologies we can build a correlator that is significantly more efficient than the existing EVN data processor at JIVE (SFXC). In this talk I will present the technologies that make this possible, discuss the current power consumption estimates and the scalability. I'll conclude with an outlook on what a future EVN correlator might look like.
