As part of the EU-funded RADIOBLOCKS project, we have created a set of highly optimized GPU libraries which can serve as building blocks for a correlator or beamformer application. We have integrated a number of these radioblocks into a fully functional VLBI correlator, which is intended to become the next production correlator of the European VLBI Network (EVN). In this talk I will present this correlator and its constituent radioblocks.

Since 2012, the operational correlator of the EVN is a distributed CPU based correlator which is deployed locally on an HPC cluster. However, the advent of next generation wideband receiver systems is expected to increase the maximum observed bandwidth in EVN observations by a factor four to eight, which will require an increase in correlator capacity by a similar amount. Expanding the current HPC cluster to meet these new requirements would be prohibitively expensive while also using a significant amount of power. However, as I will show in this talk, these requirements can be met cost-effectively using GPU acceleration, while drawing significantly less power. In the talk I will present detailed measurements of the power consumption of our GPU implementation.

The GPU correlator is not a monolithic application; it makes use of a collection of GPU libraries that we have created, which we call radioblocks. Each of these radioblocks is distributed as a stand-alone library, allowing other projects to easily incorporate the relevant radioblocks into their own code base.
For example, a core component of a VLBI correlator is the geometric delay compensation for which we have created an optimized VLBI capable radioblock. The (cross-)correlations are performed using a radioblock called the tensor core library (TCC), which uses tensor core instructions to compute the correlation function. It achieves an order-of-magnitude performance improvement over non-tensor-core GPU implementations.

Furthermore, we have implemented a number of advanced features as radioblocks. We have created a radioblock which can perform coherent de-dispersion on baseband data, this feature is crucial to process certain classes of radio transients such as Fast Radio Bursts (FRBs), and some pulsars. Another example is a radioblock which can produce multiple simultaneous phase centres, a technique to efficiently perform wide-field VLBI. It does this by producing individual narrow-field data sets for each source in the field of view, rather than producing a single monolithic wide-field dataset that encompasses the entire field of view.

I will conclude the talk by discussing a potential future application. The Space Array initiative, of which we are members, is preparing a proposal to build a space based VLBI array. Due to bandwidth limitations, the correlation will be performed onboard of the spacecrafts. I will show how our work could be applied to this space-based correlator.