Likely a machine with a group of Xeon Phi coprocessors, each a PCI card containing 64 fully-capable x86 cores: https://www-ssl.intel.com/content/www/us/en/processors/xeon/...

The idea is that you can run embarrassingly parallel workflows without needing to recode for a GPU; just launch processes with appropriate affinities. NERSC has compute clusters with thousands of these beasts. They make astronomers very happy.