Novel Technology for Reducing Wavefront Image Processing Latency

Adaptive optics is essential for the successful operation of the future Extremely Large Telescopes (ELTs). At the heart of these AO system lies the real-time control which has become computationally challenging. A majority of the previous efforts has been aimed at reducing the wavefront reconstruction latency by using many-core hardware accelerators such as Xeon Phis and GPUs. These modern hardware solutions offer a large numbers of cores combined with high memory bandwidths but have restrictive input/output (I/O). The lack of efficient I/O capability makes the data handling very inefficient and adds both to the overall latency and jitter. For example a single wavefront sensor for an ELT scale adaptive optics system can produce hundreds of millions of pixels per second that need to be processed. Passing all this data through a CPU and into GPUs or Xeon Phis, even by reducing memory copies by using systems such as GPUDirect, is highly inefficient. The Mellanox TILE series is a novel technology offering a high number of cores and multiple 10 Gbps Ethernet ports. We present results of the TILE-Gx36 as a front-end wavefront sensor processing unit. In doing so we are able to greatly reduce the amount of data needed to be transferred to the wavefront reconstruction hardware. We show that the performance of the Mellanox TILE-GX36 is in-line with typical requirements, in terms of mean calculation time and acceptable jitter, for E-ELT first-light instruments and that the Mellanox TILE series is a serious contender for all E-ELT instruments.