High-speed, multi-input, multi-output control using GPU processing in the HBT-EP tokamak

► We present a GPU based system for magnetic control of perturbed equilibria. ► Cycle times are below 5μs and I/O latencies below 10μs for 96 inputs and 64 outputs. ► A new architecture removes host RAM and CPU from the control cycle. ► GPU and DA/AD modules operate independently and communicate via PCIe peer-to-peer connections. ► The Linux host system does not require real-time extensions. We report on the design of a new plasma control system for the HBT-EP tokamak that utilizes a graphical processing unit (GPU) to magnetically control the 3D perturbed equilibrium state [1] of the plasma. The control system achieves cycle times of 5μs and I/O latencies below 10μs for up to 96 inputs and 64 outputs. The number of state variables is in the same order. To handle the resulting computational complexity under the given time constraints, the control algorithms are designed for massively parallel processing. The necessary hardware resources are provided by an NVIDIA Tesla M2050 GPU, offering a total of 448 computing cores running at 1.3GHz each. A new control architecture allows control input from magnetic diagnostics to be pushed directly into GPU memory by a D-TACQ ACQ196 digitizer, and control output to be pulled directly from GPU memory by two D-TACQ AO32 analog output modules. By using peer-to-peer PCI express connections, this technique completely eliminates the use of host RAM and central processing unit (CPU) from the control cycle, permitting single-digit microsecond latencies on a standard Linux host system without any real-time extensions.