Programming Maximum Performance Computing Systems
Programming MPC Systems
White Paper — Maxeler Technologies
Maximum Performance Computing
Maximum Performance Computing (MPC) changes the classical computer science optimization from ease-of-programming to maximizing performance and minimizing total cost of computing. We maximize performance by constructing compute engines to generate one result per clock cycle, wherever possible. Ease-of-programming is still important but takes second place to performance, computational density and power consumption.
As such MPC focuses on mission critical, long running computations with large datasets and complex numerical and intense statistical content. At Maxeler we drive MPC via ‘Multiscale Dataflow Computing’. This white paper describes the components of MPC and illustrates how we program MPC dataflow computers.
One Maxeler Dataflow Engine (DFE) combines 104 arithmetic units with 107 bytes of local fast SRAM (FMEM) and 1011 bytes of 6-channel large DRAM (LMEM). MaxelerOS allows the DFEs and CPU to run in parallel, so while the DFEs are pro- cessing the data, the CPU performs the non-time- critical parts of an application.
Our MPC programming environment comprises of MaxCompiler, MaxelerOS (running within Linux and DFEs themselves), MaxIDE, our fast DFE software simulator, and a comprehensive debug environment.
DOWNLOAD WHITE PAPER (.PDF):MPCwhitepaper (3) (1)
Sponsored by Maxeler Technologies.