World's first native multidimensional FFT technology
- Transpose hence node-to-node communication bottleneck has been removed algorithmically (not just hidden in the pipeline)
- Highly parallelizable (see animations below)
For example, an N x N x N 3-D FFT can be parallelized into (N^3)/8 threads at the top level (e.g., N=1024 gives 1/8 billion top level threads!)
Patent pending
An NxN 2-D FFT example data flow : Old vs. New (N=8)
OLD: max top level parallelism = N
NEW: max top level parallelism = (N^2)/4