https://hgpu.org/?p=11150
BIDMach: Large-scale Learning with Zero Memory Allocation