https://hgpu.org/?p=10462
D5.5.2 - Architectural Techniques to exploit SLACK & ACCURACY trade-offs