Register for free to
view the full innovation profile.
Posits are a tapered precision number system to replace IEEE floating point. It provides more precision, lower complexity, and lower power implementations than IEEE floating point.
About
Posits are a tapered floating point format that is more precise than IEEE floating point. This enables posits to use fewer bits than IEEE floating point which mitigates the memory bandwidth bottleneck of DL and HPC algorithms. Experiments demonstrate posits having a 400% memory bandwidth benefit in Deep Learning applications.
Key Benefits
Deep Learning algorithms tend to be memory bandwidth limited. Posits cut down on memory bandwidth as compared to IEEE floating point, and thus posit-based DL systems demonstrate significantly higher performance for both training and inference applications.
Applications
AI, robotics, AV, control systems, computational science, big data, statistics