Tunable Floating-Point

Tunable Floating-Point (TFP) is a floating-point format with adjustable significand and exponent fields bit-width

Features

• significand m=[4, 24] bits (including hidden bit)
• exponent e=[5,8] bits
• rounding modes
  • RTZ Round toward zero (truncation)
  • RTN Round to the nearest
  • RTNE Round to the nearest (tie to even)
    IEEE 754 roundTiesToEven mode
• subnormals flushed to zero

TFP includes binary32 (m=24, e=8), binary16 (m=11, e=5), Google's Brain-FP (m=8, e=8).

Motivation

• A flexible unit, handling a flexible format, can increase the power efficiency
  

  Power efficiency = throughput watt

• Operations can be approximated by reducing the precision
• Accuracy of reduced precision can be improved by rounding

Recent publications

• M. Franceschi, A. Nannarelli and M. Valle, "Tunable Floating-Point for Artificial Neural Networks", to appear in Proc. of 25th IEEE International Conference on Electronics Circuits and Systems (ICECS 2018), Bordeaux, France. Dec. 2018.

• M. Franceschi, A. Nannarelli and M. Valle, "Tunable Floating-Point for Embedded Machine Learning Algorithms Implementation", Proc. of 15th International Conference on Synthesis, Modeling, Analysis and Simulation Methods and Applications to Circuit Design (SMACD 2018), p. 89-92. Prague, Czech Republic. 2-5 July 2018.

• A. Nannarelli. "Tunable Floating-Point for Energy Efficient Accelerators", Proc. of 25th IEEE Symposium on Computer Arithmetic, p. 33-40, Amherst, USA. 25-27 June 2018.

Modified by Alberto Nannarelli on Friday October 12, 2018 at 17:17