Answers are grounded in this paper and cite their sources.
This paper presents a posit processing element (PPE) that integrates optimized posit adder, multiplier and Multiply Accumulate (MAC) units suitable for matrix workload aimed at ML and Deep Neural Network accelerators in the signal processing domain. The proposed architecture targets Posit (N, es) formats, with a component-partitioning stage having a leading zero counter–based regime extractor, followed by structurally modelled arithmetic cores built from Kogge–Stone, Han–Carlson adders, Modified Booth and Wallace tree multipliers. Parameterized Verilog HDL generators are developed and instantiated for 8‑, 16‑, and 32‑bit posits, with synthesis performed on a Zynq‑7000 and 90 nm ASIC flow to enable a fair comparison with existing posit cores. For a <32,3> posit adder on FPGA, the proposed Han–Carlson–based design reduces the maximum datapath delay by 1.5%, and on ASIC it reduces cell area by around 32%. The Modified Booth–based 32-bit posit multiplier achieves up to 63% lower ASIC delay but with 48% higher cell area, highlighting the speed–area trade‑off. The 32‑bit PPE and a 4‑tap FIR filter realized with the proposed posit modules achieves 23.1% lower critical-path delay. The hardware overhead in posits are amortized by the better accuracy. Evaluating rounding schemes reveals that Posit {32, 2} under truncation retains 100% accuracy up to 4 decimal places and 98.8% at 6 decimal places with a negligible Mean Relative Error of 1.0e−6, proving that rounding logic can be safely bypassed to save hardware cost in noise-tolerant DSP and ML applications. The target domain applications are MAC-dominated DSP/ML accelerator datapath involving FIR/convolution filtering and matrix–vector multiplication.
No takes yet. Share an insight, caveat, or question.
Susithra et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: