Massimult: A Novel Parallel CPU Architecture Based on Combinator Reduction
Jurgen Nicklisch-Franken, Ruslan Feizerakhmanov
TL;DR
Massimult proposes a fundamentally different CPU paradigm based on combinator reduction to achieve deep parallelism and locality, contrasting with traditional Von Neumann processors. The approach comprises a typed machine language (LambdaM) and a minimal KVY combinator base, compiled into a LambdaBase intermediate form and executed on the Matrima machine designed for massively parallel evaluation. The paper details the LambdaM/LambdaBase pipeline, Scott-encoded data types, and a concrete hardware-aware architecture featuring the CellPool, Checker, Reducer, Recycler, and Supervisor to manage parallel reductions and memory. Collectively, these components aim to deliver faster computation with lower energy consumption and better scalability, with a clear roadmap toward GPU/FPGA and eventual silicon implementations that could disrupt conventional CPU design by exploiting inherent functional-parallelism and locality.
Abstract
The Massimult project aims to design and implement an innovative CPU architecture based on combinator reduction with a novel combinator base and a new abstract machine. The evaluation of programs within this architecture is inherently highly parallel and localized, allowing for faster computation, reduced energy consumption, improved scalability, enhanced reliability, and increased resistance to attacks. In this paper, we introduce the machine language LambdaM, detail its compilation into KVY assembler code, and describe the abstract machine Matrima. The best part of Matrima is its ability to exploit inherent parallelism and locality in combinator reduction, leading to significantly faster computations with lower energy consumption, scalability across multiple processors, and enhanced security against various types of attacks. Matrima can be simulated as a software virtual machine and is intended for future hardware implementation.