SpaDA: A Spatial Dataflow Architecture Programming Language

TL;DR

SpaDA is a new programming language designed for spatial dataflow architectures like Cerebras, enabling efficient, concise, and high-performance programming of complex parallel operations on massive processing arrays.

Contribution

It introduces SpaDA, a high-level language and intermediate representation that simplifies programming for spatial dataflow architectures, with a compiler targeting Cerebras CSL.

Findings

Achieves over 260 TFlop/s on 730,000 PEs.

Allows expressing complex parallel patterns in 14.09x fewer lines than CSL.

Demonstrates effective compilation and optimization for Cerebras hardware.

Abstract

Spatial dataflow architectures like the Cerebras Wafer-Scale Engine deliver exceptional performance in AI and scientific computing by distributing scratchpad memory across hundreds of thousands of processing elements (PEs). Yet programming these architectures remains difficult: with no shared memory, data movement requires explicit configuration, and asynchronous task management introduces substantial complexity. We present SpaDA, a programming language that offers precise control over data placement, dataflow patterns, and asynchronous operations while abstracting low-level architectural details. We design and implement a compiler targeting Cerebras CSL through multi-level lowering and unique optimization passes. SpaDA functions as a high-level programming interface and an intermediate representation for domain-specific languages (DSLs), demonstrated here with the GT4Py stencil DSL. SpaDA enables concise expression of operations with complex parallel patterns -- including pipelined collective operations, multi-dimensional stencils, and dense linear algebra -- in 14.09x fewer lines than CSL, achieving over 260 TFlop/s across 730,000 PEs on a single device.