High-level code, wafer-scale performance
High-level code, wafer-scale performance
Cerebras introduced Tungsten, a language and compiler that programs its Wafer-Scale Engine at a high level while mapping code onto access pattern registers, event-driven tasks, fabric calls and loop-bearing instructions, and Aviary, a portfolio of high-performance applications and tutorials written with it. Aviary includes SNAP at 1.6 PFLOP/s on a single CS-4, HPCG at 1.04 PFLOP/s on a single system, and a wafer-scale molecular dynamics code reported as 748 times faster than Frontier. A researcher and a coding agent wrote a particle-in-cell plasma simulation in about two weeks that ran 60 million particles at 7,200 time steps per second on one system.
It documents a vendor's claim that a compiler can close the abstraction gap on wafer-scale hardware, with benchmark figures and an agent-written scientific code as evidence.
Source: Cerebras Blog · cerebras.ai