Welcome to WFA’s Documentation!

This page contains the documentation for the Wafer scale engine Field equation Application programming interface or WFA.

This package is a product of the National Energy Technology Laboratory (NETL) in collaboration with Cerebras Systems Inc.

Field equations have two properties:

  1. They have a value or set of values at every point in space time

  2. Changes to the values are only affected by the immediate surroundings (the Principle of Locality in physics)

This gives rise to several sets of common differential equations that describe almost every physical phenomenon in nature at the finest space-time scales. The only known exception to this fundamental principle appears to be quantum entanglement which seems to have the property of instantaneous information transfer. Every other phenomenon propagates through space-time with a transfer rate lower than or equal to the speed of light. Field equations are even fundamental to our understanding of the quantum realm at Plank scales. Field equations can be used to describe systems of objects which exhibit discrete behavior at intermediate scales. Due to the law of averages, ensembles of discrete objects can take on the character of a field equation at scales that are many times larger than the discrete objects (although even “discrete” objects are described as fields on the finest scales). This allows sets of field equations to adequately and accurately describe the behavior of fluids, star/galaxy clusters, and granular solids under limited circumstances. If a physical system can be written with ODE/PDE expressions at the scales of interest, it is possible to approximate them on a computer through discretization.

It should come as no surprise that an Application Specific Integrated Circuit (ASIC) can be many orders of magnitude faster and more efficient than general purpose computing hardware. To the best of our knowledge, no one has endeavored to create an ASIC for field equation modeling. However, recent developments in Artificial Intelligence (AI) ASICS are producing hardware that happens to also be very good at field equation modeling, as is the case with the Cerebras Wafer Scale Engine. The architecture allows definition of discretized values, distributed evenly across the processor array and mirrors the physical Principle of Locality in that it allows for the nearest neighbors to affect a local change in a minimum of clock cycles (on average 1-2 cycles). While the WSE architecture was not developed as an ASIC for field equations, the same properties that make it so efficient for training large AI models (ie: data flow based matrix multiplications) end up making it efficient for solving field equations with large numbers of voxels or cells.

Recent Achievement

NETL, Cerebras, and the Pittsburgh Supercomputing Center recently completed the first ever CFD simulation on a WSE at the Neocortex facility. The simulation is of Rayleigh-Benard convection using a new solution method called The Accellerated Pseudo Transient Method. The CFD solver uses the WFA to compile and run. Work is currently under way to characterize the exact speed gains and energy consumption relative to the fastest conventional hardware available. Our early benchmarks and results are available in an arXiv paper. and will be updated as new data comes in.