Tri-Labs (comprised of three main US analysis establishments – the Lawrence Livermore Nationwide Laboratory (LLNL), Sandia Nationwide Laboratories (SNL), and Los Alamos Nationwide Laboratory (LANL)) has been working with AI agency Cerebras on quite a lot of scientific issues, together with breaking the molecular dynamics (MD) timescale barrier.
There’s a paper explaining this explicit problem, which you’ll be able to learn right here, however basically it refers back to the drawback of conducting molecular dynamics simulations on a bigger timescale than would usually be doable.
The limitations listed here are twofold: computational energy and communication latency between completely different nodes of an HPC system. Historically, to compensate for the dearth of computational energy, scientists assign extra work to every node and scale up the simulation measurement with the node depend. Sadly, the gradual inter-node communication attributable to excessive latency additional exacerbates the timescale drawback.
Like a donut
MD simulations are essential to a number of scientific fields as they bridge the hole between quantum digital strategies and continuum mechanics strategies. Nevertheless, these simulations encounter timescale limitations, as they should account for atomic vibrations, which happen over very brief timescales, and different phenomena that happen over for much longer intervals.
The authors of the paper sought to beat the timescale barrier by using a extra environment friendly computational system, particularly Cerebras’ Wafer-Scale Engine.
As The Subsequent Platform explains, “The precise simulation was to beam radiation into three completely different crystal lattices made from tungsten, copper, and tantalum. In these explicit simulations, which have been for 801,792 atoms in every lattice, the thought is to bombard the lattices with radiation and see what occurs.”
Operating the simulations on Frontier, the world’s quickest supercomputer based mostly on the Oak Ridge Nationwide Laboratory in Tennessee, and on Quartz at LLNL, scientists have been solely capable of witness nanoseconds of what was occurring to the lattices as they have been bombarded with radiation. Utilizing WSE, they got tens of milliseconds of time to look at what occurred.
For the assessments, Tri-Labs used Cerebras Wafer Scale Engine 2 (WSE-2), quite than the newer, and extra highly effective WSE-3 launched earlier this yr, however as detailed above the outcomes have been spectacular. Because the paper reviews, “By dedicating a processor core for every simulated atom, we display a 179-fold enchancment in timesteps per second versus the Frontier GPU-based Exascale platform, together with a big enchancment in timesteps per unit vitality. Lowering yearly of runtime to 2 days unlocks presently inaccessible timescales of gradual microstructure transformation processes which can be vital for understanding materials habits and performance.”
The Subsequent Platform’s Timothy Prickett Morgan requested Cerebras CEO and co-founder, Andrew Feldman, what occurs while you join a number of wafer scale engines collectively and attempt to run the identical simulation and was informed “nobody is aware of but”.
Prickett Morgan went on to notice, “The proprietary interconnect within the WSE-2 methods may scale to 192 units, and with the WSE-3, that quantity was boosted by greater than an order of magnitude to 2,048 units,” however he “strongly suspects that the identical scaling rules apply to WSEs as apply to GPUs and CPUs.”
He went onto counsel, nonetheless, that there could possibly be some option to lash WSEs collectively bodily, and make a “stovepipe of squares of interconnected WSEs,” probably making a donut design with energy working on the within and cooling on the surface. As Prickett Morgan concludes, “This type of configuration couldn’t be worse than utilizing InfiniBand or Ethernet to interlink CPUs or GPUs.”