Computer manufacturer Cerebras used its AI computer on a non-AI problem: simulating “buoyancy-driven Navier-Stokes flows” that capture the dynamics of many systems in nature and the built environment. The work, the first of its kind, allows a “digital twin” of the real world, allowing scientists to make predictions and see the effects of interventions in a kind of closed loop.
Cerebras/DoE NETL 2023
Real-time simulation of the real world can offer scientists a way to make predictions based on scenarios as they unfold. This can be advantageous when dealing with extreme weather scenarios such as those associated with global warming.
AI computing pioneer Cerebras and the US Department of Energy’s National Energy Technology Laboratory on Tuesday announced an acceleration of scientific equations that they say could enable real-time simulation of ‘extreme weather conditions.
“This is a real-time simulation of how liquids of different volumes behave in a dynamic environment,” said Andrew Feldman, CEO of Cerebras.
“In real time or faster, you can predict the future,” said Feldman. “From the starting point, the actual phenomenon unfolds more slowly than your simulation, and you can go back and make adjustments.”
This type of simulation, a digital twin of real-world conditions, essentially allows for a “closed loop” that allows you to manipulate reality, Feldman said.
In prepared remarks, Dr. Brian J. Anderson, Laboratory Manager at NETL: “We are excited about the potential of this real-time simulation of natural convection as we can really speed up and improve dramatic the design process for some large projects that are essential for the mitigation of climate change and the possibility of a secure energy future of are critical – projects such as carbon capture and the production of blue hydrogen.”
Anderson added, “On a traditional supercomputer, this workload is hundreds of times slower, eliminating the possibility of real-time rates or extremely high resolution streams.”
Also: “We can solve this problem in a time that no number of GPUs or CPUs can match,” says startup Cerebras at the supercomputing conference
In a video created by the researchers, streams of hot and cold liquids flow up and down like an alien landscape.
Cerebras has made a name for itself with exotic hardware and software for large artificial intelligence training programs. However, the company has expanded its repertoire by focusing on challenging fundamental research problems that are computationally intensive and potentially unrelated to AI.
In the field of Computational Fluid Dynamics, the Cerebras machine, called CS-2, is able to simulate what is known as Rayleigh-Bénard convection, a phenomenon that occurs when a liquid is heated from below and cooled from above.
The work was made possible by running a new software package developed by Cerebras and NETL last fall called the WSE Field Equations API, a Python-based front end that describes field equations. Field equations are a type of differential equations that “describe nearly every physical phenomenon in nature at the finest space-time scales,” according to the GitHub documentation.
Basically, the field equations model everything in the known universe except quantum entanglement.
The API, described in the November article “Disruptive Changes in Field Equation Modeling: A Simple Interface for Wafer-Scale Machines” published on arXiv, was explicitly designed to take advantage of the chip Special AI computer Cerebras. Called Wafer Scale Engine or WSE, the chip was unveiled in 2019 and is the largest computer chip in the world, almost the size of an entire semiconductor wafer.
The November paper described WSE as being able to perform field equations two orders of magnitude faster than NETL’s Joule 2.0 supercomputer, built by Hewlett Packard Enterprise with 86,400 Intel Xeon processor cores and 200 GPU chip from Nvidia.
Because of their highly distributed nature, supercomputers are valued for their ability to execute components of equations simultaneously to speed up overall computation time. However, NETL scientists found that the operating field equations encountered bandwidth and latency limitations when moving data from off-chip memory to the processor and GPU cores.
Also: AI startup Cerebras has celebrated chip triumphs where others have tried and failed
The field equations API instead leveraged the large memory capacity on Cerebra’s WSE chip. WSE 2, the second version of the chip, packs 40GB of on-chip memory, a thousand times more than Nvidia’s A100 GPU chip, Nvidia’s current mainstream offering.
As the authors of NETL and Cerebras describe the issue,
While GPU bandwidth is high, latency is also high. Little’s Law dictates that when both latency and bandwidth are high, a large amount of data must be in transit to keep utilization high. Maintaining significant amounts of data in transit results in large subdomain sizes. These single device scaling features limit the iteration rates achievable on GPUs. On the other hand, the WSE has L1 cache bandwidths and a latency of one cycle, so the iteration rates that can be reached on each processor are much higher.
The simulation runs on an Excel spreadsheet type with more than 2 million cells with changing values.
While research in November found the CS-2 to be significantly faster than the Joule in the field equations, scientists at NETL have yet to report official speed comparisons for the work of the fluid dynamics announced Tuesday. That work is currently being done on a group of GPUs for comparison, Feldman said.
In the press release, NETL and Cerebras said, “The simulation is expected to run hundreds of times faster than traditional distributed computing, as previously demonstrated with similar workloads.”
Also: AI startup Cerebras has celebrated chip triumphs where others have tried and failed
The Cerebras CS-2 machine used in the project is installed at Carnegie Mellon University’s Pittsburgh Supercomputing Center as part of the Neocortex system, a “high-performance artificial intelligence (AI) computer” funded by -National Science Foundation.
The current work is not the first time that Cerebras has emerged from AI. In 2020, the WSE chip excelled in another partnership with the DoE on another set of partial differential equation problems in fluid dynamics.
Feldman, CEO of Cerebras, indicated that there are many more scientific computing opportunities coming for the company.
The simulation of buoyancy-driven Navier-Stokes flows, Feldman noted, uses “fundamental” equations in computational fluid dynamics.
“The fact that we can break it down in this simulation bodes very well for us in a variety of applications.”
Source: www.zdnet.com
