Los Alamos National Laboratory (LANL) will receive four Maverick-2 accelerators from NextSilicon, a company developing chips that can change how they use their computing resources.
Los Alamos National Laboratory (LANL) is receiving four Maverick-2 accelerators from NextSilicon, a company developing chips that can change how they use their computing resources while a program is running.
The goal is to make it easier to run demanding scientific codes without spending so much time adapting them to the hardware. The four accelerators are now installed in the lab's Darwin testbed, where researchers will test how well the technology handles their workloads.
Getting the best performance out of new hardware is not always easy, especially in high-performance computing (HPC). Scientists can spend considerable time tweaking their codes to make better use of a processor. This is even harder with large simulations, where moving data around can be as much of a challenge as the calculations themselves. NextSilicon hopes Maverick-2 can take some of that work off their hands.
"For decades, scientists have reshaped their codes to fit the hardware. Our pioneering technology in the form of the Maverick-2 accelerator reverses that completely. We are honored to work with Los Alamos to realize the benefits of hardware that adapts to the science," said Elad Raz, founder and CEO of NextSilicon.
Maverick-2 is designed to work alongside a CPU and handle the more demanding parts of an application. The technology behind it is NextSilicon's Intelligent Compute Architecture (ICA), which allows the hardware to adjust how it uses its computing resources as the workload changes. Specifically, the design allows Maverick-2 to rearrange its resources depending on what the code needs at a given point. NextSilicon says this can help extract more performance from existing applications without rewriting them.
Not every HPC workload is easy to accelerate, however. Some involve many calculations that can run at the same time, which GPUs are particularly good at handling. Others are much messier. They may need to pull data from different places in memory or keep switching between different types of calculations, making it harder to keep a processor busy and deliver the performance researchers need.
This becomes a bigger problem with multi-physics simulations, which must account for several physical processes at once, and whose computing demands are not constant throughout the simulation. Getting these codes to run well can be a real headache, even with the computing power available on today's HPC systems. Los Alamos has spent years studying these issues, which gives the lab good reason to be interested in hardware that works somewhat differently.
The plan for now is to put Maverick-2 to work on some of the multi-physics codes Los Alamos uses for national security research. The four accelerators are part of Darwin, where the lab tests new hardware before deciding what might be worth pursuing further. The arrangement also gives researchers a chance to work with NextSilicon and identify where the hardware struggles and what can be done about it.
"Working directly with technology developers is how we evaluate and affect new computing architectures for our mission. Maverick-2 gives our teams hardware to test with," said Gary Grider, senior director for Computing Technologies at Los Alamos National Laboratory.
"Maverick-2 can be reconfigured for different application requirements, with tools available to examine how codes map onto the system," said Galen Shipman, chief architect for Advanced Technology Systems at Los Alamos. "That information can inform both performance analysis and application co-design."
NextSilicon's software tools will also be tested. They let researchers see how their codes run on Maverick-2, which may reveal where performance slows down and what could be improved.
When Maverick-2 was introduced, NextSilicon claimed it could match leading GPUs on the HPCG benchmark while using about half the power. These were the company's own results and have not been independently verified. The testing at Los Alamos offers an opportunity to see how the technology holds up.