Inside a plain building located at Kirtland Air Force Base, situated in the high desert of New Mexico, liquid-cooled supercomputers operate quietly as they tackle some of the most challenging mathematical problems the U.S. government faces: modeling the movement of hypersonic nuclear weapons through the Earth's atmosphere, or determining the effects of a nuclear warhead exploding near another.
For over ten years, the chips performing this confidential and complex task were supplied by major semiconductor companies such as Nvidia (NVDA.O) or Advanced Micro Devices (AMD.O).
However, as these companies focus more on designing chips for artificial intelligence and encounter supply shortages, the system managers at Sandia National Laboratories—operating the machines at Kirtland and one of three U.S. laboratories responsible for developing and maintaining the country's nuclear weapons stockpile—are becoming increasingly uncertain about where they will secure the computing power needed for high-precision scientific research.
The current pressure we are experiencing is related to computing and also the supply chain," said Steve Monk, manager of Sandia's high-performance computing team, discussing the difficulty of obtaining chips that meet his requirements. "Looking ahead, it's somewhat stressful regarding our capacity to fulfill the mission.
New Companies Entering the Semiconductor Industry
The lab's situation highlights how the competition for superior AI chips is leading to an unexpected outcome: creating opportunities for smaller companies like NextSilicon, an Israeli startup, which is having its chips evaluated through a program at Sandia. It also emphasizes Sandia's role in fostering and influencing emerging computing technologies, as the organization has closely collaborated with Nvidia, especially during its rise in supercomputing, and continues to work with them on advanced memory solutions.
A significant issue for administrators at Sandia involves what is referred to as double-precision floating point calculation, a technical concept that enables the accurate computation of both extremely large and extremely small numbers without suffering from rounding errors. For many years, Nvidia and AMD focused on advancing this type of computing, securing supercomputing contracts with universities and government laboratories.
However, AI work does not gain the same advantages from double-precision computing as seen in physics simulations. Although AMD is launching a version of its chips designed for scientific computing, the double-precision capabilities of Nvidia's upcoming Rubin chips have dropped according to certain metrics, causing concern among many scientists in the high-performance computing sector, according to Ian Cutress, chief analyst at More Than Moore, a chip consulting company.
Daniel Ernst, senior director of supercomputing products at Nvidia, stated that the company continues to focus on scientific computing, striving to develop a versatile chip capable of handling real-world scientific tasks in addition to AI work.
However, the evolving chip market has led Sandia officials to evaluate products from new companies like NextSilicon, which employs a distinct computing method compared to graphics processing units (GPUs) or central processing units (CPUs) offered by Nvidia and AMD.
NUCLEAR SECURITY WORK
On Monday, Sandia, NextSilicon, and Penguin Solutions—the company that assisted in integrating NextSilicon's chips into a supercomputer—announced that the systems have achieved a significant technical milestone through a series of general supercomputing tests, positioning the chips as viable candidates for government applications.
This positions NextSilicon's chips for a decision this autumn regarding whether to begin evaluating the chips with more complex computing tasks that closely mirror the type of nuclear security work they will ultimately need to manage.
NextSilicon processors are capable of executing double-precision calculations and can also reconfigure themselves in real-time to operate more effectively. NextSilicon's chips conserve power through a data flow architecture, which minimizes the time and energy required to move data between the computing system's memory and processing units.
Sandia's collaboration with semiconductor companies frequently contributes to the broad adoption of technology. Liquid cooling for chips was once considered a niche concept when Sandia encouraged Intel, AMD, and Nvidia to focus on this technology over ten years ago, and it is now widely used.
James Laros, a senior scientist at Sandia responsible for evaluating new computing technologies, mentioned that collaborating with smaller companies such as NextSilicon is intended to guarantee that Sandia can consistently obtain the chips it requires, regardless of whether large chip manufacturers change their priorities.
"We must maintain available choices to finish our mission, as the mission is not a choice," Laros stated.
Provided by SyndiGate Media Inc.Syndigate.info).
No comments:
Post a Comment