TL;DR
Sandia National Laboratories developed NeuroFEM, an algorithm that runs finite element method solutions of partial differential equations on spiking neuromorphic hardware and demonstrated it on Intel's Loihi 2 chips. Tests on a 32-chip Loihi 2 system showed near-ideal strong scaling and very high parallelizability, suggesting neuromorphic approaches may offer large energy-efficiency gains for some HPC workloads.
What happened
Researchers at Sandia National Laboratories created an algorithm called NeuroFEM that maps the finite element method (FEM) for solving partial differential equations (PDEs) onto spiking neuromorphic hardware. The team ran the method on Intel's Oheo Gulch system, which is built from 32 Loihi 2 neurochips, and reported near-ideal strong scaling: doubling cores roughly halved time-to-solution. The authors characterize their implementation as about 99% parallelizable in tests and present the PDE experiments largely as a proof of concept rather than a full replacement demonstration. Sandia's broader neuromorphic program includes systems from Intel, SpiNNaker and IBM; the lab cites efficiency figures such as roughly 15 TOPS per watt for Loihi 2 and higher claimed performance-per-watt for some SpiNNaker2 deployments. The paper appeared in Nature Machine Intelligence and the team notes programmability barriers for neuromorphic systems, arguing NeuroFEM reduces the additional work required to run numerical applications on spiking hardware. The researchers also note analog neuromorphic designs could further change performance and power trade-offs.
Why it matters
- PDEs underpin many computationally intensive scientific and engineering simulations; an efficient alternative could reshape HPC energy budgets.
- Neuromorphic hardware can deliver much higher performance per watt than conventional GPUs in Sandia's comparisons, suggesting potential operational cost reductions.
- Improved programmability (via algorithms like NeuroFEM) lowers a key barrier to broader adoption of neuromorphic architectures for numerical workloads.
- If scaling and robustness hold up at larger sizes, neuromorphic systems might complement or partially replace current supercomputer architectures for select problems.
Key facts
- Sandia developed NeuroFEM to implement the finite element method on spiking neuromorphic hardware.
- The team ran PDE experiments on Intel's Oheo Gulch system, which uses 32 Loihi 2 chips.
- Tests showed near-ideal strong scaling and the implementation was reported as about 99% parallelizable.
- Sandia cites roughly 15 TOPS per watt for Loihi 2 in its deployments, about 2.5x the efficiency of some contemporary GPUs cited by the lab.
- A SpiNNaker2-based system deployed at Sandia was claimed to deliver up to 18x higher performance-per-watt than modern GPUs, per the lab's reporting.
- Loihi 2 is a digital neuromorphic chip; researchers suggest moving to analog-based neuromorphic designs could yield further speed and energy gains.
- The study was published in the journal Nature Machine Intelligence.
- Sandia has deployed neuromorphic systems from Intel, SpiNNaker and IBM in recent years.
What to watch next
- Further demonstrations that scale NeuroFEM to larger, production-grade PDE problems — not confirmed in the source.
- Development and availability of analog neuromorphic hardware that could change power/performance trade-offs — not confirmed in the source.
- Direct head-to-head comparisons of neuromorphic systems and GPUs on real-world neural networks and large-scale PDE workloads; current parity is an open question.
Quick glossary
- Neuromorphic computing: A hardware approach that mimics neural structures and signaling (often spiking) to perform computation with the aim of improving efficiency for brain-like tasks.
- Partial differential equation (PDE): A mathematical equation involving multivariable functions and their partial derivatives, used to model physical phenomena such as fluid flow, heat transfer and electromagnetic fields.
- Finite element method (FEM): A numerical technique that breaks a complex domain into smaller elements to approximate solutions to PDEs, widely used in engineering and physics simulations.
- Strong scaling: A measure of how the time to solve a fixed-size problem decreases as more compute cores or nodes are added; ideal strong scaling halves runtime when core count doubles.
Reader FAQ
What did Sandia demonstrate?
They implemented NeuroFEM to run finite element PDE solutions on spiking neuromorphic hardware and tested it on a 32-chip Loihi 2 system, reporting near-ideal strong scaling.
Did the research use real hardware or only simulations?
The experiments ran on actual neuromorphic hardware (Intel's Oheo Gulch system with 32 Loihi 2 chips).
Are neuromorphic systems already better than GPUs for all HPC tasks?
Not confirmed in the source; the paper frames the PDE runs as a proof of concept and notes open questions about outperforming GPUs on certain workloads like deep neural networks.
Will analog neuromorphic chips replace digital designs soon?
Not confirmed in the source; the researchers suggest analog designs could offer further gains but do not provide a timeline.
HPC Artificial brains could point the way to ultra-efficient supercomputers Sandia National Labs cajole Intel's neurochips into solving partial differential equations Tobias Mann Fri 9 Jan 2026 // 23:42 UTC New research from Sandia…
Sources
- Artificial brains could point the way to ultra-efficient supercomputers
- Artificial brains could point way to ultra-efficient supers
- World's first neuromorphic supercomputer nears reality with …
- Nature-inspired computers are shockingly good at math
Related posts
- Revealed: AI Startup Hark Hired Apple’s iPhone Air Designer in January
- AI Solves Erdős Problem #728, Terence Tao Reports on Mathstodon
- Accenture invests in Profitmind to use AI for retail pricing and inventory