Neuromorphic Computing Crosses the Production Threshold as Sandia Solves Physics on Brain-Inspired Hardware
Sandia demonstrates that spiking neural networks can solve physics simulations on Intel Loihi 2, Intel ships the 4nm Loihi 3 with eight million neurons, and UT San Antonio opens the first public neuromorphic computing hub in the United States.
Overview
Neuromorphic computing, the discipline of building processors that mimic the electrochemical signaling of biological neurons, has spent more than a decade as a promising but marginal branch of computer architecture. In early 2026, three developments suggest the field is crossing from laboratory curiosity into practical engineering capability. Sandia National Laboratories has demonstrated that spiking neural networks running on Intel’s Loihi 2 neuromorphic processor can solve partial differential equations, the mathematical backbone of physics simulation. Intel has released its third-generation Loihi 3 chip, fabricated on a 4nm process node with eight million neurons and 64 billion synapses. And the University of Texas at San Antonio has launched THOR, the first publicly accessible neuromorphic computing hub in the United States, giving researchers beyond well-funded national laboratories access to brain-inspired hardware for the first time.
Taken together, these milestones mark a shift from demonstrating that neuromorphic chips can process spikes to showing they can do useful work at meaningful scale.
Solving Physics With Spikes
The most technically significant result comes from Sandia, where computational neuroscientists Brad Theilman and Brad Aimone developed an algorithm that maps the finite element method, the standard numerical technique for simulating physical systems, onto spiking neural network hardware. Their paper in Nature Machine Intelligence shows that a spiking neural network can solve the large, sparse linear systems of equations at the core of finite element analysis, achieving meaningful numerical accuracy on Intel’s Loihi 2 neuromorphic platform.
The finite element method underpins simulations of fluid dynamics, electromagnetic fields, structural mechanics, and nuclear weapons physics. It works by dividing a continuous physical domain into a mesh of discrete elements and solving a system of equations across that mesh. Conventional supercomputers handle this by brute-forcing matrix algebra through billions of floating-point operations per second. Theilman and Aimone’s approach instead encodes the mesh structure directly into the connectivity of a spiking neural network, where each neuron corresponds to a mesh node and synaptic weights carry the coefficients of the governing equations.
The approach achieves close to ideal scaling on the Loihi 2 platform and extends to irregular mesh geometries in two and three dimensions, as well as to different classes of partial differential equations including linear elasticity. “You can solve real physics problems with brain-like computation,” Aimone stated in a Sandia press release, adding that the algorithm “retains the structure and dynamics of cortical networks in the brain.”
The discovery also carries a surprising neuroscience implication. The cortical network model that Theilman and Aimone adapted for their algorithm was introduced 12 years ago, and the connection between its dynamics and partial differential equations had gone unnoticed. If the relationship holds broadly, it could offer new computational frameworks for understanding how biological brains process spatial and temporal information, with potential relevance to research on neurological conditions such as Alzheimer’s and Parkinson’s disease.
Loihi 3: An Eightfold Leap in Neuromorphic Density
Intel’s Loihi line has served as the primary research platform for academic and government neuromorphic computing projects since the first chip shipped in 2018. The Hala Point system, deployed at Sandia in 2024, aggregates 1,152 Loihi 2 processors into the largest neuromorphic system ever built, supporting 1.15 billion neurons and 128 billion synapses across 140,544 processing cores while consuming a maximum of 2,600 watts. The system achieves up to 20 petaops and delivers over 15 trillion 8-bit operations per second per watt when running conventional deep neural networks.
Loihi 3, released in January 2026, represents an eightfold increase in per-chip density over Loihi 2. Fabricated on Intel’s 4nm process, each chip integrates eight million neurons and 64 billion synapses while operating at a peak power draw of just 1.2 watts, a 250-fold reduction in energy consumption compared to equivalent GPU-based inference for robotics and autonomous navigation workloads.
The architectural advance that most distinguishes Loihi 3 from its predecessors is its support for 32-bit graded spikes. Earlier neuromorphic chips relied on binary spike signals, limiting them to workloads purpose-built for spiking neural networks. Graded spikes allow the hardware to bridge the gap between traditional deep neural networks and spiking neural networks, enabling developers to port mainstream AI models onto neuromorphic silicon without wholesale algorithm redesign. This is a pragmatic concession: rather than waiting for the spiking neural network software ecosystem to mature, Intel is meeting conventional AI developers where they are.
THOR: Democratizing Access to Brain-Inspired Hardware
Access to neuromorphic hardware has historically required either a relationship with Intel’s research division, a national laboratory affiliation, or participation in a small number of academic consortia. The THOR Neuromorphic Commons, launched on February 23 at UT San Antonio, aims to change that model. Funded by a $4 million grant from the National Science Foundation, THOR provides open access to large-scale neuromorphic hardware for researchers, students, and even K-12 educators across the United States.
The hub is built around the SpiNNaker2 system, developed in partnership with SpiNNcloud, and is led by Dhireesha Kudithipudi, founding director of the MATRIX AI Consortium for Human Well-Being. The initiative operates as a multi-university project giving researchers in AI, physics, life sciences, and neuroscience the ability to run experiments on hardware that most institutions could not independently afford or maintain.
THOR’s significance extends beyond hardware access. Neuromorphic computing currently lacks the standardized toolchains, benchmark suites, and educational curricula that define mature computing platforms. By providing a shared facility with training materials and community resources, the project aims to build the human infrastructure that the field needs alongside the silicon.
Emerging Hardware From the Broader Ecosystem
Intel is not the only company shipping neuromorphic silicon. BrainChip’s Akida platform, designed for ultra-low-power edge inference, targets commercial deployments in vision, audio, and sensor processing. IBM’s NorthPole chip, which achieves 25 times the energy efficiency of leading GPUs on image recognition benchmarks, is designed for deployment at the network edge where power budgets are tight and latency must be minimal.
Meanwhile, fundamental materials research continues to expand the design space. Researchers at Osaka University have created ultrasmall pores approaching the width of individual atoms, published in Nature Communications, that mimic biological ion channels and produce electrical current spikes consistent with neuromorphic signaling. While far from commercial application, such work points toward a future where neuromorphic processors could be built from materials that more closely replicate the physics of biological neurons rather than approximating them in conventional transistor logic.
What Neuromorphic Computing Is Not
It is worth stating what neuromorphic computing does not currently offer. These chips are not general-purpose replacements for CPUs or GPUs. They do not run large language models, train foundation models, or handle the dense matrix multiplications that dominate today’s AI data center workloads. Loihi 3’s eight million neurons, while impressive by neuromorphic standards, represent a tiny fraction of the parameter counts in modern transformer architectures.
The value proposition is narrower and more specific: extremely low power consumption for inference tasks that involve sparse, event-driven data. Robotics, autonomous navigation, sensor fusion, anomaly detection, and certain classes of scientific simulation fit this profile. The ANYmal D Neuro, a quadruped inspection robot running on Loihi 3, reportedly achieves 72 hours of continuous operation on a single charge, a ninefold improvement over GPU-powered models performing equivalent tasks.
Looking Ahead
The neuromorphic computing market is projected to reach $17.2 billion by 2030, growing at a compound annual rate exceeding 50 percent. Whether the field achieves that trajectory depends on several factors that remain unresolved: whether graded spikes and other hybrid approaches can attract a critical mass of software developers, whether shared facilities like THOR can train enough researchers to populate the field, and whether neuromorphic advantages in power efficiency translate into cost advantages compelling enough to justify the engineering effort of porting workloads away from GPUs.
Sandia’s demonstration that spiking hardware can solve partial differential equations is perhaps the most consequential data point. It suggests that neuromorphic computing’s utility extends beyond pattern recognition and into the domain of numerical simulation, where the energy costs of exascale supercomputing are becoming a first-order constraint. If brain-inspired hardware can handle physics at a fraction of the power budget, the implications reach well beyond the neuromorphic computing community.