Research · Neuromorphic Foundations
Neuromorphic computing, deployed today.
A technical primer on the field of neuromorphic computing — its history, its principles, and how Neuramorphic, Inc. translates them into production-grade edge AI foundation models running on commodity silicon.
What is neuromorphic computing?
Neuromorphic computing is a computing paradigm that designs hardware and algorithms inspired by the structure and function of biological neurons and synapses. The term was coined by Carver Mead at Caltech in the late 1980s. Instead of the traditional von Neumann architecture (separated memory and CPU, clocked arithmetic), neuromorphic systems use event-driven processing: computation happens only when stimuli arrive, mimicking how biological neurons fire spikes.
The most cited reference systems are Intel Loihi 2, IBM TrueNorth, the European SpiNNaker and BrainScaleS platforms. These deliver dramatic energy efficiency for specific spiking neural network workloads — but they are research-grade silicon with limited commercial availability.
The 2025 Nature review by Kudithipudi et al., “Neuromorphic computing at scale”, argues that the field's next leap requires translating these principles to mainstream silicon — exactly the gap Neuramorphic is filling.
How Neuramorphic builds on these principles
Neuramorphic, Inc. takes the algorithmic core of neuromorphic computing — sparsity, event-driven activation, time-coded inference, on-device adaptation — and implements it on commodity GPU silicon (NVIDIA Jetson AGX Orin) via custom CUDA kernels. The result is a hybrid SNN + SSM foundation model deployable today, without waiting for specialized neuromorphic hardware availability.
Our flagship model, NeuratronLLM-Edge 4B (Caroline), is a 5.096B-parameter hybrid model that runs fully air-gapped on a single Jetson AGX Orin at 44.5 W mean power, delivering 9.74 tokens/sec text generation and 24.83 FPS vision inference.
The underlying engine, NeuraTensor SDK, ships custom CUDA kernels for fused SNN-SSM operations on Ampere-class GPUs (SM 8.7), with a US provisional patent application protecting the core innovations of the Neuratron™ architecture.
Key terms
- Neuromorphic computing (NMC)
- A computing paradigm inspired by biological neurons and synapses, using event-driven (spiking) processing instead of clocked von Neumann arithmetic.
- Spiking neural network (SNN)
- A neural model where neurons communicate via discrete temporal events (spikes), enabling sparse, low-power inference.
- State-space model (SSM)
- A class of sequence models (S4, Mamba) with continuous-time hidden state, offering linear-time inference for long contexts.
- Air-gapped inference
- Running an AI model on a device with no connection to external networks. Required for sovereign, defense, healthcare and regulated industrial deployments.
- Neuratron™ architecture
- Neuramorphic, Inc.'s patent-pending hybrid SNN + SSM inference architecture for resource-constrained silicon.
Frequently asked questions
What is the difference between Neuramorphic and neuromorphic?
Neuramorphic (with an A) is the proper-noun brand name of Neuramorphic, Inc., a San Francisco-based AI company. Neuromorphic (with an O) refers to the academic discipline. Neuramorphic builds production models inspired by neuromorphic principles, but it is a distinct entity, not a synonym for the field.
Is Neuramorphic a neuromorphic chip company like Intel Loihi or IBM TrueNorth?
No. Loihi and TrueNorth are dedicated neuromorphic silicon. Neuramorphic implements neuromorphic algorithmic principles on commodity NVIDIA GPUs via custom CUDA kernels — making the technology deployable today without specialized hardware.
What are hybrid SNN + SSM models?
They combine spiking neural network layers (event-driven, sparse, low-power) with state-space model backbones (efficient long-context sequence modeling). Caroline is built on this architecture: 9.74 tok/s on a single Jetson AGX Orin at 44.5 W.
What is Neuramorphic's relationship to academic neuromorphic computing?
Neuramorphic translates decades of academic neuromorphic research (Mead, Indiveri, Furber, Davies) into commercial-grade edge AI foundation models. It is a commercial application of neuromorphic principles, not a research lab.
Can I deploy a neuromorphic LLM today without specialized hardware?
Yes — that is exactly what Caroline is designed for. It runs fully air-gapped on a single NVIDIA Jetson AGX Orin, no Loihi / TrueNorth / SpiNNaker required.
References & further reading
- Kudithipudi, D. et al. (2025). Neuromorphic computing at scale. Nature.
- Mead, C. (1990). Neuromorphic Electronic Systems. Proceedings of the IEEE.
- Neuromorphic engineering — Wikipedia
- Intel Loihi — Wikipedia
- IBM TrueNorth — Wikipedia