Hardware for spiking neural networks

For a research lab building a neuromorphic computing platform for spiking neural networks (SNNs), I would focus on companies developing purpose-built neuromorphic processors, event-driven AI chips, and scalable spiking-compute systems—rather than conventional AI accelerators that simply run neural networks efficiently.

As of October 2026, the ecosystem includes established research platforms, commercially oriented chipmakers, and specialized sensor-processing companies. The most useful distinction is whether you need programmable SNN research hardware, deployable low-power chips, or a large-scale neuromorphic system.www.intel.com+2

1. Key companies to investigate

• 

1. Intel

(Research platform)

Hardware: Loihi 2, Kapoho Point, Hala Point

Intel develops programmable neuromorphic processors using asynchronous, event-driven SNNs. Its Lava software framework supports developing and mapping neuro-inspired algorithms to compatible hardware. Hala Point is a large-scale research system built from Loihi 2 processors.www.intel.com+1

Best research fit: Programmable SNN experimentation, algorithm–hardware co-design, online adaptation, and scaling experiments.

Website:Intel Neuromorphic Computing

---

• 

2. BrainChip

(Edge AI chips)

Hardware: Akida neuromorphic processor family

BrainChip develops event-based, low-power AI hardware for edge inference, with support for neuromorphic processing and on-chip learning capabilities. Its ecosystem is oriented toward embedded devices rather than only large research installations.open-neuromorphic.org+2

Best research fit: Embedded SNN experiments, always-on sensing, event-based inference, and edge deployment.

Website:BrainChip

---

• 

3. SynSense

(Event-driven sensing)

Hardware: Speck and DYNAP product families

SynSense develops digital and mixed-signal neuromorphic processors, including hardware for event-based vision and low-power SNN processing. Speck integrates neuromorphic vision sensing and event-driven computation in a compact system-on-chip.open-neuromorphic.org+2

Best research fit: Event-camera processing, visual perception, robotics, and low-power sensor fusion.

Website:SynSense

---

• 

4. Innatera

(Commercial SNN hardware)

Hardware: Pulsar neuromorphic microcontroller; earlier T1 processor

Innatera builds spiking neural processors for ultra-low-power, real-time sensor intelligence. Pulsar combines SNN computation with a RISC-V CPU and conventional signal-processing and CNN accelerators. The company announced Pulsar's commercial launch in May 2025.www.innatera.com+1

Best research fit: Wearables, industrial sensors, acoustic or radar sensing, and battery-powered edge systems.

Website:Innatera

---

• 

5. SpiNNcloud

(Scalable systems)

Hardware: SpiNNaker2-based computing systems

SpiNNcloud commercializes brain-inspired computing infrastructure based on the SpiNNaker research lineage. Its architecture uses large numbers of processing cores and event-based communication, targeting scalable, energy-efficient workloads.spinncloud.com+5

Best research fit: Large-scale spiking networks, distributed neural simulation, and experiments that require more compute capacity than a small edge chip.

Website:SpiNNcloud

2. Additional companies and research platforms

These are also relevant, but their hardware differs in how directly it supports SNN research.

• 

Heidelberg University BrainScaleS — BrainScaleS-2

An analog/digital neuromorphic system with configurable spiking neurons, plastic synapses, and accelerated neural dynamics. Particularly relevant for studying neural models and synaptic plasticity. This is a university research platform rather than a conventional commercial chip vendor.www.bmm.uni-heidelberg.de+1

BrainScaleS research platform

---

• 

IBM Research — TrueNorth and NorthPole

IBM has developed landmark brain-inspired chips. TrueNorth used spiking, event-driven computation; NorthPole instead emphasizes distributed processing and memory-near-compute. NorthPole is relevant to neuromorphic-adjacent architecture research, but should not be assumed to be a directly programmable SNN platform.research.ibm.com+1

IBM Research: brain-inspired computing

---

• 

GrAI Matter Labs

Develops brain-inspired edge AI processors, with an emphasis on low-latency, energy-efficient processing. It is worth investigating for event-driven edge architectures, although its hardware should not automatically be treated as a general-purpose programmable SNN accelerator.www.mordorintelligence.com+1

Website:GrAI Matter Labs

3. How the platforms compare for your lab

Sources: manufacturer documentation and research-platform descriptions.www.intel.com+5

4. What I would evaluate first

For a lab building its own platform, I would begin with three complementary tracks:

• Programmability and SNN research: Intel Loihi 2 and Heidelberg BrainScaleS-2.
• Commercial chips and evaluation kits: BrainChip Akida, SynSense, and Innatera.
• Scale-out and distributed computing: SpiNNcloud's SpiNNaker2 systems.

Before committing to hardware, check these criteria:

1. Network scale: Supported neurons, synapses, connectivity, and number of chips.
2. Learning support: On-chip plasticity, online learning, and support for training versus inference.
3. Software access: APIs, simulators, model conversion, Python integration, and compiler availability.
4. Hardware access: Whether research access, evaluation boards, development kits, or production parts are actually available to your institution.
5. Performance measurement: Energy per inference or spike, end-to-end latency, and performance on your own SNN workloads.

One important caveat: neuromorphic computing and brain-inspired AI are not interchangeable. Some chips implement explicit spiking neurons and event-driven computation; others borrow ideas such as parallelism and memory-near-compute without implementing SNNs directly.

For your platform, the most important question is whether you need hardware to execute and study spiking neural networks, or hardware to deploy energy-efficient AI inspired by the brain. That distinction will substantially narrow the field.

The primary commercial and research-grade vendors providing hardware for spiking neural networks (SNNs) and brain-inspired AI include Intel, BrainChip, SynSense, SpiNNcloud Systems , and IBM.youtube.com+2

Key Neuromorphic Hardware Vendors and Platforms

• Intel (Loihi / Hala Point)
  • Platform/Chip: Loihi 2 and the Hala Point large-scale neuromorphic research system.
  • Focus: Asynchronous, event-driven, digital neuromorphic processing designed for real-time edge intelligence, online learning, and massive-scale spike simulation.supercomputing.news
• BrainChip (Akida)
  • Platform/Chip: Akida (1st and 2nd generation event-driven neural processors).
  • Focus: Ultra-low-power, commercial edge AI accelerators supporting direct spike-based processing, one-shot learning, and compatibility with standard PyTorch/TensorFlow training pipelines.youtube.com
• SynSense
  • Platform/Chip: Speck, DYNAP series.
  • Focus: Mixed-signal (analog/digital) neuromorphic vision and audio sensors and processors built for ultra-low-latency, low-power edge deployment.
• SpiNNcloud Systems (SpiNNaker2)
  • Platform/Chip: SpiNNaker2 (developed originally at TU Dresden / University of Manchester lineage).
  • Focus: Massively parallel, core-mesh architectures combining small Arm cores with dedicated accelerators for SNNs and deep neural networks, capable of large-scale brain modeling and high-efficiency inference.supercomputing.news
• IBM (NorthPole)
  • Platform/Chip: NorthPole.
  • Focus: Advanced architecture with colocated memory and computation that mimics aspects of biological efficiency, achieving high throughput and low power on neural workloads.ibm.com
• Fathom Radiant
  • Platform/Chip: Human brain-scale neural training hardware.
  • Focus: Specialized physical network architecture built to scale neural network training to human brain-like capacities.revenuebase.ai

Comparison of Leading Platforms

If you'd like to narrow this down, tell me:

• Are you targeting edge deployment (low-power, embedded) or large-scale cluster simulation in the lab?
• Do you prefer fully digital CMOS chips or mixed-signal/analog designs?

I can help outline the software toolchains (such as snnTorch, NXTK , or proprietary SDKs) associated with each vendor.genomics.ucsc.edu

For a lab developing a neuromorphic platform for spiking neural networks (SNNs), I'd divide the hardware landscape into three groups: research-grade programmable neuromorphic systems, commercially oriented low-power SNN processors, and large-scale neural simulation platforms. The right vendors depend on whether you need to experiment with neuron dynamics and learning rules, deploy inference on physical devices, or scale up network simulations.

As of October 2026, the key names to evaluate are Intel, BrainChip, SynSense, Innatera, SpiNNcloud, and the academic teams behind BrainScaleS-2 and SpiNNaker. IBM is also relevant to the broader brain-inspired computing landscape, although its NorthPole architecture is not equivalent to a fully programmable SNN research chip.open-neuromorphic.org+3

Key vendors and platforms

• 

1. Intel — Loihi 2

(Research platform)

Best suited to: Programmable SNN research, learning rules, and hardware–software co-design.

Loihi 2 supports asynchronous, event-driven spiking networks and programmable neuron dynamics. Intel's Lava framework provides the development environment, and researchers can apply through the Intel Neuromorphic Research Community (INRC) for access to hardware.

Access: Primarily research-community and shared-system access, rather than ordinary retail purchasing.www.intel.com+1

Intel neuromorphic computing

---

• 

2. BrainChip — Akida

(Commercial hardware)

Best suited to: Deploying low-power, event-based AI on embedded and edge devices.

BrainChip offers Akida processor IP, development tools, chips, and reference platforms. Its AKD1500 reached commercial availability with initial production shipments announced in June 2026. Check which supported network types and learning mechanisms fit your SNN design; not every Akida workflow is equivalent to arbitrary, biologically detailed SNN simulation.

Access: Commercial product and IP licensing options.investor.brainchip.com+1

BrainChip products and developer tools

---

• 

3. SynSense — Speck, Xylo, DYNAP

(Sensor-focused hardware)

Best suited to: Event-based vision, audio processing, and low-power sensor intelligence.

The portfolio includes Speck for neuromorphic vision and Xylo for temporal signal processing. DYNAP-family research hardware is also relevant for investigating asynchronous spiking circuits and low-power neural dynamics.

Access: Commercial products and development kits, with research-oriented options varying by platform.www.synsense.ai+3

SynSense official site

---

• 

4. Innatera — Pulsar / T1

(Mixed-signal edge computing)

Best suited to: Always-on sensing, low-latency inference, and power-constrained embedded systems.

Innatera combines an SNN engine with conventional processing elements in a neuromorphic microcontroller. Its hardware is particularly relevant if the lab plans to move trained models into real-time sensor applications rather than simulate very large neural populations.

Access: Commercial hardware and its Talamo development software.open-neuromorphic.org+1

Innatera official site

---

• 

5. SpiNNcloud — SpiNNaker 2

(Scalable neural simulation)

Best suited to: Large-scale SNN experiments, neural population simulations, and distributed network execution.

SpiNNaker 2 combines many programmable processing cores with dedicated neural-computation support. It is attractive when network scale and parallel execution matter more than fitting the entire workload onto a tiny edge device.

Access: Research and larger-system deployments; discuss availability and system configuration directly with the provider.open-neuromorphic.org+4

SpiNNcloud official site

---

• 

6. Heidelberg University — BrainScaleS-2

(Academic research system)

Best suited to: Studying analog neuron dynamics, synaptic plasticity, and accelerated neural experiments.

BrainScaleS-2 uses mixed-signal hardware, with analog circuits for neural dynamics and digital components for control and experimentation. It is especially interesting if your research concerns the physical implementation of neural dynamics rather than only digital spike processing.

Access: Research infrastructure and academic collaborations, rather than a standard off-the-shelf product.open-neuromorphic.org+1

BrainScaleS research platform

Other organizations worth tracking

How I'd shortlist hardware for your lab

If you're inventing new SNN algorithms

Start by investigating Intel Loihi 2 and its Lava ecosystem. Add BrainScaleS-2 if analog dynamics or accelerated physical experiments are central to your research.

---

If you're scaling to millions of neurons or large networks

Evaluate SpiNNaker 2 through SpiNNcloud and compare it with research-system access to Intel's larger Loihi deployments.

---

If you're building an embedded neuromorphic product

Compare BrainChip Akida, SynSense, and Innatera on power, latency, sensor integration, supported neuron models, and the ability to deploy your trained networks.

What to compare before committing

For a serious platform evaluation, I would request benchmarks and SDK access against the same representative network from each vendor. Focus on:

• Model flexibility: neuron models, synaptic precision, connectivity, refractory behavior, and support for custom learning rules.
• Learning: on-chip plasticity and online learning versus inference on pretrained weights.
• Scale: neuron and synapse capacity, inter-chip communication, and practical network partitioning.
• Software portability: Python APIs, training pipelines, simulators, and export paths from frameworks such as PyTorch.
• Measured efficiency: joules per inference or spike, latency, idle power, and energy at the system level—not just the chip.
• Availability: development-kit access, cloud or remote research access, licensing restrictions, and production lead times.

The key distinction is that neuromorphic does not always mean general-purpose SNN hardware. Some platforms are designed for flexible neuroscience research, while others accelerate particular classes of event-driven inference.

One question that would help narrow the list: Is your lab primarily developing new SNN algorithms and learning rules, scaling large neural simulations, or building a deployable low-power neuromorphic processor?