Z2 hardware · NeuroOS software

Low-cost neuromorphic computing at cluster scale.

Z2 is a spiking neural network computer that connects directly to the sensors it reads. It brings modular hardware, managed workloads and agent APIs into one system. Designed and built in Texas. Shipping to early-access partners.

96compute nodes per Z2
192processor cores per Z2
2Ustandard rack format
≤180 Wsystem design envelope
Illustrative spike simulationactive neurons —spikes/s —
NeuroFab Z2 with the lid removed, showing 96 node cards across six backplanes
Six banks of NeuroFab compute cards with illuminated status indicators
Inside the Z2

96 nodes. Six backplanes. One system.

In use by research groups atThe University of Texas at Austin
Research collaborationRadboud University
Backed byWorldQuant Foundry

Meet NeuroOS

The hardware.
The workloads.
One clear view.

Choose a model or an experiment. NeuroOS assigns capacity and tracks the work. See which nodes hold each model, follow requests through the queue, and inspect the completed results.

Explore the workspace →
The current NeuroOS Machine dashboard
Real NeuroOS interface · engineering system, November 2026
Automatic placementShared resource managementFly impulse APISandia science APIRecorded results
Why spiking networks

Compute when the signal changes.

A spiking network represents information as events in time. A neuron sends a spike when its state crosses a threshold. This lets sparse signals drive sparse work, with the result shaped by both the model and the hardware.

Event-driven

Neurons send spikes when their state crosses a threshold. Sparse activity can reduce computation. The power used depends on the model, the hardware and the input.

Native to time

Spike timing carries information. Models can use temporal patterns in EEG, vibration, audio or touch. Input preparation depends on the application.

Robust at the edge

Local processing keeps sensor data close to the source. Models can be trained for noise and missing inputs, then tested on the signals they will meet in use.

See how a spiking neuron computes →
Z2 Neuromorphic Computer

A complete spiking computer. Built to work as one system.

The brain computes with sparse spikes in time. Z2 brings that approach to a modular system, with memory and sensor connections at each node.

0
node cards, field-replaceable
0
cores, Arm Cortex-M33 and Hazard3 RISC-V
0+
concurrent sensor inputs
0W
design power envelope, whole system
16,384
neuron-state slots per qualified node
1.57M
neuron-state slots per 96-node Z2

Capacity assumes 96 qualified nodes at 16,384 neuron-state slots each. Actual model size also depends on synapses, state and available memory.

Z2 Early Access Program

A limited allocation of Z2 systems for university and industry research labs, shipping with the model library, training toolkit and documentation.

Z2 Neuromorphic Computer

The Z2 neuromorphic computer.

A complete spiking neural network system in a standard 2U rack unit: 96 node cards, 192 cores, memory at every node and a sensor bus on every card, for more than 1,000 concurrent sensor inputs. Rack it, connect Ethernet, deploy a model.

NeuroFab Z2 from the front, lid off, with its hexagonal vent panel and power ring
Architecture

A fabric that scales like a cortex.

Each backplane carries sixteen node cards and a controller. Nodes talk over a 16-bit parallel matrix bus with hardware-assisted signalling; backplanes talk over Ethernet. The hierarchy scales like a cortex: local when it can be, global when it must be.

  • Every node is a complete computer with its own memory, clock and sensor bus.
  • Controllers coordinate node communication. The host manages workloads that span backplanes.
  • Controller OLED diagnostics show backplane state at a glance.
Z2 backplane with node cards, controller OLED displays and interconnect16-node backplane · controller OLEDs
Node cards

Modular node cards, serviceable in the field.

Each card carries a dual-core embedded processor with Arm Cortex-M33 and Hazard3 RISC-V cores, programmable I/O state machines, 8 MB of local memory and 16 MB of flash. Cards are field-serviceable. When a better node exists, you swap cards, not systems.

  • Dedicated sensor bus per card for direct sensor attachment and fusion at the edge.
  • A single card runs standalone for development, then drops into the fabric.
  • Roadmap path to FPGA and ASIC node cards in the same chassis and software.
A single Z2 node card running standalone in front of a Z2 unitNode card · standalone development mode
Deploy

Train on a workstation. Deploy to the whole fabric with one command.

The Z2 training toolkit runs on an ordinary workstation. Train a spiking model with hardware-aware quantization, package it, and push it to the whole fabric with NeuroOS. Firmware updates travel the same path, so a Z2 in the field is never frozen at the version it shipped with.

  • Managed transport carries model data and inference requests between the host and controllers.
  • Model packaging with on-device verification after every deploy.
  • The current engineering catalog includes 29 entries, with automatic placement and execution records.
Z2 node cards seen from above with the lid removedZ2 · lid off
Model library

Earlier model benchmarks, measured on hardware.

Every Z2 ships with five spiking models trained for the fabric and validated on the fabric, from scalp EEG to spoken digits. These are the numbers the system produces on the bench, not in a simulator.

Latency is end-to-end per sample on NeuroFab hardware. Energy per inference measured for the seizure model.

WorkloadAccuracyLatency
Seizure detectionScalp EEG · 71.7 µJ per inference94.3%19.6 ms
Handwritten digitsMNIST93.0%4.8 ms
Human activity recognitionWearable inertial sensors92.3%5.0 ms
Spoken digitsFSDD audio92.0%105 ms
Arrhythmia detectionSingle-lead ECG88.2%6.3 ms
Z2 front panel
Ten Z2 units · 20U · photographed front panel
Rack scale

Scale the fabric over Ethernet.

Z2 units connect over standard Ethernet. Ten fully qualified 96-node units provide 15.7 million neuron-state slots at the current 16,384-slot configuration. Synapse storage, model state and communication determine how much of that capacity a workload can use.

0
neuron-state slots per ten-unit rack
0kW
design power for the full rack
Specifications

Z2 specifications

Form factor
2U rack-mount, 19-inchRear-fan cooling, central power distribution
Compute
96 node cards · 192 coresDual-core Arm Cortex-M33 / Hazard3 RISC-V per card, programmable I/O state machines, up to 300 MHz
Topology
6 backplanes × (16 node cards + 1 controller)Modular, field-serviceable cards
Memory
8 MB PSRAM + 16 MB flash per nodeApproximately 2.3 GB aggregate across the unit
Fabric
16-bit parallel matrix busMulti-master, any node to any node, hardware-assisted signalling
Sensor I/O
1,000+ concurrent sensor inputsDedicated sensor bus per node card; direct attachment and fusion at the edge
Networking
Ethernet per backplanePersistent binary streaming transport for spike I/O
Diagnostics
Controller OLED status, per-node RGB indicatorsFront-panel status ring
Power
≤ 180 W design envelopeEvent-driven operation keeps idle draw low
Software
NeuroOS, workstation training toolkit, model libraryDashboard, model catalog, managed deployment and agent APIs
Hardware

Z2 Neuromorphic System

Put a Z2 in your lab.

The Early Access Program places systems with a small number of research groups working on sparse, temporal, sensor-proximate problems.

Technology · Platform

Event-driven computing, from neuron to rack.

Spiking neural networks are the third generation of neural computation. They replace continuous activations with discrete events in time, which is how a brain does remarkable work on twenty watts. NeuroFab makes that model practical in hardware you can buy.

The neuron

The leaky integrate-and-fire neuron.

A spiking neuron accumulates incoming events on a membrane potential that slowly leaks away. When the potential crosses threshold it emits a single spike and resets. Sparse spike activity can reduce synaptic work. The hardware still maintains neuron state and communication.

Drive the neuron below. Raise the input rate and watch output spikes appear only when evidence accumulates faster than it leaks.

0output spikes/s
Dense vs. spiking

Dense networks versus spiking networks.

Dense networks and spiking networks organize work differently. Dense layers apply tensor operations. Spiking layers exchange discrete events. Sparse inputs can reduce synaptic work; the gain depends on the model and its implementation.

Dense networksSpiking networks
ActivityDense tensor operations and memory movementSparse events; work follows activity
TimeRecurrence or windowing added onNative to spike trains and neuron state
RobustnessDepends on training and model designDepends on training and spike encoding
Best forTraining and batch inferenceSensors, control, event streams
Dense network

Every unit, every step

0
multiply-accumulates so far
Spiking network

Only what fired

0
synaptic events so far
Same 100-unit layer, same input stream. The spiking side does — of the work.Illustrative simulation, not a benchmark.
Architecture principles

System architecture first. Custom silicon when it earns its place.

Most neuromorphic efforts start with a chip and hope for a system. NeuroFab started with a system. Off-the-shelf embedded processors, a custom fabric and a real operating system let us ship hardware now and learn from real workloads before committing designs to silicon.

No global clock
Nodes run asynchronously and communicate through events.
Coordinated communication
Controllers manage communication within each backplane. The host coordinates supported workloads across backplanes.
Memory in the connections
Weights live next to the neurons that use them, on every node.
Sensory democracy
Any node can own a sensor. Understanding emerges from many partial views.
Row of Z2 node cards with OLED displaysEvery node is a complete computer
Toolchain

Train, package, deploy.

The Z2 toolchain meets researchers where they already work: train spiking models on a workstation, quantize with the hardware in the loop, and deploy with NeuroOS, the same tools you use to manage the cluster.

TRAIN

Hardware-aware training

Train on your workstation with constraints that mirror the fabric: quantized weights, bounded spike queues, fixed-point dynamics. Validate the deployed model against the workstation reference.

PACKAGE

Model packaging

Compile topology and weights into a compact package for the fabric. Supported deployment adapters map models to available capacity; you can inspect or override placement.

DEPLOY

One-command deploy

Use the managed deployment flow to load a supported model. The platform records placement and execution. Firmware updates are a separate operation.

Go to NeuroOS →
Direction

Platform roadmap.

Z2 is the first shipping system, not the last. The research agenda is learning on the hardware itself and driving energy per synaptic event down by orders of magnitude.

On-hardware learning

Spike-timing dependent plasticity running on the node, so a deployed system keeps adapting to its own environment without a training server.

Sparsity-preserving traversal

Synaptic pointers that visit only downstream neurons that were actually activated, so cost tracks events all the way through the network.

Next-generation nodes

A hardware revision with energy conservation as the primary design objective, and a path to FPGA and ASIC node cards in the same chassis.

Technology · NeuroOS

One workspace.
The whole fabric.

NeuroOS is the operating system for the Z2 fabric. It turns every node, backplane and controller into one managed system. Choose a workload. The platform finds capacity, deploys the model and records the result.

NeuroOS Machine view showing the two-box fleet, node placement and capacity
Machine view · captured from the engineering system, November 2026. Counts and placement are a recorded snapshot.

Resource management

Ask for the work.
Let the system place it.

Automatic placement is the default. A controller or node override is available when your experiment needs a fixed location.

01

Find capacity

The scheduler first checks free, qualified nodes. It can reclaim eligible idle models when it needs space. Hard reservations stay protected.

02

Deploy and run

Requests move through queued, deploying and running states. Resource leases keep workloads from changing nodes that another job owns.

03

Return the result

Jobs retain their status and execution records. Applications show completed outputs, and API clients can retrieve them by run ID.

Model library

A growing set of workloads.

Image and signal models share a workspace with scientific experiments. The current engineering catalog contains 29 entries, with automatic placement available across the catalog.

The NeuroOS model library with automatic placement controls
Model library · current engineering UI. Available models and capacity depend on the installed release.
MNIST & Fashion MNISTEEG & seizure signalsSpoken digitsHuman activityECGScientific computing

Agent workflows

Submit a job.
Receive a hardware result.

Use HTTP APIs from an application, a script or an LLM agent. Submit a request, keep its job ID, then retrieve the completed result. A stable idempotency key lets a client retry a submission without creating another logical job.

Fly connectome

Impulses in. Spike events out.

Send current injections to an adapted fly connectome. Receive per-tick neuron IDs and spike counts for analysis or client-side animation. The host coordinates ticks and relays spikes between hardware shards.

POST /live/api/apps/fly-brain/queue
Idempotency-Key: fly-example-001

{
  "steps": 3,
  "keep_loaded": true,
  "impulses": [[[1200, 250]], [], []]
}

Poll the returned job ID. After success, retrieve the run export. These endpoints run on your NeuroFab host.

Sandia workloads

Scientific experiments, through an API.

Run supported heat and particle diffusion experiments with explicit workload sizes. The scheduler selects placement when you omit a box. Completed runs provide scientific results and execution records.

POST /live/api/apps/sandia/inference
Idempotency-Key: science-example-001

{
  "problem": "heat", "workers": 1,
  "side": 4, "steps": 8,
  "particles": 128
}

Read the workload contract, submit the job, then fetch the result by run ID. An accepted request may wait for capacity.

A clear view of the machine

See what is loaded.
See what is running.

Model placement

Node segments show recorded model residency. Select a node to see its models and assigned jobs.

Work in progress

Distinct effects show deployment, inference and clearing phases. The UI marks stale observations and provides fleet refresh controls.

Machine control

Stop and free cancels work and releases models that can be safely unloaded. Admissions remain paused until you resume them.

NeuroOS ships with every Z2.

Early-access partners receive the workstation toolkit, model library and documentation, with direct access to the engineers who built them.

Request access →
Research & applications

Research and applications.

NeuroFab works with university and clinical researchers to put spiking hardware against problems that batch pipelines have never served well: continuous, noisy, time-critical signals at the edge.

Clinical neuromonitoring

Neuromonitoring alerts in under 20 milliseconds.

During spine and brain surgery, intraoperative neurophysiological monitoring watches evoked potentials for signs of nerve compromise. Today's workflows can take tens of seconds to raise an alert. With researchers at Radboud University, NeuroFab is developing spiking models that flag neural instability in under 20 ms, directly from the monitoring stream.

  • Continuous, event-driven analysis of evoked potentials and EEG.
  • On-hardware seizure detection today: 94.3% accuracy at 19.6 ms and 71.7 µJ per inference.
  • A path toward adaptive monitoring that learns the patient during the case.
ILLUSTRATIVE SIGNAL · DETECTOR EVENTS IN AMBER
19.6 msalert latency on hardware
In the lab

NeuroFab systems in research.

Research groups at The University of Texas at Austin are using NeuroFab systems to develop training methodologies for spiking neural networks, to evaluate the robustness of spiking classifiers, and to examine applications in robotics.

  • Training methods that close the gap between simulation and spiking hardware.
  • Robustness of spiking classifiers under noise, drift and adversarial inputs.
  • Robotics: distributed sensing and reflexive control on the fabric.
Z2 with the lid removed on a benchZ2 · research unit
Application areas

Sparse, temporal, sensor-proximate.

If your problem has these three properties, it is a spiking problem. These are the domains where we are placing early systems.

Biomedical signals

EEG, ECG, EMG and evoked potentials. Continuous classification at the bedside and on the body, at energy budgets that fit a wearable.

Robotics & tactile sensing

Robot skin, force and slip detection, event cameras. One node per patch of sensors, fused across the fabric.

Industrial monitoring

Vibration, acoustic and electrical signatures of machines, watched continuously, alerting only on anomalies.

Audio & keyword spotting

Always-listening acoustic models that spend energy only when sound is present.

Time-series prediction

Forecasting and anomaly scoring over streaming telemetry without windowing or retraining cycles.

Autonomous systems

Low-power perception and reflexive control for platforms that cannot carry a GPU or a cooling budget.

Research partner program

We place Z2 systems with groups who will publish. Partners get hardware, the training toolkit, direct access to our engineers, and a say in what the platform becomes.

Compute services

From neural circuits to scientific models.

The Fly application runs an adapted connectome on hardware and returns spike events. The Sandia application runs supported diffusion experiments and returns scientific results. Both expose job-based APIs for scripts and agent workflows.

Explore the API workflows →
Team

Leadership.

Founders who have built and scaled AI companies, a pioneer of parallel and GPU computing, and operators and investors who have backed deep technology for decades.

Portrait of Amir Husain

Amir Husain

Founder

Entrepreneur, author and AI leader. Founder of SparkCognition, SkyGrid (acquired by Boeing) and Argon, and Chairman of WorldQuant Foundry. Named on roughly fifty patents; former member of the NATO Maritime Unmanned Systems Innovation Advisory Board.

Portrait of Professor Don Fussell

Prof. Don Fussell

Trammell Crow Regents Professor, UT Austin

Chair of the Department of Computer Science at The University of Texas at Austin and Director of the Laboratory for Real-time Graphics and Parallel Systems. A pioneer of parallel systems research whose students have taken key roles across the GPU industry.

Portrait of Steve Lau

Steve Lau

Managing Director, WorldQuant Ventures

Leads deep-tech investments across fintech, AI, biotech, robotics and quantum, and has backed more than ninety companies since 2014.

Portrait of Jesús Mantas

Jesús Mantas

Independent Director, Biogen

Former IBM Global Managing Partner who led Business Transformation Services. Serves as an Independent Director and Chair of the Compensation & Management Development Committee at Biogen.

Backing

Backed by WorldQuant Foundry.

NeuroFab is funded by WorldQuant Foundry and a group of distinguished individual investors. The company holds patent filings on its fabric architecture and neuromorphic methods, with additional filings underway.

WorldQuant Foundry
Wide view of the Z2 node arrayAustin, Texas

Build the third generation of computing with us.

We are hiring in Austin and remotely for embedded firmware, neuromorphic research, hardware design and developer tooling.

Contact

Request early access.

Tell us about the signal you are working with. We allocate Z2 systems to research groups and industry partners whose problems are sparse, temporal and close to the sensor.

Research partnerships

University labs and clinical groups planning to publish on spiking hardware.

Industry pilots

Teams with a continuous sensor stream and a latency or power problem a GPU cannot solve.

Investors & press

Write to us at hello@neurofab.com.

Opens your mail client with the request pre-filled.