MagStack™ Clustering

Name: Thox OS
Author: Thox.ai LLC

Revolutionary magnetic stacking technology for distributed AI inference

Proprietary Technology - U.S. Patent Pending

This documentation describes proprietary trade secrets of Thox.ai LLC. Unauthorized reproduction, implementation, or reverse engineering is strictly prohibited.

MagStack™ Technology

U.S. Patent Pending | Proprietary Trade Secret

Revolutionary magnetic stacking technology that enables multiple Thox.ai devices to combine RAM and compute power for running larger AI models.

Stack More. Compute More.

How MagStack™ Works

1. Physical Stacking

Simply place Thox.ai devices on top of each other. The N52 neodymium magnets automatically align the devices with sub-millimeter precision. No cables required.

2. Auto-Discovery

NFC antennas and Hall effect sensors detect the connection. Devices exchange capabilities and form a cluster via mDNS within ~10 seconds.

3. Leader Election

The cluster automatically elects a leader using a deterministic algorithm. Priority: external power → highest battery → lowest UUID.

4. Distributed Inference

AI models are partitioned using pipeline parallelism. Each device processes specific layers, with activations flowing via 10 Gbps link.

The Stacking Process

When you stack two Thox.ai devices, a precise 5-step sequence automatically forms a compute cluster:

Step A: NFC Approach Detection

As devices approach within 30mm, the NFC antennas (ST25DV64K, 13.56 MHz) detect each other. This triggers the handshake sequence before physical contact.

Range: 30mm detection, 4mm handshake initiation

Step B: NFC Handshake

At 4mm range, devices exchange critical information via NFC:

Device UUID and firmware version
Available RAM and GPU cores
Cluster capabilities and current state
Security certificates for authentication

Step C: Magnetic Alignment

8× N52 neodymium magnets (4 per surface) with alternating N-S-N-S polarity guide the devices into perfect alignment. Self-centering corrects up to 5mm offset.

~25N pull force|±0.5mm accuracy|±2° rotational tolerance

Step D: Physical Connection

The 12-pin pogo connector engages, establishing:

10 Gbps USB 3.2 Gen 2 data link for model layer activations
USB Power Delivery up to 100W passthrough
I2C bus for power negotiation and telemetry

Pogo pins: 2.54mm pitch, 50-100 gf spring force, <20mΩ contact resistance

Step E: Cluster Active

The software layer completes cluster formation:

Leader elected (external power → highest capacity → lowest UUID)
AI models automatically partitioned across available memory
Health monitoring begins (heartbeat every 2 seconds)
Cluster ready for inference in ~10 seconds

Distributed AI Inference

Pipeline Parallelism

MagStack uses pipeline parallelism to distribute large AI models across devices. Each device processes specific layers, with intermediate activations flowing between devices via the 10 Gbps pogo pin connection.

# Example: 70B model on 2-device stack

70B Model (Q4 quantized)

├─ Layers 0-16 (35GB) → Device 1 (16GB RAM)

└─ Layers 17-32 (35GB) → Device 2 (16GB RAM)

# Data flow

INPUT → Device 1 → [10 Gbps Link] → Device 2 → OUTPUT

Token Latency

~200ms

per token (70B model)

Generation Speed

5-8 tok/s

70B model on 2x stack

Cluster Efficiency

50-60%

sub-linear scaling (2 devices)

Cluster Configurations

Configuration	Devices	Combined RAM	Compute	Max Model	Speed
Single Device	1x	16GB	100 TOPS	32B	20-72 tok/s
2x Stack	2x	32GB	200 TOPS	70B	25-45 tok/s
4x Stack	4x	64GB	400 TOPS	100B+	15-30 tok/s
8x Stack	8x	128GB	800 TOPS	200B+	10-20 tok/s

Recommended Models

2-Device Stack (32GB)

thox-cluster-nano - 30B, 1M context (Recommended)
thox-cluster-70b - 70B parameters

4-Device Stack (64GB)

thox-cluster-100b - 100B+ parameters

8-Device Stack (128GB)

thox-cluster-200b - 200B+ frontier model

Special: 1M Context

The thox-cluster-nano model offers 1 million token context for processing entire codebases in a single session.

Cluster CLI Commands

# Check cluster status

$ clusterctl status

# View cluster topology

$ clusterctl topology

# Load a cluster model

$ clusterctl load thox-cluster-nano

# Check recommended model

$ clusterctl recommend

# Monitor real-time metrics

$ clusterctl monitor --realtime

Key Features

Magnetic Alignment

4x N52 neodymium magnets with alternating polarity ensure perfect alignment every time.

Auto-Discovery

Devices automatically discover each other via mDNS and form clusters when stacked.

Combined Resources

Stack 2-8 devices to aggregate RAM (32GB-128GB) and compute (200-800 TOPS).

Leader Election

Deterministic leader election ensures consistent cluster coordination.

Anti-Shear Design

Precision alignment pins prevent lateral movement for stable stacking.

Hot-Swappable

Add or remove devices from the stack without interrupting the cluster.

Technical Specifications

magnets: 8x N52 Neodymium (Ø10×5mm)
magnet Polarity: Alternating N-S-N-S pattern
pull Force: ~25N combined per interface
alignment Pins: 2x Ø3mm stainless steel
dock Detection: AH3362 Hall effect sensor
nfc Module: ST25DV64K (13.56 MHz)
nfc Range: 30mm detection, 4mm handshake
pogo Connector: 12-pin (2.54mm pitch)
data Link: 10 Gbps USB 3.2 Gen 2
power Passthrough: USB-PD up to 100W
interconnect: Wi-Fi 6E / 2.5GbE
max Stack Height: 8 devices
stacking Tolerance: ±0.5mm accuracy
cluster Formation: ~10 seconds

Intellectual Property Notice

MagStack™ is a trademark of Thox.ai LLC. U.S. Patent Pending.

Patent Protection

The MagStack™ magnetic stacking interface system, including the magnetic alignment mechanism, NFC-based device discovery protocol, pogo pin data interconnect, automatic cluster formation algorithm, and distributed inference methodology, is proprietary technology protected by U.S. Patent Pending and international patent applications. This technology constitutes valuable trade secrets of Thox.ai LLC.

Prohibition Notice

NOTICE: Unauthorized manufacture, use, sale, offer for sale, or importation of devices or systems implementing MagStack™ technology, or any substantially similar technology, is strictly prohibited and may result in civil liability and criminal prosecution. Reverse engineering, disassembly, or decompilation of MagStack™ hardware or software is expressly forbidden.

CONFIDENTIAL AND PROPRIETARY INFORMATION

This documentation is provided for informational and operational purposes only. The specifications and technical details herein are subject to change without notice. Thox.ai LLC reserves all rights in the technologies, methods, and implementations described.

Nothing in this documentation shall be construed as granting any license or right to use any patent, trademark, trade secret, or other intellectual property right of Thox.ai LLC, except as expressly provided in a written agreement.

Patent Protection

The MagStack™ magnetic stacking interface technology, including the magnetic alignment system, automatic cluster formation, NFC-based device discovery, and distributed inference me...

Reverse Engineering Prohibited

You may not reverse engineer, disassemble, decompile, decode, or otherwise attempt to derive the source code, algorithms, data structures, or underlying ideas of any Thox.ai hardwa...

Thox.ai™, Thox OS™, MagStack™, and the Thox.ai logo are trademarks or registered trademarks of Thox.ai LLC in the United States and other countries.

NVIDIA, Jetson, TensorRT, and related marks are trademarks of NVIDIA Corporation. Ollama is a trademark of Ollama, Inc. All other trademarks are the property of their respective owners.