Unified hardware device for the hosting, real-time orchestration and securing of artificial intelligence agents, the device comprising a multi-processor computational architecture and a native aiot gateway

A unified hardware device with a multi-processor system and native ALoT gateway addresses fragmentation issues in AI architectures, providing secure and efficient execution of diverse AI agents with adaptive resource management and resilience.

WO2026235542A1PCT designated stage Publication Date: 2026-11-12ABA TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/MA2026/000009
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-08
Filing Date
2026-06-03
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

Existing hardware architectures are fragmented, leading to significant latency, resource contention, and inefficiency in managing diverse multi-agent AI systems, and lack a unified platform for secure and efficient execution of heterogeneous AI agents.

Method used

A unified hardware device integrating a multi-processor system with a CPU, GPU, and NPU, a native ALoT gateway, and a hardware orchestrator for adaptive resource management, secure isolation, and real-time data preprocessing, along with a scalable and resilient design for dynamic workload distribution.

Benefits of technology

Enables secure, efficient, and adaptive hosting of multiple AI agents with reduced latency and energy consumption, ensuring coherent task orchestration and proactive management across shared resources, with integrated security and resilience.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure MA2026000009_12112026_PF_FP_ABST
    Figure MA2026000009_12112026_PF_FP_ABST
Patent Text Reader

Abstract

The patent relates to an integrated hardware device comprising a unified computational platform specifically designed for the secure hosting and dynamic, autonomous and contextual orchestration of a plurality of intelligent agents (AI). The device incorporates a heterogeneous hardware architecture comprising general-purpose processing units (CPUs), specialised graphics processing units (GPUs) and specialised neural processing units (NPUs), operating under the control of a real-time hardware orchestrator ensuring the dynamic, adaptive and contextual distribution of the loads according to the precise needs of the hosted AI agents. The device also integrates a native AloT gateway compatible with various industrial protocols to allow direct and secure acquisition of loT streams without an external intermediary. Moreover, the device comprises an integrated zero-trust hardware cybersecurity subsystem and an intelligent predictive and contextual energy management module, the module ensuring high performance coupled with a significant reduction in the overall energy footprint, with the reduction being adapted to the strict requirements of industrial and critical environments.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] A unified hardware system for hosting, real-time orchestration, and securing artificial intelligence agents, comprising a multi-processor computing architecture and a native ALoT gateway.

[0002] DETAILED DESCRIPTION

[0003] The state of the art in hardware platforms dedicated to artificial intelligence highlights a persistent fragmentation of computing resources, making traditional architectures poorly suited to the simultaneous, secure, and efficient execution of diverse multi-agent AI systems. In particular, existing architectures generally rely on the structural separation of computing units, such as central processing units (CPUs), graphics processing units (GPUs), and neural processing units (NPUs), each operating within distinct subsystems, often interconnected by generic I / O links (e.g., PCIe, QPI, or other inter-socket protocols). This introduces significant latency, contention on shared buses, and a growing inadequacy to meet the dynamic and contextual needs of heterogeneous intelligent agents.

[0004] Because these units are designed to handle specific workloads, their isolated deployment in monolithic or distributed architectures hinders coherent task orchestration, efficient resource sharing, and integrated energy performance monitoring. This segmentation also makes it difficult to proactively manage the coexistence of intelligent agents of varying technological nature (e.g., generative AI, neurosymbolic AI, autonomous cognitive AI, physical perception-action agents, etc.), thus impeding the development of unified "agent-hosting" platforms.Furthermore, it is observed that artificial intelligence technologies are deployed according to distinct and incompatible computational paradigms at the hardware level: generative AIs, often based on Transformer-type architectures, require massive memory resources and tensor accelerations; agentic AIs rely on event-oriented decision models, requiring fluid management of priorities and logical interconnections; physical AIs rely on rapid sensory acquisition and very low-latency processing cycles; while neuromorphic AIs seek to reproduce spike-timing dependent plasticity models on very low-power silicon.

[0005] Currently, no known hardware architecture offers a unified platform enabling the shared hosting of these technological types within a coherent, secure, and hardware-orchestrated framework. Existing solutions rely on disparate clusters composed of specialized accelerators, assembled in heterogeneous architectures but lacking a centralized and adaptive hardware orchestrator capable of contextualizing the activity of AI agents according to their computational and environmental profiles.

[0006] The invention described in this application precisely addresses these major limitations by proposing an integrated, autonomous, and scalable hardware device, specifically designed as a unified physical substrate for hosting multiple, interoperable, and secure artificial intelligence agents. This device aims to materialize a hardware infrastructure dedicated to the real-time coexistence of heterogeneous algorithmic models, ensuring their isolated execution, controlled interconnection, and dynamic allocation across shared and adaptive physical resources. The resulting architecture enables a technological breakthrough in the design of embedded distributed intelligence machines, providing an "AI-native" infrastructure at the hardware layer itself.

[0007] In a preferred embodiment, the hardware specifically includes a server-class, multi-processor central processing unit (CPU) that handles general logic functions, interprets inferences returned by hosted AI agents, manages internal interrupts, and provides overall control of the intra-system control flow. The CPU acts as an arbiter for inter-agent supervision operations, particularly in cases of time synchronization or adaptive redeployment following the detection of critical events (e.g., overconsumption, urgent prioritization, preemptive termination of a failing agent).

[0008] In a complementary embodiment, the device integrates at least one specialized, next-generation graphics processing unit (GPU) architected around a massively parallel model optimized for high-rank tensor processing, enabling, in particular, the accelerated execution of deep learning algorithms, generative adversarial networks (GANs), multimodal transformer architectures, and large language models (LLMs). The GPU unit is capable of leveraging Tensor Core or equivalent hardware libraries, in combination with multi-stream queue planners that ensure the asynchronous execution of computationally intensive AI agents, while retaining recontextualization capabilities via the host CPU.

[0009] In a complementary embodiment, the device also integrates at least one specialized neural processing unit (NPU), directly integrated on the same chassis or on a coherent computing module (SoC or interposer), optimized for lightweight inferences with very low latency and power consumption. This NPU enables local processing of real-time sensory streams, the execution of compact convolutional neural network topologies, and the deployment of sparse or quantized topology models, typical of embedded environments. The NPU particularly supports bio-inspired models such as spiked neural networks, whose operations are triggered by discrete electrical events, and for which a conventional GPU-type unit would be oversized and energy-inefficient.Because the hardware device directly incorporates a native hardware ALoT gateway, this gateway features specific hardware circuits ensuring direct compatibility with the main industrial protocols for real-time data acquisition and communication, including but not limited to: Ethernet TSN (Time Sensitive Networking), industrial CAN bus, Modbus TCP / RTU, LPWAN protocols such as LoRa, and the latest generation of cellular infrastructures such as 5G URLLC (Ultra-Reliable Low Latency Communications). This gateway enables the direct acquisition of signals from physical sensors, actuators, programmable logic controllers (PLCs), and distributed cyber-physical devices, without the need for external converters or software middleware.

[0010] The native ALoT gateway integrates a hardware-based, on-the-fly streaming edge preprocessing unit, including a Contextual Sensorial Stream Normalization Unit (CSSNU), which dynamically standardizes, compresses, or filters incoming data according to the specifications of the AI ​​agents subscribed to those streams. This hardware capability for sorting, enriching, and structuring data on the fly significantly reduces the load on CPU / GPU / NPU units, while simultaneously enhancing the security and confidentiality of the raw streams.

[0011] In one particular embodiment, the hardware system includes a self-contained hardware orchestrator, designed as the real-time decision-making core of the platform. This component is preferably implemented as hardware-programmable logic, typically a ruggedized FPGA (Field-Programmable Gate Array) circuit for industrial environments, or as an embedded real-time microcontroller such as an ARM Cortex-R / M or RISC-V real-time microcontroller. This hardware orchestrator is specifically configured to provide contextual, adaptive, and hierarchical allocation of internal processing resources (CPU, GPU, NPU), based on an embedded analytics engine that continuously receives indicators of load, priority, criticality, and temperature.

[0012] The orchestrator is coupled with low-latency shared memory and an internal hardware bus, preferably with photonic or optical interconnect, designed to enable direct communication between the different processing units at speeds exceeding 100 Gbit / s, with inter-process latency times of less than one microsecond. This bus supports inter-accelerator DMA communication modes, as well as hardware priority queue management, allowing simultaneous processing of instructions from different AI agents while maintaining consistency in the dynamic allocation of hardware threads.

[0013] This hardware orchestrator also includes an embedded contextual analysis module, based on a local micro-inference engine (e.g., a decision tree engine or embedded small neural network), enabling proactive anticipation of immediate or future computational resource needs. This predictive logic is fed in real time by metrics extracted from running AI agents: model call frequency, observed latency, instantaneous energy consumption, and impending thermal overload. Based on these inputs, the orchestrator can pre-distribute workloads, allocate idle CPU cores, start or stop shared GPU / NPU blocks, or temporarily disable certain less critical AI agents. In a reinforced embodiment, the hardware device is equipped with an integrated cybersecurity subsystem, operating according to a hardware-based zero-trust approach.Each AI agent hosted on the platform runs within a dedicated hardware enclave, physically separated from those of other agents by a robust isolation structure. This isolation is achieved through hardware segmentation of memory (e.g., with Intel SGX or AMD SEV-SNP modules) and hardware partitioning of the address bus, preventing any unauthorized memory or peripheral access.

[0014] The system also includes a hardware-integrated Trusted Platform Module (TPM) configured to provide continuous authentication and cryptographic attestation of the state of each active AI agent. Each AI agent launch is preceded by a hardware validation process based on a SHA-2 / 3 hash and a comparison with a signed digital signature stored in secure non-volatile memory. This verification prevents the introduction of any malicious code, even through a software or firmware update chain.

[0015] Because the device incorporates a hardware-based Secure Boot mechanism, it ensures that only verified, encrypted, and electronically signed code can be executed from the platform's boot process until the AI ​​agents are deployed in their respective enclaves. This Secure Boot relies on a root chain of trust anchored in a secure hardware component (e.g., an HSM chip or TPM module), guaranteeing that the bootloader, low-level firmware (BIOS / UEFI), and initial orchestrator modules remain unaltered.

[0016] In a preferred embodiment, the device also includes an integrated intelligent adaptive power management module, combining physical sensors, an embedded local decision engine, and dynamic redistribution logic. The system incorporates a dense network of thermal sensors and electrical current probes distributed throughout each active sub-component (CPU, GPU, NPU, VRM, DRAM, etc.), enabling real-time mapping of heat dissipation areas, power consumption peaks, and computational stress profiles.

[0017] This data is processed locally by a hardware predictive analytics module based on an embedded ACAPM (Adaptive Context-Aware Power Management) algorithm. This module cross-references real-time information with usage history and AI agent profiles to proactively determine the optimal load distribution, engage units in power-saving mode, or activate secondary cooling circuits. All of this is accomplished without software intervention, through instructions microcoded within the hardware control logic. This power management module is also capable of dynamically reconfiguring operating voltages (Dynamic Voltage and Frequency Scaling - DVFS), disabling unused cores, or migrating certain inferences to the NPUs in case of GPU overload, thus ensuring an optimal balance between processing performance, component longevity, and total energy consumption.

[0018] In an advantageous configuration, the proposed hardware offers scalable modular expansion capabilities, enabling the dynamic integration of additional resources via the Compute Express Link (CXL) hardware interface. This emerging high-performance interconnect standard allows for consistent, ultra-low-latency communication between heterogeneous computing units and shared memory components, while ensuring memory address space uniformity between internal resources and hot-swapped expansions. Because the device integrates this interface, it can dynamically aggregate additional GPU modules, high-capacity RAM blocks (such as DDR5 attached to CXL), or specialized NPU / FPGA units without rebooting or complex software reconfiguration.

[0019] The embedded system is equipped with a hardware-integrated CXL controller that automatically manages the detection, registration, and secure initialization of added resources. These integrated resources are automatically incorporated into the coherent shared memory space and made available to the hardware orchestrator. The orchestrator, in conjunction with the contextual power management module, can redirect certain heavy or critical workloads to these extensions, while simultaneously adjusting their energy allocation. This capability ensures high scalability for the platform, allowing it to adapt to increasing workloads or specialized applications without significant human intervention.In a distributed implementation, the hardware is configured to operate in clusters ("Clustered Unified Hardware Nodes - CUHN"), with each cluster consisting of several unified hardware devices interconnected via a secure network infrastructure, preferably with very low latency (e.g., a dedicated TSN Ethernet network, InfiniBand, or a secure campus optical network). This clustering allows for intelligent distribution of AI agents across the different physical nodes, taking into account real-time available resources, local thermal constraints, and the criticality levels associated with certain agents or application flows.

[0020] In such a configuration, each device acts as an autonomous node capable of hosting a subset of AI agents, while also participating in a global distributed orchestration. Load balancing, redundancy of critical agents, and adaptive relocation of certain functions in the event of a node failure are ensured by the local hardware orchestrator, in coordination with a specific inter-node exchange protocol called FNAP (Federated Neural Aggregation Protocol). This protocol enables intelligent federated orchestration, leveraging secure hardware consensus techniques to organize, synchronize, and enable the distributed units to cooperate according to a common global processing plan. FNAP operates independently of the network software stack and utilizes dedicated secure hardware channels (hardware AES 256 encryption, local PKI signature).In a reinforced embodiment, the hardware is also equipped with an internal advanced operational resilience module, designated by the acronym AORM (Advanced Operational Resilience Module). This embedded hardware module is responsible for continuously monitoring the health of the computing units (CPU, GPU, NPU) through performance, power, and temperature sensors, as well as critical load probes. When a risk of failure is detected (e.g., overheating, localized power failure, abnormal latency), the system can automatically reallocate ongoing workloads to pre-configured or newly activated backup units, preserving the state of active AI agents (stateful failover) through continuous duplication of the execution context.

[0021] This hardware resilience makes it possible to avoid any service interruption in critical environments (e.g., industrial control, medical monitoring, mobile robotics), while optimizing device lifespan through balanced hardware stress management. The AORM module can also integrate predictive migration logic, allowing certain tasks to be moved to neighboring clusters (in a CUHN mode) before a failure occurs, based on an anticipated degradation profile.

[0022] Finally, in an integrated configuration, the device's native ALoT gateway includes a specialized hardware preprocessor for intelligent structuring of raw sensor streams, designated by the acronym CSSNU (Contextual Sensorial Stream Normalization Unit). This subsystem performs online transformation of incoming signals, their statistical normalization (min-max scaling, z-score standardization), the extraction of first-order features (e.g., moving average, envelope curves, peak detection), and their distribution to data-consuming AI agents according to criteria of latency, priority, or relevance. This hardware preprocessing, performed without CPU load, ensures rapid event detection, reduces the workload on central processing units, and protects against the injection of unnecessary noise or redundancy into the embedded machine learning circuits.Thus described in all its aspects, the hardware device proposed by the present invention constitutes an innovative, comprehensive, and scalable response to the contemporary challenges of embedded artificial intelligence. It combines secure multi-agent hosting, contextual hardware orchestration, proactive energy management, dynamic hardware scalability, distributed and resilient operation, and native readiness for interactions with complex physical environments, positioning this invention as a next-generation technological building block for autonomous intelligent systems. LIST OF DIAGRAMS AND TECHNICAL DRAWINGS:

[0023] • Figure 1: General hardware architecture of the unified device.

[0024] • Figure 2: Detailed architecture of the real-time autonomous hardware orchestrator.

[0025] • Figure 3: Operation of the zero-trust cybersecurity hardware subsystem.

[0026] • Figure 4: Intelligent module for predictive contextual energy management.

[0027] • Figure 5: Modular scalable CXL interfaces for hardware expansion.

[0028] • Figure 6: Scalable operation in clusters via FNAP protocol.

[0029] • Figure 7: Scenarios for the interoperability of multi-technology AI agents within the device.

Claims

DEMANDS 1. A unified, autonomous hardware device for hosting, orchestrating, and securing intelligent agents, characterized in that it comprises, within a single, integrated hardware architecture: • At least one multi-processor central processing unit (CPU) specifically dedicated to the rapid execution of general logical functions, the control of interoperability between agents, and the real-time interpretation of the results of inferences performed by the different hosted intelligent agents; • At least one specialized graphics processing unit (GPU), utilizing a high-density parallel architecture to support intensive deep learning, massive inference, and advanced neural computing operations; • At least one specialized neural processing unit (NPU), hardware integrated, optimized to perform lightweight, contextual, ultra-low latency and very low power consumption neural inferences, particularly in the context of intelligent agents using neuromorphic models or spiking neural networks; • A native hardware ALoT gateway, directly ensuring the integration and secure preprocessing of sensor streams from industrial IoT equipment through multiple standardized protocols including Ethernet TSN, CAN, Modbus, LoRa, and industrial 5G, without the need for dedicated external intermediate equipment; • An autonomous real-time hardware orchestrator, preferably implemented via dedicated hardware programmable logic of the FPGA type or on a real-time industrial microcontroller, configured to dynamically and contextually allocate internal resources (CPU, GPU, NPU) according to the immediate and specific needs of the different intelligent agents, taking into account in particular their technological typology (generative AI, agentic AI, physical AI, robotics) and their anticipated energy requirements; • An integrated and strictly isolated hardware cybersecurity subsystem based on a zero-trust model, comprising: o Independent physical enclaves for each intelligent agent; o A secure boot module; o An integrated TPM module for continuous hardware authentication of active agents. • An integrated hardware module for predictive and adaptive energy management, configured to continuously measure internal thermal, energy and operational parameters, and to dynamically redistribute the tasks of intelligent agents according to their energy profile, maximum permissible latency, and actual thermal constraints anticipated in real time.

2. Hardware device according to claim 1, characterized in that the autonomous hardware orchestrator comprises an ultra-fast shared internal bus with photonic or optical interconnection, enabling extremely low latency internal communication and high-speed exchanges between CPU, GPU, NPU and ALoT gateway.

3. Hardware device according to claim 1, characterized in that the zero-trust cybersecurity hardware subsystem comprises secure hardware enclaves specific to each intelligent agent, physically preventing any unauthorized sharing or access to the processed data between agents or from outside.

4. Hardware device according to claim 1, characterized in that the hardware module for contextual and predictive energy management uses an embedded hardware adaptive algorithm of the type "Adaptive Context-Aware Power Management (ACAPM)", capable of proactively anticipating the energy consumption profiles of each hardware unit and ensuring optimal dynamic distribution of workloads in real time.

5. Hardware device according to claim 1, characterized in that it integrates a low-power neuromorphic coprocessor dedicated to permanent intelligent sensory monitoring tasks ("Always-On Neuromorphic Surveillance Processor - AONSP"), configured to ensure very low-power active standby and to dynamically trigger the main computing units according to the proactive detection of critical events.

6. Hardware device according to claim 1, characterized in that it comprises a hardware mechanism for modular expansion scalable via Compute Express Link (CXL) interfaces, enabling dynamic, scalable and transparent integration of complementary external hardware resources (additional GPUs, extended memories, specialized storage) directly into the shared coherent memory space of the device.

7. Hardware device according to claim 1, characterized in that it enables scalable clustered operation (“Clustered Unified Hardware Nodes - CUHN”), allowing the dynamic and secure federation of several similar hardware devices via a high-performance secure internal network, ensuring a balanced and autonomous distribution of the intelligent agents hosted between the physical nodes grouped in a cluster, in order to globally optimize the resources available at the scale of an extended industrial or operational site.

8. Hardware device according to claim 7, characterized in that the clustered federated hardware devices implement secure inter-node communication via a specific hardware protocol for federated neuronal aggregation (“Federated Neuronal Aggregation Protocol - FNAP”), enabling intelligent distributed orchestration of intelligent agents based on the actual and available capabilities of each hardware device in the cluster.

9. Hardware device according to claim 1, characterized in that it integrates an internal hardware capability for advanced operational resilience (“Advanced Operational Resilience Module - AORM”), capable of automatically and dynamically reallocating critical loads from intelligent agents to backup hardware units, thus proactively preventing any interruption or significant degradation of service in the event of a partial failure or overload detected within the main units (CPU, GPU, NPU).

10. Hardware device according to claim 1, characterized in that the native AloT gateway integrates a specific hardware preprocessor for contextual normalization of sensory streams (“Contextual Sensorial Stream Normalization Unit - CSSNU”), enabling the intelligent real-time structuring of captured data in order to optimize its immediate use by the hosted intelligent agents.