Configurable multi-mode on-chip routing method

Abstract

The invention relates to the technical field of network-on-chip and brain-like computing, in particular to a configurable multi-mode on-chip routing method, which adopts'broadcast-unicast mixed mode + boundary trigger dynamic switching ': in a local communication area, a source node sends a broadcast data packet carrying local neuron IDs (Identity), nodes in the area receive the broadcast data packet in a matched manner through a local ID table, and the local neuron IDs are sent to the source node; high fan-out efficient diffusion is ensured; and when the data packet arrives at the region boundary, the router automatically switches to a unicast mode through a built-in boundary triggering mechanism, adds a global unique synapse ID to the head of the data packet, and transmits the data packet to a remote target node based on a point-to-point path, thereby ensuring cross-region accurate low-redundancy communication. According to the method, a large-scale routing table is not needed, hardware is easy to implement, local high bandwidth and global low congestion are both considered, and the method can be widely applied to on-chip network communication of large-scale brain-like computing chips.

Description

Technical Field

[0001] This invention relates to the fields of on-chip networks and neuromorphic computing technology, and in particular to a configurable multi-mode on-chip routing method. Background Technology

[0002] With the rapid development of neuromorphic computing technology and the continuous expansion of system scale, requiring the integration of billions of neurons and synapses, on-chip networks have become a key communication foundation for connecting large-scale computing units. Their routing methods directly determine the system's communication efficiency, storage overhead, and scalability. Existing on-chip network routing methods mainly include broadcast, multicast, and unicast.

[0003] Broadcast mode is simple to implement. The source node sends data to all nodes simultaneously, and the target node determines whether to receive it based on the address information. However, the spread of data to all nodes will bring a lot of redundant traffic and seriously consume bandwidth.

[0004] Multicast mode sends data to multiple specified nodes by selectively branching and forwarding data through routers. Compared with broadcast, it can reduce redundancy, but it requires storing a large number of multicast routing tables in the router, resulting in extremely high storage and control overhead, and is no longer suitable for large-scale system applications.

[0005] Unicast mode uses precise point-to-point transmission with no redundant traffic, but in high fan-out communication scenarios, the source node needs to repeatedly send a large number of data packets, resulting in excessive communication load.

[0006] The three routing methods mentioned above all have obvious shortcomings. They cannot simultaneously meet the core requirements of low redundant communication, low storage overhead, and high scalability. They are also difficult to adapt to the hybrid communication characteristics of local dense connections and global sparse connections in neuromorphic computing. They have become a key bottleneck restricting the development of large-scale neuromorphic computing systems, and a new on-chip routing method is urgently needed to solve the above problems. Summary of the Invention

[0007] The purpose of this invention is to overcome the problems of the prior art and provide a configurable multi-mode on-chip routing method to solve the problems of existing on-chip network routing methods in neuromorphic computing scenarios, including high redundant traffic caused by global data diffusion in broadcast mode, high storage and control overhead caused by the need for a large number of routing tables in multicast mode, heavy load on source nodes caused by repeated transmission in unicast mode high fan-out scenarios, and the technical problem that existing methods cannot simultaneously satisfy high efficiency of local communication, accurate remote communication, low storage overhead and high scalability, making it difficult to support large-scale neuromorphic computing systems.

[0008] The above objectives are achieved through the following technical solutions: A configurable multi-mode on-chip routing method is provided, applicable to network-on-chip (NoC) and neuromorphic computing scenarios. This method combines broadcast and unicast modes and dynamically switches between the two modes based on a boundary triggering mechanism. Specifically, the method includes the following steps: (1) Within a preset local communication area, data packets are transmitted in broadcast mode: the source node generates a broadcast data packet carrying the local neuron ID and broadcasts it to all nodes in the local area; each node in the area quickly matches the local neuron ID through a pre-stored local ID table. If the match is successful, the data packet is received; if the match fails, the data packet is discarded. (2) In the scenario of remote communication across local areas, the routing mode is switched from broadcast mode to unicast mode: when the data packet is forwarded to the boundary node of the local area, the router of the boundary node automatically detects the cross-regional transmission requirement, appends a globally unique synapse ID to the header of the data packet, and calculates the point-to-point transmission path based on the globally unique synapse ID; (3) Remote transmission is completed through unicast mode: The router accurately transmits the data packet with the globally unique synaptic ID attached to the remote target node according to the calculated point-to-point transmission path. The target node confirms the receiving identity by parsing the globally unique synaptic ID, thus realizing accurate communication without global flooding.

[0009] As a further optimization of this method, the local neuron ID is a unique identifier of a neuron within a local communication area, used for fast matching of nodes within the area; the local ID table is a list of local neuron IDs that need to be received, pre-stored by nodes within the area, used to achieve data packet reception filtering.

[0010] As a further optimization of this method, the globally unique synapse ID is a globally unique identifier for synapses in the on-chip network, used for precise addressing of nodes across regions; the globally unique synapse ID is pre-stored in the synapse lookup table (Synapse LUT) of the border node router, providing a basis for path calculation.

[0011] As a further optimization of this method, the mode switching is automatically completed by the boundary triggering mechanism without manual intervention, ensuring the real-time performance and adaptability of the communication.

[0012] As a further optimization of this method, the implementation logic of the boundary triggering mechanism includes: The router has built-in area boundary detection logic and pre-stores the boundary identification information of the local area. When data packets are forwarded, the boundary detection logic determines in real time whether the current node is the boundary node of the local area and whether the next hop forwarding will exceed the local area range; If the cross-regional conditions are met, the router is triggered to switch from broadcast mode to unicast mode; otherwise, it continues to forward in broadcast mode.

[0013] As a further optimization of this method, the router includes a broadcast processing module, a unicast processing module, a boundary detection module, and a packet format conversion module; wherein: The broadcast processing module is used to parse the local neuron ID of the broadcast data packet and call the local ID table to complete the matching; The unicast processing module is used to parse the globally unique synapse ID and calculate the point-to-point transmission path; The packet format conversion module is used to convert the header of the broadcast data packet into a unicast header and attach a globally unique synapse ID when switching modes. The boundary detection module is used to realize regional boundary detection and mode switching triggering.

[0014] As a further optimization of this method, both broadcast and unicast data packets include a header and a payload, and the payload structure of the two types of data packets is identical; wherein: The header of the broadcast data packet contains a local neuron ID and a packet type identifier, which are used to identify the broadcast mode; The header of the unicast data packet includes a globally unique synapse ID, a globally unique chip ID, and a packet type identifier, which are used to identify the unicast mode.

[0015] As a further optimization of this method, the division of the local communication region is based on the neuron connection density of the neuron-like computing system, dividing the computing chip where the densely connected neuron clusters are located into the same local communication region.

[0016] As a further optimization of this method, the method is specifically applied to on-chip network communication of neuromorphic computing chips to realize the interconnection and communication between neurons and synapses in neuromorphic computing chips.

[0017] As a further optimization of this method, the method supports neuronal interconnection in large-scale neuromorphic computing systems, balancing high bandwidth for local communication with low congestion for global communication.

[0018] The configurable multi-mode on-chip routing method provided by this invention achieves adaptive adaptation between efficient local communication diffusion and accurate long-distance communication transmission through a hybrid broadcast-unicast mode and boundary-triggered dynamic switching. By abandoning large-scale multicast routing tables and relying solely on local ID tables and synaptic lookup tables, it significantly reduces storage and hardware overhead. By adapting to the locally dense and globally sparse communication characteristics of neuromorphic computing, it achieves a balance between high local bandwidth and low global congestion, effectively supporting large-scale neuronal interconnections. Specific beneficial effects are as follows: 1. Adaptive communication capability: The routing mode can be dynamically switched according to the communication range (local / cross-regional) - the broadcast mode is used in local areas to meet the high fan-out and efficient diffusion, and the unicast mode is used in cross-regional areas to meet the remote and accurate transmission. No manual configuration of mode parameters is required, which is suitable for the hybrid communication characteristics of "local dense connection + global sparse connection" in neuromorphic computing.

[0019] 2. Balancing low redundancy and low load: Local broadcast mode avoids duplicate transmission of unicast, cross-regional unicast mode avoids global flooding of broadcast, and at the same time abandons the large-scale routing table of multicast, achieving the dual goals of "local efficiency and remote accuracy".

[0020] 3. Low storage and low hardware overhead: No need to store multicast routing tables, only local ID tables and synapse lookup tables are required, significantly reducing storage overhead; the hardware logic of the boundary triggering mechanism is simple, the overall hardware complexity of the router is low, and it is easy to integrate on a large scale.

[0021] 4. High scalability: In large-scale neuromorphic computing systems, local communication areas can be expanded on demand, and storage overhead and hardware complexity do not increase significantly with the growth of system size. It can simultaneously take into account local high bandwidth and global low congestion, and support large-scale neuronal interconnection. Attached Figure Description

[0022] Figure 1 The diagram illustrates broadcast mode, unicast mode, and a configurable multi-mode on-chip routing method according to the present invention. It shows a comparison of the routing characteristics of broadcast, unicast, and the method of the present invention: In broadcast mode, the source chip sends data to all chips, resulting in a large amount of redundancy; in unicast mode, the source chip sends data to the target chip one by one, resulting in excessive load; the method of the present invention uses broadcast mode for local communication and switches to unicast mode for cross-regional communication, balancing low redundancy and low load, and achieving optimal overall performance. Figure 2 This diagram illustrates the packet format of the broadcast-unicast hybrid routing method in the configurable multi-mode on-chip routing method described in this invention. It includes two packet formats, both consisting of a header and a payload: the broadcast packet format header contains a local neuron ID and a packet type identifier, used for broadcast transmission within a local area; the unicast packet format header contains a global chip ID, a global synapse ID, and a packet type identifier, used for unicast transmission across regions. The payload structures of both packet formats are identical, ensuring data transmission compatibility. Figure 3This diagram illustrates the boundary triggering mechanism in a configurable multi-mode on-chip routing method according to the present invention. The left side represents local region A, the right side represents local region B, and the middle represents the boundary chip. Data packets are transmitted in broadcast mode within region A. Upon reaching the boundary chip, the boundary triggering mechanism is activated, switching to unicast mode and transmitting the data to the target chip in region B via a point-to-point path. The diagram also shows a Synapse LUT, used to store target node information corresponding to global synapse IDs, supporting unicast path calculation. Figure 4 This is a schematic diagram of the modular processing of the router in a configurable multi-mode on-chip routing method according to the present invention; it includes a broadcast processing module, a unicast processing module, and a payload processing module; the broadcast processing module is used to parse the local neuron ID of the broadcast data packet and complete the matching, the unicast processing module is used to parse the global synapse ID and calculate the unicast path, and the payload processing module is used to extract the valid data in the data packet; the modules work together to realize the dynamic switching of broadcast-unicast mode and the efficient transmission of data. Detailed Implementation

[0023] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. The described embodiments are merely some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] This solution provides a configurable multi-mode on-chip routing method. It ensures efficient distribution through local area broadcasting and accurate transmission through cross-area unicasting. Furthermore, it achieves adaptive mode switching based on a simple boundary triggering mechanism, eliminating the need for complex control logic. Figure 1 As shown, this includes local area broadcast mode, cross-region unicast mode, and boundary triggering mechanism, as detailed below: The local area broadcast mode in this embodiment is suitable for high fan-out communication within a local communication area. The core is to achieve efficient and low-redundancy transmission by using "local neuron ID + local ID table matching". The specific steps are as follows: Step (101) Local communication region division: Based on the neuron connection density of the neuron computing system, the computing chip where the densely connected neuron clusters are located is divided into the same local communication region to ensure the communication efficiency of the local broadcast mode.

[0025] Step (102) Broadcast data packet generation: When the source node generates a communication request, it generates a broadcast data packet; such as Figure 2As shown, the packet header contains the local neuron ID and packet type identifier, and the payload is the neuron communication data; among them, the local neuron ID is a unique identifier of the neuron in the local communication area, ensuring that nodes in the area can be matched and identified through this ID.

[0026] Step (103) In-region broadcasting and matching reception: The source node broadcasts data packets to all nodes in the local area through the on-chip network; such as Figure 4 The "broadcast processing module" allows routers at each node within the region to parse the local neuron ID in the packet header and quickly match it using a pre-stored local ID table. The local ID table is a list of local neuron IDs that nodes within the region need to receive. If a matching local neuron ID exists in the table, the data packet is received and the payload is extracted. If no matching entry exists, the data packet is discarded to avoid invalid data consuming resources.

[0027] The local area broadcast mode in this embodiment does not require storing a global routing table, but only uses a local ID table to complete the matching, reducing storage overhead; at the same time, local broadcast avoids repeated transmission in unicast mode, and the source node only needs to send a data packet once to cover all target nodes, which is suitable for high fan-out communication scenarios.

[0028] The cross-regional unicast mode in this embodiment is suitable for remote communication across local regions. The core is to achieve accurate and low-redundancy transmission by using "globally unique synapse ID + point-to-point path". The specific steps are as follows: Step (201) Cross-regional demand identification: When the target node of a data packet does not belong to the current local region, the data packet is forwarded to the boundary node within the local region, such as... Figure 3 The "boundary chip" in the context.

[0029] Step (202) Mode switching and unicast packet generation: as follows Figure 3 As shown, after the router at the border node identifies cross-regional needs through border detection logic, the trigger mode switches from broadcast to unicast; simultaneously, the router's packet format conversion module (such as...) Figure 4 (As shown) Convert broadcast packets into unicast packets; such as Figure 2 The "unicast packet format" appends a globally unique synapse ID and a packet type identifier to the packet header. The globally unique synapse ID is a globally unique identifier for synapses within the on-chip network, used for precise addressing of nodes across regions; for example... Figure 3 As shown, the boundary node queries the target node information corresponding to the globally unique synapse ID through the pre-stored Synapse Lookup Table (Synapse LUT), providing a basis for path calculation.

[0030] Step (203) Point-to-point unicast transmission: such as Figure 4The "unicast processing module" calculates the point-to-point path from the boundary node to the target node based on the target location information in the globally unique synapse ID. The router forwards unicast data packets according to the path, and each node along the way only forwards the data packets to the next hop node and does not spread them to irrelevant nodes. The router of the target node resolves the globally unique synapse ID, confirms that it is the target node, receives the data packets and extracts the payload to complete the remote communication.

[0031] In this embodiment, the cross-regional unicast mode transmits data across regions without global flooding, reducing redundant traffic; the point-to-point path ensures the accuracy and reliability of transmission, adapting to remote communication needs.

[0032] The cross-regional boundary triggering mechanism in this embodiment is the core of realizing "dynamic switching between broadcast and unicast". Its hardware implementation is simple and does not require complex control logic. The specific design is as follows: Core components: The router has built-in area boundary detection logic, which pre-stores the boundary identification information of the local area to determine whether data packets need to be transmitted across areas.

[0033] Detection logic: When a data packet is forwarded to any node, the boundary detection logic determines in real time whether the current node is the boundary node of the local area and whether the next hop forwarding will exceed the local area range; that is, by identifying the location of the current node and the area to which the target node belongs, it is confirmed whether the cross-regional transmission conditions are met.

[0034] Switching trigger: If it is determined to be a cross-regional transmission requirement, the boundary detection logic generates a mode switching control signal and sends it to the broadcast processing module and the unicast processing module. The matching logic of the broadcast processing module is turned off, and the path calculation and forwarding logic of the unicast processing module is activated to complete the mode switching. If it is determined to be a local transmission, the broadcast processing module remains active and the broadcast mode continues.

[0035] Hardware characteristics: The boundary triggering mechanism does not rely on complex control logic, its hardware implementation is simple, and it will not significantly increase the hardware overhead of the router.

[0036] As a specific embodiment of this solution, and in conjunction with the application scenario of large-scale neuromorphic computing systems, the specific implementation process of this method is described in detail, including: 1. System Configuration Local region partitioning: The neuromorphic computing system is divided into multiple local communication regions, each containing several computing chips, and each computing chip integrates a large number of neurons; the partitioning of local regions is based on the neuron connection density to ensure dense neuron communication within the region and to adapt to the efficient transmission of broadcast mode.

[0037] Router configuration: Each computing chip is equipped with one router as described in this invention. The router integrates a broadcast processing module, a unicast processing module, a boundary detection module, and a packet format conversion module to realize mode switching and data forwarding.

[0038] Pre-stored data: Nodes within the region pre-store a local ID table, storing the local neuron IDs that the node needs to receive; boundary nodes pre-store a synapse lookup table, storing the correspondence between globally unique synapse IDs and target nodes, as well as boundary identification information of the local region.

[0039] 2. Local broadcast communication Scenario: A source node within a local area needs to send communication data to multiple target neurons within the area.

[0040] The implementation steps are as follows: S1: The source node's router generates a broadcast data packet. The packet header contains the local neuron ID and packet type identifier, and the payload is the neuron data to be transmitted. S2: The source node broadcasts the data packet to all nodes in the local area via the on-chip network; S3: Routers at each node within the region parse the local neuron IDs through the broadcast processing module and call the local ID table for matching; S4: The node that successfully matches (the target node) receives the data packet and extracts the payload, while the node that fails to match discards the data packet, thus completing the local communication.

[0041] 3. Cross-regional unicast communication Scenario: A source node in a local area needs to send communication data to a target node in another local area.

[0042] The implementation steps are as follows: S1: Broadcast data packets generated by the source node are forwarded within the local area until they reach the boundary node; S2: The router at the boundary node identifies cross-regional needs through boundary detection logic and generates mode switching control signals; S3: The packet format conversion module converts broadcast packets into unicast packets and appends a globally unique synapse ID and packet type identifier to the packet header; S4: The unicast processing module queries the target node information based on the synaptic lookup table and calculates the point-to-point transmission path; S5: Unicast data packets are forwarded along the computation path, and are only forwarded to the next hop at each node until they reach the target node in the target area. S6: The router at the target node resolves the globally unique synapse ID, confirms it, receives the data packet, extracts the payload, and completes cross-regional communication.

[0043] 4. System performance verification Through the above implementation process, the communication performance of the large-scale neuromorphic computing system achieves the following results: (1) Local communication has no global redundancy, the load on the source node is significantly reduced, and it is suitable for high fan-out communication requirements; (2) Cross-regional communication is flooded, which greatly reduces redundant traffic and improves transmission accuracy and reliability; (3) Router storage overhead is low, no large-scale routing table is required, and hardware implementation is simple; (4) The system supports large-scale neuron interconnection, taking into account both local high bandwidth and global low congestion, to meet the needs of expanding the scale of brain-like computing.

[0044] The above description is merely illustrative of the embodiments of the present invention and is not intended to limit the present invention. For those skilled in the art, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A configurable multi-mode on-chip routing method, characterized in that, The method is applied to on-chip networks and neuromorphic computing scenarios. By combining broadcast and unicast modes and implementing dynamic switching between the two modes based on a boundary triggering mechanism, it specifically includes: Within a pre-defined local communication area, data packets are transmitted in broadcast mode: the source node generates a broadcast data packet carrying the local neuron ID and broadcasts it to all nodes within the local area; each node within the area quickly matches the local neuron ID using a pre-stored local ID table, and if the match is successful, it receives the data packet; if the match fails, it discards the data packet. In remote communication scenarios across local areas, the routing mode is triggered to switch from broadcast mode to unicast mode: when a data packet is forwarded to the boundary node of the local area, the router of the boundary node automatically detects the cross-regional transmission requirement, appends a globally unique synapse ID to the header of the data packet, and calculates the point-to-point transmission path based on the globally unique synapse ID. Remote transmission is accomplished via unicast mode: the router accurately transmits the data packet with the globally unique synaptic ID attached to it to the remote target node according to the calculated point-to-point transmission path, and the target node confirms the recipient's identity by parsing the globally unique synaptic ID.

2. The configurable multi-mode on-chip routing method according to claim 1, characterized in that, The local neuron ID is a unique identifier for neurons within a local communication area, used for rapid matching of nodes within the area; the local ID table is a list of local neuron IDs that nodes within the area need to receive, used to filter and receive data packets.

3. The configurable multi-mode on-chip routing method according to claim 1, characterized in that, The globally unique synapse ID is a globally unique identifier for synapses in the on-chip network, used for precise addressing of nodes across regions; the globally unique synapse ID is pre-stored in the synapse lookup table of the border node router, providing a basis for path calculation.

4. The configurable multi-mode on-chip routing method according to claim 1, characterized in that, The mode switching is automatically completed by the boundary triggering mechanism without manual intervention, ensuring the real-time performance and adaptability of the communication.

5. A configurable multi-mode on-chip routing method according to claim 1 or 4, characterized in that, The implementation logic of the boundary triggering mechanism includes: The router has built-in area boundary detection logic and pre-stores the boundary identification information of the local area. When data packets are forwarded, the boundary detection logic determines in real time whether the current node is the boundary node of the local area and whether the next hop forwarding will exceed the local area range; If the cross-regional conditions are met, the router is triggered to switch from broadcast mode to unicast mode; otherwise, it continues to forward in broadcast mode.

6. The configurable multi-mode on-chip routing method according to claim 1, characterized in that, The router includes a broadcast processing module, a unicast processing module, a boundary detection module, and a packet format conversion module; wherein: The broadcast processing module is used to parse the local neuron ID of the broadcast data packet and call the local ID table to complete the matching; The unicast processing module is used to parse the globally unique synapse ID and calculate the point-to-point transmission path; The packet format conversion module is used to convert the header of the broadcast data packet into a unicast header and attach a globally unique synapse ID when switching modes. The boundary detection module is used to realize regional boundary detection and mode switching triggering.

7. A configurable multi-mode on-chip routing method according to claim 1, characterized in that, Both broadcast and unicast data packets include a header and a payload, and the payload structure of the two types of data packets is identical; wherein: The header of the broadcast data packet contains a local neuron ID and a packet type identifier, which are used to identify the broadcast mode; The header of the unicast data packet includes a globally unique synapse ID, a globally unique chip ID, and a packet type identifier, which are used to identify the unicast mode.

8. The configurable multi-mode on-chip routing method according to claim 1, characterized in that, The division of the local communication region is based on the neuron connection density of the neuron-like computing system, dividing the computing chip containing the densely connected neuron clusters into the same local communication region.

9. A configurable multi-mode on-chip routing method according to claim 1, characterized in that, The method is specifically applied to on-chip network communication of neuromorphic computing chips, and is used to realize the interconnection and communication between neurons and synapses in neuromorphic computing chips.

10. A configurable multi-mode on-chip routing method according to claim 1, characterized in that, The method supports neuronal interconnection in large-scale neuromorphic computing systems, balancing high bandwidth for local communication with low congestion for global communication.

Images