Neural Command Processor for Low-Latency Task Scheduling

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing neural processing units face inefficiencies in data transmission and computational speed due to repeated communication with host systems, limiting their effectiveness in high-workload deep learning and inference tasks.

Innovation Solution

A command processor and neural processing system that minimizes data transmission to host systems by using doorbells to differentiate between workload data types, allowing for parallel processing and efficient task management without redundant buffer access, and incorporating a workload manager, context manager, and task descriptor generation for optimized neural processor utilization.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Speed

If the neural processing unit repeatedly communicates with the host system for data transmission, then the data can be processed, but the computing speed and latency are reduced

Engineering Contradiction:
Improvecomputing speedVSAvoidlatency
Core Design Contradiction:
SpeedVSLoss of time

Solution Approach 1:

The patent implements a command buffer that pre-loads and stores multiple commands before execution. The workload manager prepares task descriptors and command buffers in advance, allowing the neural processing unit to execute sequences of operations without repeatedly pausing for host system communication. This preliminary preparation of data and commands reduces latency and accelerates computing by eliminating repeated communication overhead.

Inventive Principle:
Principle #10Preliminary action

2Productivity

If the workload manager accesses command buffers for every doorbell interrupt, then data can be retrieved, but processing efficiency is reduced due to redundant access

Engineering Contradiction:
Improveprocessing efficiencyVSAvoidbuffer access time
Core Design Contradiction:
ProductivityVSLoss of time

Solution Approach 1:

The patent implements a mechanism where the workload manager checks a doorbell register to determine whether new workload data has arrived. When no new data is present, the manager continues processing existing task descriptors without accessing the command buffer again. This continuous processing approach eliminates redundant buffer access operations, maintaining high productivity by keeping the processing pipeline continuously active rather than repeatedly stalling for buffer reads.

Inventive Principle:
Principle #20Continuity of useful action

3Device complexity

If the neural processing unit uses centralized command processing, then task management is simplified, but parallel processing capability is reduced

Engineering Contradiction:
Improvetask management complexityVSAvoidparallel processing capability
Core Design Contradiction:
Device complexityVSProductivity

Solution Approach 1:

The patent segments the command processing function into multiple command managers (first command manager, second command manager, etc.), each capable of independently processing different task descriptors. This segmentation allows parallel processing of multiple workloads simultaneously while maintaining simplified individual manager logic. Each command manager handles its own queue of tasks, eliminating the bottleneck of centralized sequential processing and improving overall system productivity through parallel execution.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentUS20250217212A1Command processor, neural processing system and method for transmitting data thereof
Publication Date: 2025.07.03 REBELLIONS INC
  • US20250217212A1 patent drawing
  • US20250217212A1 patent drawing
  • US20250217212A1 patent drawing

AI summary

An apparatus comprising neural processors, a command processor, and a shared memory. The command processor receives a context start signal indicating a start of a context of a neural network model from a host system. The command processor determines whether neural network model data is entirely or partially updated based on the context start signal. The command processor updates the neural network model data in the shared memory based on a determination on whether neural network model data is entirely or partially updated based on the context start signal. The command processor generates a plurality of task descriptors describing neural network model tasks based on the neural network model data, and distributes the plurality of task descriptors to the neural processors.