3results about How to "Increase resource consumption" patented technology

This invention belongs to the field of time measurement technology and discloses a resource-saving short-chain time-to-digital converter and its conversion method. It utilizes a delay unit and conversion circuit to build a local oscillator, which flips the output state at fixed time intervals during a continuous high-level input signal. The time at which the local oscillator causes the level to flip is known. The time of the last flip is obtained using TDL and TDC, and the start and last flip times are obtained by recording the number of flips of the local oscillator. The resource-saving short-chain time-to-digital converter provided by this invention has good linearity, completely avoiding nonlinearity problems caused by inconsistent line lengths introduced across multiple resource blocks. The resource-saving short-chain time-to-digital converter has a simple and flexible structure, low resource consumption, and high utilization rate. It has good robustness and can be directly ported between different channels or even devices, and can correct temperature-induced drift online.

The invention relates to the technical field of artificial intelligence, and discloses a multi-language barrier-free conference room and an implementation method thereof, and the method comprises the steps that an application server, an AI model server and a plurality of user terminals cooperatively execute the method, and the method comprises the following steps: the application server creates a conference room instance and generates a corresponding access link, the user terminals are connected to the application servers through the access links and send target language information selected by the user terminals to the application servers; the application server receives the audio data stream from the user terminal in real time and splits the audio data stream into continuous audio stream segments; the application server sends the audio stream segment to an AI model server; the AI model server carries out voice recognition on the received audio stream segment to generate a corresponding source language text segment; and the AI model server translates the source language text fragments into translated text fragments of multiple target languages in parallel through one-time model reasoning based on a target language list provided by the application server.

This invention discloses a method for processing matrix multiplication weight data in model inference, adapted to large-scale matrix multiplication in BFP format for large models in the Transformer architecture. Addressing the pain points of existing architectures regarding BFP exponent and mantissa splitting, bandwidth, and resource consumption, this method employs (64,64) weight blocks, dual address generator 1:4 scheduling, waterline flow control, and double buffering techniques to achieve full-bandwidth read / write between DDR, BRAM, and computational units. Through multi-core parallelism and multi-round multiplication and accumulation, performance is improved by more than 3 times compared to FP16, reducing memory usage and design complexity while ensuring inference accuracy, providing an efficient hardware acceleration solution for LLM inference in resource-constrained environments.