Industrial Technology Research Institute

Jing-Jia Liou

Juin-Ming Lu

Chih-Tsun Huang

Yao-Hua Chen

Chun-Chen Chen

Che-Wei Hsu

Chun-Hung Lai

Ting-Shuo Hsu

Jih-Sheng Shen

Yu-Xiang Yen

Wei-Shiang Lin

Wan-Shan Hsieh

A processing element architecture adapted to a convolution comprises a plurality of processing elements and a delayed queue circuit. The plurality of processing elements includes a first processing element and a second processing element, wherein the first processing element and the second processing element perform the convolution according to a shared datum at least. The delayed queue circuit connects to the first processing element and connects to the second processing element. The delayed queue circuit receives the shared datum sent by the first processing element, and sends the shared datum to the second processing element after receiving the shared datum and waiting for a time interval.

Processing elements array that includes delay queues between processing elements to hold shared data

An adaptive learning power modeling method includes: sampling at least one of a plurality of network components to form a power consumption evaluation network according to at least one parameter within a parameter range; evaluating a predictive power consumption of a to-be-measured circuit by the power consumption evaluation network; training and evaluating an actual power consumption and the predictive power consumption of the to-be-measured circuit by the power consumption evaluation network to obtain an evaluation result; and performing training according to the evaluation result to determine whether to change the power consumption evaluation network.

Hardware structure aware adaptive learning based power modeling method and system

A DNN hardware accelerator and an operation method of the DNN hardware accelerator are provided. The DNN hardware accelerator includes: a network distributor for receiving an input data and distributing respective bandwidth of a plurality of data types of a target data amount based on a plurality of bandwidth ratios of the target data amount; and a processing element array coupled to the network distributor, for communicating data of the data types of the target data amount between the network distributor based on the distributed bandwidth of the data types.

Deep neural networks (DNN) hardware accelerator and operation method thereof

A memory transaction-level modeling method and a memory transaction-level modeling system are provided. The memory transaction-level modeling method is used for simulating the operation of outputting at least one command to the memory. The memory includes a plurality of banks each of which corresponds with a bank status table. The memory transaction-level modeling method includes the following steps: An event is received. Whether one of the bank status tables is needed to be updated is determined. If one of the bank status tables is needed to be updated, this bank status table is recovered according to a TMP queue. A command is outputted to the memory according to a command queue. The outputted command is stored in the TMP queue. Some of the bank status tables are updated and others of the bank status tables are kept unchanged.

Memory transaction-level modeling method and system

A NoC timing power estimating method includes: estimating a plurality of transmission timing of a plurality of transmission units of at least a packet, the transmission timing indicating respective time points at which the transmission units enter/leave a plurality of passing elements of the NoC; based on the transmission timing of the transmission units, estimating respective circuit states and respective power states of the passing elements of the NoC, the circuit state indicating an operation state of the passing element and the power state being related to the circuit state; and based on the power states of the passing elements of the NoC, estimating power consumption of the NoC.

Growing community of inventors

Hsinchu, Taiwan

Jing-Jia Liou