Industrial Technology Research Institute

Tay-Jyi Lin

Chih-Wei Liu

Po-Han Huang

Chein-Wei Jen

Pi-Cheng Hsiao

I-Tao Liao

Shih-Chieh Chang

Shyh-Shyuan Sheu

Yuan-Hua Chu

Chih-Sheng Lin

Gin-Kou Ma

Fu-Cheng Tsai

Chi-Hung Lin

Wei-Sheng Huang

Chan-Hao Chang

Yi-Ching Kuo

Pi-Chen Hsiao

Chou-Kun Lin

Chie-Min Chao

A data feature augmentation system and method for a low-precision neural network are provided. The data feature augmentation system includes a first time difference unit. The first time difference unit includes a first sample-and-hold circuit and a subtractor. The first sample-and-hold circuit is used for receiving an input signal and obtaining a first signal according to the input signal. The first signal is related to a first leakage rate of the first sample-and-hold circuit and the first signal is the signal generated by delaying the input signal by one time unit. The subtractor is used for performing subtraction on the input signal and the first signal to obtain a time difference signal. The input signal and the time difference signal are inputted to the low-precision neural network.

Data feature augmentation system and method for low-precision neural network

A performance scaling device, a processor having the same, and a performance scaling method thereof are provided. The performance scaling device includes an adaptive voltage scaling unit, a latency prediction unit, and a variable-latency datapath. The adaptive voltage scaling unit generates a plurality of operation voltages and transmits the operation voltages to the variable-latency datapath. The variable-latency datapath operates with different latencies according to the operation voltages and generates an operation latency. The latency prediction unit receives the operation latency and a system latency tolerance and generates a voltage scaling signal for the adaptive voltage scaling unit according to the operation latency and the system latency tolerance. The adaptive voltage scaling unit outputs and scales the operation voltages thereof according to the voltage scaling signal.

Performance scaling device, processor having the same, and performance scaling method thereof

An embodiment of a processing device includes a function unit and a control unit. The function unit receives input data and performs a specific operation to the input data to generate result data. The control unit receives the result data and generates an output signal. The control unit latches the result data according to a first clock signal to generate first data and latches the result data according to a second clock signal to generate second data. The control unit compares the first data with the second data to generate a control signal and selects the first data or the second data to serve as data of the output signal according to the control signal. The second clock signal is delayed from the first clock signal by a predefined time period.

Processing device for determining whether to output a first data using a first clock signal or a second data using delay from the first clock signal according to a control signal

A method and corresponding apparatus for compiling high-level languages into specific processor architectures are provided. In this embodiment, the specific processor is encapsulated in a virtual processor interface with simple instruction set architecture, and a compiler translates application programs into corresponding assembly codes. Further, the difficulty of the compiler design is reduced.

Method and corresponding apparatus for compiling high-level languages into specific processor architectures

A pipelined datapath with dynamically reconfigurable pipeline stages is provided, having a pipeline controller which generates clock signals and selects signals based on a system clock and a valid data signal to control each of the registers and each of the multiplexers in the pipeline circuit. In other words, when a valid datum is being processed, the pipeline register is activated to latch the output of the combinational logic circuit; otherwise, when an invalid datum is received, the register is not activated and the datum bypasses the register through a multiplexer. Therefore, the pipeline stages of the pipelined datapath are dynamically reconfigured to save the power dissipation effectively.

Dynamically reconfigurable stages pipelined datapath with data valid signal controlled multiplexer

An inter-cluster communication module using the memory access network is provided, including a plurality of clusters, a memory subsystem, a controller and a switch device. When some clusters issue a load instruction and some clusters issue a store instruction of an identical memory address concurrently, the controller controls the switch device which connects the clusters and the memory banks of the memory subsystem, so that the data item is transmitted from the cluster issuing the store instruction to the cluster issuing the load instruction through the switch device, thereby achieving data exchange between the clusters. Herein, the data item is selectively stored in the memory module depending on the address. Furthermore, the data item is also transmitted between the memory and the clusters over the switch device.

Inter-cluster communication module using the memory access network

Growing community of inventors

Hsinchu, Taiwan

Tay-Jyi Lin