Sample shipment of extremely low power (XLP) LSI chip and the application of the chip to AI


May 1, 2017
Semiconductor Energy Laboratory Co., Ltd.



  Semiconductor Energy Laboratory Co., Ltd. (SEL, Atsugi, Japan, President Shunpei Yamazaki) and United Microelectronics Corporation (UMC, Taiwan)*1 plan to start sample shipment of crystalline oxide-semiconductor large-scale integrated chips (OS-LSI), which we have been developing, in autumn 2017.
  A 60-nm-node prototyped chip realizes extremely low-power consumption (XLP*2) by efficient power-off operation, and aims at "consumed power lower by three orders of magnitude" than that of existing silicon LSI (Si LSI).
  The crystalline OS-LSI with extremely low off-state current characteristics and favorable compatibility with Si LSI can be widely applied to an artificial intelligence (AI) chip and a microcontroller unit (MCU), a field programmable gate array (FPGA), and a memory (a memory capable of both digital and analog operations), as well as digital operation necessary for the era of Internet of Things (IoT) and big data.


1. OS-LSI
  The most significant features of the crystalline OS-FETs are extremely low off-state current of 70 yA (yoctoampere, yocto is a prefix denoting 10-24) at 85°C and high on/off ratio (16 digits or higher) compared with that of Si FET (6 to 8 digits). Owing to these features, the refresh rate of the memory using OS-FETs can be reduced to several times per hour, or even per year, whereas that of a conventional dynamic random access memory (DRAM) is once in less than several milliseconds. "Normally-off" CPU that shuts off the power when a circuit does not operate is also provided (FIG. 2), thereby leading to extremely low power CPU.
  Aiming at the practical use of crystalline OS-LSI chips, we adopt a 3D hybrid process where a crystalline OS FET is stacked on a Si FET (illustrated in FIG. 1). It is confirmed that the prototyped processor fabricated by this process achieved low power consumption (FIG. 2).



FIG. 1 Cross-sectional micrograph of crystalline OS-LSI*3



FIG. 2 Power consumption*3


2. AI (Artificial Intelligence)
The crystalline OS with its extremely low off-state current can be applied to an OS memory that can store a number of data (levels) (multi-level OS memory). The OS memory can also be used as an analog memory.
We suggest the effectiveness of OS memories to AI.
Neural networks resembling human brains are widely used for AI. In the neural networks, analog data is processed by a mutiply-accumulate operation of a weight coefficient (coupling factor or multiplier) and input data (multiplicand), and large-scale parallel arithmethic is needed.
In the case of arithmethic by digital circuits, accumulation in a memory of data as a result of the arithmethic and numerous circuits are necessary; thus access to the memory is problematic to the operation speed.
Mixed analog-digital processing is expected to bring an efficient AI, but so far there has been no ideal memory that meets the demand of size and accuracy.
SEL succeeded in fabricating a crystalline OS-LSI using a high-accuracy analog memory, which enables analog arithmetic in neural network. Therefore, it is possible to incorporate an AI function into a chip with scaled-down circuitry and low power consumption.

The OS memory using the crystalline OS-FETs has an accuracy of 6 bits or more necessary for neural network, i.e., can store data of 64 levels or more. Moreover it is a four-terminal device and easy to control, unlike two-terminal devices such as a magnetoresistive random-access memory (MRAM) and a ferroelectric RAM (FRAM).

Our prototyped OS memory functions as an arithmetic device integrating a local memory and a multiplier circuit, and constitutes an efficient multiply-accumulate circuit (FIG. 3). This multiply-accumulate circuit enables reductions in power consumed and area occupied by the local memory, unlimited rewrite cycles (easy learning), and parallel arithmetic. In addition, digital circuits and analog circuits are easily formed together to be mixed in the same process, thereby easily mounting an analog memory constituting neural network, a logic (controller), and an analog-digital interface. Therefore, a mixed AI chip can be readily fabricated.



FIG. 3 Analog multiplier (OS memory corresponds to an arithmetic device integrating a local memory storing a weight coefficient and a multiplication circuit, and the sum of current corresponds to addition)
*1: UMC is a leading global semiconductor foundry headquartered in Hsinchu, Taiwan. Source: www.umc.com (the second-largest foundry in the world).
*2: XLP: eXtremely Low Power
*3: S.H. Wu et al., Symp. VLSI Technology, 58 (2016).