According to foreign media reports, Synopsys has launched the latest generation of embedded visual processor, equipped with deep neural network (DNN) accelerator, providing the industry-leading computing performance for AI intensive edge applications – 35 trillion operations per second (TOPS).
Based on arcv2risc instruction set architecture, the new designware arc ev7x visual processor adopts 1, 2 or 4-core heterogeneous architecture, integrates vector DSP, vector FPU and neural network accelerator, and can realize all kinds of intelligent automobile and consumer applications integrated with AES encryption. The optional DNN accelerator can be extended from 880mac to 14080mac, which enables the system to provide up to 35tops performance in the 16 nm fin field effect transistor (FinFET) process technology under typical conditions, four times the performance of the previous generation arc ev6x processor.
Gordon Cooper, product marketing director of EV processor of new thinking technology, said that ev7x processor is the optimized version of ev6x processor, but it’s not only a fourfold increase in MAC. Increasing MAC is just a simple part of enhancing the graphics performance of CNN (convolutional neural network). The real key is to improve the memory bandwidth required for external memory access, so as to minimize its power consumption. The processing bandwidth of the new processor can be extended to 100tops, so the DRAM interface with lower cost can be used.
In addition, advanced image mapping tools are needed to divide CNN graphics on the increased MAC. Compared with the EV6x processor, the frame throughput per second of the EV7x processor is increased by 65%.
The heterogeneous multi-core architecture of designware arc ev7x visual processor includes up to four high-performance VPUs. Each ev7x Vpu consists of a 32-bit scalar unit and a 512 bit wide vector DSP, which can be configured as 8-bit, 16 bit or 32-bit operations for multiplication accumulation on different data streams at the same time. According to Xinsi technology, the DNN accelerator adopts a special architecture, which can access memory faster, perform higher and perform better than other neural network IP. In addition to supporting CNN, the DNN accelerator also supports batch LSTM (short and long term memory), which is used for applications that need to be based on time series, such as predicting the position of pedestrians according to the observed path and speed.
In addition, the visual engine works in parallel with the DNN accelerator, making EV7x particularly efficient in the application of autopilot and ADAS, because such applications require multiple cameras and visual algorithms to run simultaneously. The new ev7x processor also attaches great importance to security. The data transmission from the chip memory to the vision engine and DNN accelerator will be protected by the optional aes-xts encryption engine. The engine can prevent the use of training data sets and personal biological data such as face recognition and retinal scanning.
The embedded arc ev7x visual processor combines high performance visual engine, DNN accelerator and programming tools, which can be applied to various visual applications.