At the 31st Hot Chips conference, Huawei, along with leading chipmakers represented by AMD, Intel, and ARM, revealed the details of the upcoming 32-core DaVinci chip. It is manufactured using a 7-nanometer process technology and actively interacts with neural networks.
Back in October last year, Huawei introduced the Ascend 310 and Ascend 910 chips, designed specifically for the Internet of Things, unmanned cars and smartphones. The main feature of the platforms was the support of deep learning technologies. Announced in June, the Kirin 810 mobile processor was the company's first chip based on DaVinci microarchitecture.
The goal of DaVinci was to accelerate the work of artificial intelligence technologies with the ability to scale them. Huawei has been engaged in the new microarchitecture for 11 months and now plans to release about 100 million gadgets with these processors.
“We want the chips used in modern devices to be designed for both compact gadgets and large data centers, being a universal solution,” said Henry Liao, Research and Development Manager for Huawei.
Huawei’s flagship solution today is a 32-core chip, whose crystal area is 182 square millimeters, and the peak power is 256 teraflops when calculating 16-bit floating-point operations. The consumption of such a processor is 350 watts. A server built by the company with eight chips and 1.5 TB of DRAM received the power of 6000 watts. Server computing speed is 512 petaflops.
Huawei plans to increase the energy efficiency of the new generation of chips in two ways. Firstly, due to the increase in DRAM-memory. Secondly, by increasing the density of the tracks of the chips themselves. During testing, Huawei's solution bypassed the graphics accelerators from NVIDIA Tesla V100 and Google TPU v3 in terms of performance.