In the traditional X86, GPU and other chip areas, domestic companies have fallen behind, but in aI chips, the domestic semiconductor industry is expected to catch up, Alibaba’s PingTouge last year launched a performance comparable to the GPU AI chip – Hanguang 800. Light from the “LieZi” article has a “the ancient three swords” chapter, meaning not exposed, light but not bright, symbol of light 800 invisible but strong calculation.
Hanguang 800 is an AI chip, more emphasis on reasoning. According to reports, a Hanguang 800 calculation force is equivalent to 10 GPUs, currently based on the hanguang 800 AI cloud service has been launched in Alibaba Cloud.
Official data show that the specific score of the Hanguag 800 on the chip test standard platform Resnet 50 is: performance 78563 IPS, is the second place (15012) 5 times, energy efficiency than 500 IPS/W, is the second place (150) 3.3 times.
Simply put, AI chips such as the Hanguang 800 are to compete with GPU-accelerated AI chips, which are more energy efficient and perform better because of the more shifting functions.
At the ISSCC 2020 conference, the editor of the Wikichip website saw the hanguang 800 chip and accelerator displayed by Alibaba’s Pingtouge, as follows:
You can see that these acceleration cards have AliNPU Vision logos, and apart from not having a video output interface, they look similar to the graphics card — as is the usual form of the GPU chip’s AI accelerator, which does not require video output.
After the radiator is disassembled, you can see the light-bearing 800 chip itself and the PCB.
Hanguang 800 chip is using TSMC 12nm process production, the core area of up to 709mm2, equivalent to the highest end of the NVIDIA GPU chip, the current GV100 core core area is only 815mm2, so Hanguang 800 is still a considerable AI chip.
In addition, from the PCB can be seen, the accelerator used 16 phase power supply, higher than the high-end graphics card power supply Ali has not announced TDP and other indicators, from the power supply alone, the power consumption is estimated to be 250W level, after all, the performance is also ten times the GPU.
On the right you can see the two power supply interfaces, which are a combination of 8 plus 4pin, which is a bit strange, unlike the 6 plus 8pin interface commonly used in graphics cards.