A chess game of hundreds of billions: why is the video processing chip VPU so important?

Recently, ChatGPT has triggered a global AI fever, and with it, AI computing and AI chips closely related to AI algorithms have once again become popular. But in fact, the growth of computing power that big language models can drive is limited. According to relevant data, 70%-80% of AI computing tasks in the future will occur in the field of machine vision and multimodality. Therefore, the real AI era is actually the era of video computing.

The continuous popularity of live broadcast and short video, and the demand for camera and video capabilities in various equipment forms and industry scenarios, make dedicated video computing capabilities and processing units become the rigid needs of the times.

We are, and will live for a long time, in a world where there is no screen and no video everywhere. People's demand for high-definition, high-frame rate video has always increased, and the CPU + GPU system in the classical computing structure can no longer meet the processing needs of massive video services, so the video processing chip VPU (Video Processing Unit) came into being.

What exactly can a VPU do? What is the stage of its application direction and market structure? Let's open this little chip and enter the exquisite chess game of VPU together.

Observation: Why do you need a VPU?

At the beginning of this chess game, we need to understand what the inevitability of the emergence of the VPU is.

At this stage, whether it is C-end or B-end users, there is a clearer and stronger demand for high-definition video (4K/8K). As ordinary users, we look forward to the continuous upgrading of high-definition experience in live broadcast, short video, and VR, while enterprises need to improve high-definition video capabilities in online meetings, industrial vision, autonomous driving, and other fields. Overall, we are at a critical juncture towards 4K ubiquitous and 8K landing applications.

However, the sprint to high-definition video will bring higher computing power consumption and network costs in video encoding and decoding. Especially for computing power, when massive video codec requirements appear, traditional CPU computing methods lack pertinence, processing efficiency is low, and waste energy consumption and deployment space will be brought about, and the comprehensive cost is too high. The GPU has a better ability to process video codecs, but its resource utilization and floating flexibility are still insufficient. In particular, the use of GPUs to handle codec tasks often leads to problems where drivers are incompatible with processors, resulting in a large number of tasks not running properly.

Professional chips do professional things, is the core proposition of the development of the semiconductor industry in this era. Since video tasks make the CPU or GPU a little difficult to process, the VPU codec is more flexible and the processing power is more professional.

What the VPU is aiming for is actually a very big opportunity.

Drop: Application and test of VPU

As far as the current situation is concerned, VPU has rich application fields and related scenarios. This is because video is becoming a ubiquitous form of basic information that pervades every corner of life and all areas of industry. In the cloud, short video and live streaming platforms handle massive video encoding and decoding tasks every day, while on the device side, various devices such as cameras, drones, mobile phones, and PCs have the functions of shooting and playing videos.

This also makes the VPU fall mainly in two areas. On the one hand, it becomes part of the data center in the cloud to cope with the accelerated demand for massive video data; On the other hand, on the terminal side, it is combined with mobile phones, automobiles, security, drones, and AR\VR equipment to improve the video encoding and decoding performance of the device and reduce the power consumption and delay required for video processing. Today, with the explosive growth of video processing demand, the industry has also put forward a series of new requirements for VPUs, which we can summarize into three capabilities:

1. HD encoding and decoding capabilities: VPUs need to cope with 4K or even 8K codecs to meet the landing requirements of ubiquitous HD. Nowadays, online meetings, industrial quality inspection, security and other applications are also upgrading 4K, and the ubiquitous generalization of 4K has become a common and extensive demand for video processing capabilities from all walks of life.

2. Software environment compatibility: VPU should be able to flexibly adapt to different operating systems and software environments, so as to ensure the software and hardware compatibility of video applications and avoid resource allocation and compatibility problems caused by software problems.

3. Hardware scenario adaptability: VPUs need to flexibly adapt to different hardware features to reduce end-to-end latency and improve codec processing efficiency, so as to meet the needs of different hardware systems for VPUs.

Under the framework of two major fields and three capabilities, the chess game of VPU has become clearer. This chess game has indeed attracted players from all walks of life to join.

Competition: A chess game of hundreds of billions of dollars

At this stage, VPU's industrial practice and market development are in a period of rapid explosion, and Internet companies and technology companies in various fields have joined the layout of this key technology. Among them, the more representative is that in April 2021, Google released its self-developed Argos VCU. After that, enterprises with advantages in the fields of streaming media, short video, and live broadcast have begun to deploy VPUs.

For example, Tencent's self-developed video transcoding chip "Canghai" has been successfully tape-out in March 2022. Internet giants such as Meta, ByteDance, and Kuaishou have launched VPU layouts. In particular, it is worth noting that in the past, there was not much investment in semiconductors, but ByteDance and Kuaishou, which are highly dependent on video services, have laid out VPUs, which shows that the industry attaches great importance to VPU technology. Some data believe that VPU will grow into a market of 100 billion yuan in the future, which is very likely at present.

However, the layout of VPUs by Internet and streaming media manufacturers is still a cross-border advance layout, and its effect will take several years to appear. At this stage, the industry needs more mature and stable VPU products to support.

For example, Arm Technology previously released the second-generation VPU product "Linglong" series V6/V8. These two products correspond to the mainstream 4K\8K real-time decoding needs in the market. It has the characteristics of flexible and customizable configuration, excellent encoding and decoding performance, and small area. In terms of hardware scene adaptation, Linglong V6/V8 can not only face data center scenarios, but also be combined with cameras, automobiles, smart homes, and mobile devices. In the software compatibility issue, it has fully implemented Android and Linux drivers, as well as some upper application framework support, which can meet the needs of various consumer products, cloud and embedded scenarios.

In general, the VPU market application is in the stage of giant entry, high-speed outbreak, and extensive industrial application, which basically indicates the potential value of the VPU track in the future.

In the future, high-definition video and machine vision will be ubiquitous, and VPU may also be the same.