Detailed explanation of "computing power" of Black Sesame Intelligent Open Core Classroom Intelligent Driving Chip

Introduction: From the beginning of the birth of automobiles, the main evaluation system has been based on horsepower. With the development of the automobile industry for a century, the attributes of automobiles have slowly transformed from production tools to the third space. More and more new functions are coming to the car, requiring the support of more and more chips. With the development of electronic and electrical architecture, chips have also developed from the previous single function to the current high-performance SoC (System on Chip) stage. The previous "Core" class introduced it from the perspective of tool chain and application. Today we will introduce the basis for realizing the above functions-computing power.

Definition and classification of computing power

Computing power, simply put, is the ability to calculate. An SoC generally includes computing units such as CPU, GPU, NPU and DSP. Similar to the learning ability of a top student in Chinese, painting, mathematics and comprehensive science. Simply put, the CPU mainly processes logical calculations and its unit is DMIPS. The GPU focuses on image processing and its unit is FLOPS, which represents floating point computing power. As a neural network processor, NPU is mainly used for the reasoning work of neural networks. Currently, the computing power of mainstream NPU chips with medium and large computing power can reach TOPS level, which means 1 trillion operations per second. As a more flexible computing unit, DSP can provide both fixed-point and floating-point computing capabilities. Therefore, everyone can feel that as an excellent SoC chip, it cannot be partial to one subject and must have good results in multiple subjects. So what kind of functions can computing power support? How much computing power do different levels of intelligent driving require?

Computing power requirements corresponding to different levels of autonomous driving

Autonomous driving is mainly divided into levels 0-5, with Level 3 as the dividing line. The following is intelligent driving, and the above is automatic driving. At present, the more mature and mass-produced products on the market are mainly Level 3 and below. The current main perception of intelligent driving relies more on vision, which also promotes the convolutional neural network accelerator, NPU, to occupy a relatively important position in the field of intelligent driving.

Why are car companies scrambling for computing power?

We can see that as the level of autonomous driving increases, its requirements for computing power are also constantly increasing. Computing power has also replaced horsepower as an important parameter in the automotive industry. The automobile industry has officially entered the era of computing power, and high-computing power chips have begun to become an important selling point for new models of major car companies.

On the one hand, while the number and types of vehicle-mounted sensors are rapidly increasing, their accuracy is also greatly improved to ensure the implementation of more and more complex functions, which places higher requirements on the computing power of the chip. On the other hand, new algorithm models are constantly being updated, and more large models are being applied to actual solutions, further increasing the demand for computing power. However, we have also discovered an interesting phenomenon. Currently, the computing power of flagship models of major car companies is getting higher and higher, often hundreds or thousands of TOPS. At the same time, we have also seen the investment of chip companies in chips with medium computing power. It is also getting bigger and bigger, and the penetration rate of intelligent driving in mid- to low-end models is also further increasing. While chip computing power is being increased, sensor reuse and computing power reuse are gradually being applied to more mass production projects. Car companies and chip manufacturers seem to have reached a certain consensus, and chip computing power is also increasing. Return to reason.

The higher the computing power of autonomous driving, the better

While the computing power of chips is soaring, the market is also returning to rationality. With the development of software and algorithms, the computing power utilization of chips is also further improved. At the same time, we cannot ignore another very important concept, which is the frame rate. The frame rate can more truly reflect the actual computing power of the chip. Therefore, computing power as a theoretical value can reflect the basic computing power of the chip, while the frame rate can truly reflect the actual computing efficiency of a chip. Does that mean that the higher the frame rate, the better? It can only be said that in actual intelligent driving, it is half right. We do need a higher frame rate, provided that the algorithm model corresponding to this frame rate can truly be used by intelligent driving. Serve. Without a given algorithm model and its corresponding resolution, simply proposing the frame rate cannot provide true feedback on the actual computing power of a chip.

Black Sesame Smart Huashan Series Chip Solution

The Huashan series of smart driving chips launched by Black Sesame Intelligence not only strives for computing power, technology, and investment, but also strives for innovation. Its single chip has a computing power of 58 Tops and has the characteristics of platformization, high performance, and low power consumption. Based on Two independently controllable core IPs have built core competitive advantages, including car-grade image processing ISP and car-grade deep neural network accelerator NPU, allowing vehicles to "see clearly" and "understand". Integrating forward + surround-view cameras, front + angle radar and ultrasonic sensors, it is the first domestic single SoC parking and parking integrated domain controller chip platform that complies with vehicle regulations and reaches mass production status. The overall solution is simple, efficient and cost-effective. Rich sensor interfaces can support L2+/L3 level autonomous driving solutions, helping users improve their smart driving experience while also providing car manufacturers and parts suppliers with the possibility of significant cost reductions.