Qualcomm autonomous driving platform Snapdragon Ride analysis

On December 30, 2020, Great Wall Motors held a conference on upgrading its smart driving strategy and officially released the brand-new “331 Strategy” for smart coffee driving. According to the plan, Great Wall will be equipped with the advanced intelligent driving function NOH (Navigation On Highwaypilot, Chinese name “high-speed automatic pilot assisted driving”) on the Wey brand in early 2021. At the meeting, Great Wall also reached a strategic partnership with Huawei and Qualcomm respectively, and decided to carry Huawei’s MDC platform and Qualcomm Snapdragon Ride platform on mass-produced vehicles.

Great Wall Motor will launch the world’s first L4-level mass-produced vehicle based on Qualcomm’s Snapdragon Ride platform in 2022. Equipped with IBEO’s 4D all-semiconductor true solid-state LiDAR, that is, Flash LiDAR, the longest effective distance is 300 meters. The global shutter is used for one-time imaging at short distances, and the zoom vertical scanning is used at long distances. The core of Qualcomm Snapdragon Ride platform consists of two chips, SM8540 and SA 9000B accelerator. The AI ​​computing power of the single board is 360TOPS, which is estimated to be INT8 precision. The overall power consumption is 65 watts and 5.5TOPS/W. It can be increased to 4 sets of computing platforms through PCIe switches. The total AI computing power of the four accelerators reaches 1440TOPS.

The technology in the field of autonomous driving is changing rapidly, and currently only Toyota’s new hydrogen-powered Mirai and Lexus’ new LS 500h are actually on the market, equipped with the TAD system. The starting price of the former is 8.6 million yen (approximately RMB 515,000), and the starting price of the latter is 17.94 million yen (approximately RMB 1.075 million). Only 100 of Honda’s L3s were built, and they were only leased and not sold. Audi’s A8 L3 features are not enabled, and Tesla is still standard L2. L3/L4 requires not only the intelligence of the car, but also the cooperation of high-precision maps and V2X infrastructure. The sales volume of L3/L4 cars will not exceed 500,000 in 5 years, and will not exceed 3 million in 10 years. The market is not large. , the competition is fierce. L3/L4 car sales are insignificant compared to mobile phones with hundreds of millions of sales, and are also insignificant compared to game consoles with tens of millions of sales. There is no doubt that a single development of a L3/L4 automotive chip will be a loss, because L3/ L4-level autonomous driving chips must use at least a 7-nanometer process, which will lead to an astonishing development cost. A single tape-out fee for 7-nanometer excluding design cost is about 30 million US dollars, and the tape-out fee for 12-16 nanometer FinFET is much lower, which is 5 million US dollars. to $8 million. In addition, there are hard expenditures for purchasing various IPs, such as microkernel architectures such as ARM A78, and IP of automotive-grade ASILs, which cost tens of millions of dollars. The largest design fee has not yet been calculated, so if someone develops chips for the L3/L4 autonomous vehicle market alone, it is destined to suffer huge losses, at least for more than 7 consecutive years. Both Qualcomm and Nvidia are well aware of this. Nvidia has considered the multi-purpose use of Orin from the very beginning, while Qualcomm, from the perspective of the cockpit, has changed its mobile phone chips to vehicle-grade grades for reuse. Qualcomm has not disclosed the information of SM8540 plus SA 9000B, but it can be speculated from other chips that it is extremely unlikely that Qualcomm will develop new chips for L3/L4 autonomous driving alone. After all, mobile phones are the main source of profit for Qualcomm. So we make a bold guess.

Qualcomm will officially commercialize an AI 100 edge computing kit in the first half of 2021, using the Snapdragon 865 as the application processor and the AI ​​100 as the accelerator. Under the M.2 edge interface, the computing power is 70TOPS, and under the PCIe interface 16 cores , the computing power can reach 400TOPS.

According to the promotional pictures of Great Wall, 8540 and 9000 are both 7nm, AI 100 and Snapdragon 865 are also 7nm, PCIe can also be seen on Great Wall’s promotional pictures. Of course, for the sake of car regulations, a little performance must be sacrificed, and power consumption is reduced by frequency reduction to achieve car regulations, so the performance is reduced to 360TOPS. The Snapdragon 865 is the top of Qualcomm’s 7nm chips. The frequency of 870 is higher, up to 3.2GHz, and the power consumption is bound to be higher. Therefore, the 8540 is most likely to be the car version chip of the Snapdragon 865. Of course, the Modem of the X55 can be used. remove. Qualcomm has only one accelerator, and the SA9000B is likely to be the car version of the AI ​​100.

The AI ​​100 has a maximum of 16 cores. Qualcomm has always taken the DSP route, and the number of cores is not high. The SRAM of each AI core is 9MB, 16 is 144MB, the Tesla FSD is 64MB, basically the AI ​​100 is twice that of the Tesla. The Qualcomm kit uses 12GB of LPDDR5, and the Tesla FSD can only correspond to LPDDR4.

The software environment of Tesla AI100 supports autonomous driving applications.

Of course, Qualcomm will not only provide hardware, Qualcomm will provide a complete set of solutions, including tools and simulation.

The partners of Qualcomm’s autonomous driving platform especially pointed out the visual perception and driving strategy software stack Arriver. In fact, Arriver is the software brand of Veoneer. Acquires Veoneer for about $5 billion. Parking is mainly Valeo, and Park4u is the name of Valeo’s parking system. The main partner in DMS is Seeing Machines, the supplier of Cadillac DMS.

Based on lidar detection and trajectory tracking, it seems that 360-degree bird’s-eye lidar is still used.

Qualcomm’s future self-driving car Electronic architecture plan has also been thought out for the car factory.

Others include sensor calibration, deep learning datasets, big data management, and Qualcomm provides solutions.

Car companies obediently surrender their souls, Qualcomm has thought of everything for you.

The Links:   G150XTN030 6MBI100S-120