GM's Ultra Cruise with LiDAR and 5-nanometer chip revealed

Consumer Reports, an authoritative American automotive review media, once compared the SuperCruise of General Motors Cadillac CT6 with several other intelligent driving systems including Tesla FSD. Super Cruise surpassed Tesla and received the most praise. Super Cruise adds lidar scanning maps, high-precision positioning modules and high-precision satellite positioning services compared to Tesla FSD. This is the key factor for Super Cruise to take the first place. This month, General Motors announced Ultra Cruise, an advanced version of Super Cruise. UltraCruise will be officially launched in 2023 and used in luxury cars, while mid- and low-end cars will continue to use Super Cruise. The upgrade mainly adds lidar and uses Qualcomm's 5nm chip SA8295 as the main processor.

Image source: Internet

Ultra Cruise can only be used in North America, that is, the route shown in the picture above. The route is the original 130,000 miles. The blue line is newly added and includes part of the Canadian route.

Image source: Internet

This is a spy photo of Ultra Cruise. The lidar is located behind the windshield and is very small and almost invisible. The system on the roof is a verification test system, and the mass-produced car will not be any different from a traditional vehicle in appearance.

Ultra Cruise's lidar supplier is Cepton. Founded in Silicon Valley in 2016, Cepton (also known as "Cepton Technology"), a high-performance MMT (Micro Motion Technology) lidar solution company, announced in August this year that it had signed a final corporate merger agreement with GrowthCapital, a publicly traded special acquisition purpose company, and a private placement-related subscription agreement totaling US$58.5 million. After the transaction is completed, the merged company will be renamed "Cepton, Inc" and listed on the Nasdaq trading market with a new stock code of "CPTN". This means that Cepton, which was founded five years ago, has completed its goal of landing in the secondary market through a "backdoor listing". Unlike any other lidar company, Cepton is positioned as an L2/L3 lidar, not an L4. Cepton's investors include Nvidia and Japan's KOITO (Koito Manufacturing), the world's largest supplier of headlights. Cepton is a Chinese company with CEO Pei Jun and most of the team members from Velodyne.

Image source: Internet

KOITO invested 50 million US dollars in Cepton, which is extremely rare for conservative Japanese companies. Japanese small and medium-sized enterprises almost never invest in venture companies. KOITO cooperated with Cepton to create a headlight and lidar integrated device, putting lidar into the headlight. In addition to the GM project, Cepton also has two platforms. In 2024, eight cars will use Cepton's lidar.

Not only in the automotive field, Cepton has also explored the field of urban infrastructure services with Qualcomm. On October 18, 2021, Cepton joined the Qualcomm® Smart Cities Accelerator Program. Through the Qualcomm IoT Service Suite, Cepton and its key partner in smart space, The Indoor Lab, plan to work with Qualcomm to provide "smart venues as a service" using a lidar-based crowd analysis system (which can both protect privacy and optimize space utilization). Cepton worked with TheIndoor Lab to deploy a lidar-based crowd analysis solution in a terminal at Orlando International Airport in 2020. As a pilot project, the solution can provide anonymous footstep tracking data to help maintenance personnel carry out targeted cleaning work while helping travelers avoid crowded areas.

Image source: Internet

Cepton uses a voice coil drive solution, and the principle is exactly the same as that of a speaker. The patent explanation diagram is as above. The laser radar includes a fixed transmitting lens 130 and a receiving lens 140. And a laser source 110a~110b is basically set on the focal plane behind the transmitting lens 130. The laser source 110a is used to emit a laser pulse 120. The laser pulse is collimated by the transmitting lens 130 and directed to the object 150 in front of the LiDAR sensor. Then, the laser pulse 122 reflected from the object is directed to the receiving lens 140. It is focused on the corresponding photodetectors 160a~160b through the receiving lens 140, and the photodetectors 160a~160b are basically set on the focal plane of the receiving lens 140. Finally, the flight time (ToF) of the laser pulse 120 from emission to detection is determined by the processor 190 coupled to the laser source 110a and the photodetector 160a, and then the distance between the LiDAR sensor and the object is calculated.

Image source: Internet

As shown in the above diagram of the key scanner patent, four laser sources 110a~110d and four photodetectors 160a~160d are mounted on the same rigid platform 230. The rigid platform 230 is connected to the first substrate 210 through two flexure elements 220a and 220b. The flexure elements 220a and 220b can be deflected to the left or right by using a single actuator (such as a voice coil 250 and a permanent magnet 260, or by a piezoelectric actuator, etc.). The first substrate 210 can be connected to the second substrate 212 by two flexure elements 270a and 270b. The flexure elements 270a and 270b can be deflected forward or backward by using a single actuator (such as a voice coil 252 and a permanent magnet 262, or by a piezoelectric actuator, etc.). Therefore, through the left and right movement of the flexure elements 220a and 220b, and the forward and backward movement of the flexure elements 270a and 270b, the laser sources 110a~110d and the photodetectors 160a~160d can perform two-dimensional scanning in the focal planes of the transmitting lens 130 and the receiving lens 140, respectively. Voice coil laser radar technology meets the three elements of large-scale commercialization of laser radar: high performance, low cost, and high reliability. The light source and detector are not static, but are placed on the electromagnet moving coil. After providing current, they can be used to scan an overall image or environment. The light source and detector are not single, they have multiple pathways and multi-channel arrays. This unique imaging method has two major advantages: first, it draws on relatively mature speaker technology. Unlike motors and bearings, it does not produce friction and component loss; second, the light source and detector are both directly in and out of the lens, without mirrors, to avoid attenuation and polarization. The most important point is low cost. The supply price to GM is about US$500-700.

Image source: Internet

Voice coil LiDAR naturally has disadvantages. The first is that the FOV is relatively narrow because the voice coil has limited space for movement. Anyone who has played with a speaker knows that once a certain power limit is exceeded, the voice coil movement is nonlinear, that is, distortion. The advertising slogans of speakers are all large voice coils and long strokes. Another is that the angular resolution is slightly low. The voice coil is not a motor, and it is difficult to achieve many levels of control accuracy. These are not considered disadvantages when used in L2/L3. Voice coil LiDAR is very similar to MEMS LiDAR, except that voice coils are used instead of MEMS. The disadvantages are that the cost is high, there are moving parts, the light cannot go in and out directly, and the signal-to-noise ratio is slightly low. The advantages are that the FOV can be wider and the angular resolution can be higher.

UltraCruise will use Qualcomm's fourth-generation automotive cockpit chip, the 5-nanometer SA8295

   Image source: Internet

Qualcomm currently has only one 5nm chip, the Snapdragon 888, model SM8350, and the improved version Snapdragon 888 Plus, model SM8350-AC. The difference between the two is that the 888 Plus's super-large core Cortex X1 runs at a higher frequency of 2.995GHz, and has stronger AI computing power, reaching 32TOPS, while the 888 is 26TOPS. If GM wants to launch Ultra Cruise in 2023, it must get a sample of the 5nm chip in 2021, and can only use the automotive version of the SM8350 chip, namely SA8295.

The shipment volume of high-end cars is almost negligible compared to that of mobile phones. It is too low. The one-time cost of a 5nm chip is about $300 million, and Qualcomm does not need to develop chips separately for the automotive field where the shipment volume is negligible.

The Snapdragon 888 uses an octa-core design, with one Cortex-X1 super core, running at 2.84GHz. The car version should be a little lower, estimated to be 2.5GHz, three A78 large cores, running at 2.4GHz, a little lower than the car version, estimated to be 2.1GHz, and four A55 high-efficiency cores, running at 1.8GHz. It is estimated that the car version will also be reduced by another 100MHz.

According to ARM's internal data, taking the middle value, the computing power of A76 is 11.55 DMIPS/MHz. A75 also takes the middle value, which is 8.85 DMIPS/MHz. ARM has not announced the computing power of A77 and A78. Based on the performance improvement of about 20% in each generation, it is estimated that the computing power of A78 is 16.632 DMIPS/MHz. Considering that ARM performance is not only about computing power, it is estimated that the computing power of A78 is most likely 15-16DMIPS/MHz, taking the middle value of 15.5DMIPS/MHz. CortexX1 is 22% stronger than A78, that is, 18.91DMIPS/MHz. Compared with Samsung MediaTek, Qualcomm's CPU is not strong, but its GPU and NPU have absolute advantages. I believe Ultra Cruise is enough to challenge Tesla's FSD 2.0 (ie HW4.0).