Once Nvidia Orin "has something to do", who can replace the domestic self-driving chip?

After the accident of high-end GPU, "who can replace NVIDIA Orin" has become a problem that must be paid attention to now.

Nvidia Orin, is it really that critical?

At present, NVIDIA Orin is indeed unique in terms of technological advancement, performance indicators, and mass production delivery capabilities.

In terms of performance, Orin uses a 7-nanometer process and consists of an Ampere architecture GPU, an ARM Hercules CPU, a second-generation deep learning accelerator DLA, a second-generation vision accelerator PVA, a video codec, and a wide dynamic range ISP.

At the same time, a vehicle-grade safety island design was introduced.

Orin supports 204GB/s memory bandwidth and up to 64GB DRAM. The high-speed I/O interface is compatible with the interface of the previous generation Xavier SoC. It can achieve INT8 computing power of 275TOPS, 7 times that of Xavier, and consumes 55W.

From a technical point of view, the hardware configuration used by Orin is the biggest difference from the previous generation in the introduction of Tensor Core, which supports sparse computing (170 TOPS), which is a fine-grained computing structure that can double throughput and reduce memory usage. .

In addition, in terms of deep learning acceleration, memory and communication, CPU performance and other indicators, it is 1-2 times higher than the previous generation.

In terms of software, NVIDIA provides a software development kit SDK (Software Development Kit).

Mainly the board level support package (BSP), including the bootloader Bootloader, Linux kernel, driver driver, tool chain Tool chain and Ubuntu-based reference file system, BSP also supports various security functions (secure boot, trusted execution environment) , disk and memory encryption, etc.).

On top of BSP, there are multiple user-level libraries for accelerating applications, including Deep Learning Acceleration Library (CUDA, CuDNN, Tensor RT), Accelerated Computing Library (cuBLAS, cuFTT), Computer Vision and Image Processing Library (VPI) , multimedia and camera libraries (libArgus and v4l2).

Enterprises that have the strength to develop themselves may not necessarily use the development tools provided by NVIDIA, but this shows the highly open and customized features of NVIDIA Orin.

At the mass production delivery level, NVIDIA Orin has begun to deliver stably to OEMs and autonomous driving companies.

Currently listed models equipped with Orin chips, including Weilai ET7, ES7 (equipped with 4 pieces of NVIDIA Orin), and subsequent models including almost all models aimed at high-level assisted driving.

Such as Xiaopeng G9, Jidu, Weimar, Zhiji, Ideal L9 and so on.

Autonomous driving companies, especially players who focus on L4, are almost inseparable from the underlying computing platform supported by NVIDIA Orin, whether it is test development or commercial implementation.

Therefore, the importance of NVIDIA Orin is almost impossible to replace at present:

Among all mass-produced autonomous driving chips, Orin has the highest single-chip computing power, the most advanced technology, and the fastest mass production pace.

The surface is the strongest, and car companies are rushing to grab it.

And what should Orin do if something happens?

Who can replace Nvidia Orin

In addition to its deep accumulation of technical strength in the GPU field, Nvidia dominates the autonomous driving chip market. Another main reason is that it entered the game early.

In 2015, Nvidia launched the first chip PX for autonomous driving.

But the latecomers are not without opportunities. In fact, some alternatives to Nvidia have gradually surfaced.

possible alternative, but not entirely possible

What I am talking about here mainly refers to two other foreign manufacturers: Qualcomm and Mobileye.

Among them, Mobileye belongs to the old predecessors of autonomous driving. Its software and hardware integrated autonomous driving system was once the only choice for mass production of passenger cars.

At present, the EyeQ 5 chip, which is widely mass-produced, has a computing power of 24TOPS, which is at the same level as NVIDIA's previous generation Xavier.

The largest user in China is Geely Automobile. Almost all of its models with intelligent assisted driving function use Mobileye solutions.

But Mobileye's limitations are also here. The "black box" model that does not open data rights to OEMs, and the "bundled sale" that requires a set of software and hardware to make it narrower and narrower.

It was abandoned by Weilai, Ideal and other manufacturers successively.

Although the next-generation EyeQ 6 chip is expected to match or even surpass Nvidia in terms of technical indicators, the Mobileye solution can only be said to be an alternative, but not the best choice.

Qualcomm, which has dominated smart cockpit chips, has produced products that surpass Nvidia as soon as it makes a move in the field of autonomous driving.

SnapDragon Ride chip, 7nm process, 360TOPS computing power under INT8 precision, overall power consumption 65w.

The performance surpasses that of Nvidia Orin, and it has been mass-produced.

The latest version of Great Wall's Wei brand Mocha DHT-PHEV, the first mass-produced Qualcomm SnapDragon Ride, will be delivered by the end of the year. According to the official announcement, it can realize the auxiliary function of city navigation beyond ordinary L2.

Qualcomm is of course an option to replace Nvidia, but as a foreign manufacturer, Qualcomm faces the same risks as Nvidia.

Self-replacement, who can choose

The loudest is the domestic manufacturer Horizon, and the upcoming mass-produced Journey 5.

Horizon Journey 5 is based on TSMC's 16nm process, and the AI computing power can reach 128TOPS.

In terms of core architecture, the CPU part of the Horizon Journey 5 chip adopts an 8-core ARM Cortex A55, and the AI computing unit adopts a dual-core Horizon Bayesian architecture BPU (Brain Processor Unit).

At the same time, the Journey 5 chip also has 2 ISP cores, computer vision engine, 2 DSP cores, and video coding and decoding units.

In terms of mass production progress, Journey 5 has been delivered to OEMs for development and testing, and the official mass production time is set in 2023.

Beyond the horizon, Huawei is another important player.

MDC 810, with a computing power of 400TOPS, has been mass-produced. The MDC 810 is not equipped with a GPU that does not support general computing, but uses Ascend, an AI chip with a "specific domain architecture", to be responsible for computing.

BAIC Polar Fox αS Hi version, Changan Avita 11, and GAC's upcoming new cars will all be equipped with Huawei MDC 810.

The other two domestic autonomous driving chips with promising mass production are Huashan No. 2 and Jiazhixin V9.

They are from domestic start-up companies Black Sesame Intelligence and Xinchi Technology.

Black Sesame's Huashan No. 2 A1000 is on its way to mass production. The computing power of a single chip with INT8 precision reaches 58TOPS. It will be mass-produced for the first time on the new Sihao model of Jiangqi Group.

Xinchi Technology puts the mass production target of autonomous driving chips directly in the stage for high-end intelligent driving in the future.

In the second half of this year, Xinchi will launch a dedicated chip for autonomous driving with a computing power of more than 200TOPS.

The mass production plan, according to the general laws of the semiconductor industry, will not be earlier than 2024.

Alternative pros and cons

According to the latest "Autonomous Driving Chip Industry Research Report (2022)" analysis by CITIC Securities, these different player solutions that can replace NVIDIA Orin have obvious advantages and disadvantages in terms of performance and mass production pace.

Horizon Journey 5:

The biggest advantage is localized R&D and service, as well as the most advanced performance parameters in China.

The same goes for the downside: still not as good as the Nvidia Orin.

Black Sesame Intelligence, CITIC believes in the research report that its A1000's maximum computing power is still far behind that of Nvidia. In the computing platform scheme, four A1000s are used to achieve a computing power of about 250T, which is similar to that of an NVIDIA single board, and there is also a generation difference.

But the advantage is also that mass production is imminent. If it goes well, the new models of Jiangqi will be launched and delivered by the end of this year.

As for Huawei, the real advantage is not how much computing power and how fast mass production is, but the system engineering capabilities for smart cars.

Including chip, algorithm, cloud, V2X, operating system and so on.

So this is why in the MDC solution, Huawei does not use a general-purpose GPU, but chooses the Ascend chip that is closely linked to Huawei's overall strategic layout.

CITIC Securities believes that Huawei has the ability to integrate the resources of major giants in the process of smart car and autonomous driving research and development, greatly speeding up the landing speed of real smart cars, and may soon exceed the horizon.