As a large-scale multi-modal model with human performance, GPT-4 is regarded as an important step towards AGI, which marks the deep change of innovation paradigm and the redefinition of productivity, and will also bring more product migration.
Up to now, more than/kloc-0.0 million start-ups around the world claim to use this secret weapon to create new products, which will completely change almost all fields from law to stock trading, from games to medical diagnosis.
Although many of them are marketing bubbles, like all technological breakthroughs, there will always be hype cycles and unexpected long-term effects.
In fact, on the other hand, it is also very lively to enter the smart car field in 2023.
Intelligentization has become the biggest focus of Shanghai Auto Show. In addition to the single-point breakthrough of key sensors such as lidar, the giants also demonstrated the full product matrix of intelligent driving, and the mass production of urban scene-assisted driving accelerated.
What's more noteworthy is that computer hot words such as BEV, big model and supercomputer center are quickly arranged and combined with intelligent driving focuses such as automatic driving, parking and parking integration, and urban NOA, which have the flavor of mutual blending and two-way travel.
Behind this, on the one hand, the continuous upgrading of intelligent driving and intelligent cabin in recent years has continuously put forward higher requirements for data, algorithms and computing power of AI in automobile scenes. On the other hand, AGI's major breakthrough has also extended its reach to smart cars, which are regarded as an important scene to realize closed-loop applications. Many companies have made quite high-profile layouts.
A few days ago, Shangtang Technology Shangtang Technology held a technical exchange day to share the strategic layout of "big model+big computing power" to promote the development of AGI, and announced the "SenseNova" big model system under this strategy.
With the blessing of "big model+great computing power", Shang Tang's unique integrated product system of driving, cabin and cloud was fully unveiled at this Shanghai Auto Show, with nearly 30 cooperative production models on display. Shang Tang once again shared AGI's new thinking of landing in the era of smart cars.
Some amazing cooperative models unveiled at this year's Shanghai Auto Show are on display.
0 1, algorithm: AI officially entered the era of big models.
As Wang Xiaogang, co-founder and chief scientist of Shangtang Technology and president of Jueying Smart Car Group, said, "AGI has given birth to a new research paradigm, that is, based on a powerful multi-modal basic model, by strengthening learning and human feedback, the new capabilities of the basic model are continuously unlocked, thus solving massive open tasks more efficiently. 」
The general large model is not born for autonomous driving, nor is it designed to meet the specific task requirements of autonomous driving. However, many new requirements of intelligent driving development have promoted the rapid evolution of the algorithm from a special small model to a general large model.
The first is the urgent need to deal with massive data processing and corner cases.
Algorithm development needs to face a large amount of data for small targets that appear at low frequency but are vital to the perception system and their potential security risks. It will be difficult for the traditional AI small model to handle a large amount of data and high-complexity tasks at the same time. The general large model can be used in the preliminary screening process of long tail targets, and the processing of superimposed corpus words has achieved good results.
Another example is the demand of intelligent driving algorithm development for automatic data labeling and reducing labor costs. Compared with manual labeling, the general large model will automatically label massive data, which greatly reduces the time cost of labeling data acquisition and its own money cost, thus shortening the research and development cycle and improving the cost performance.
For similar reasons, in recent years, giant enterprises at home and abroad have launched their own intelligent driving layouts around the big model.
After Google put forward the application of Transformer structure in CV image classification on 20 17, the large model has been proved to be powerful in GPT-2, GPT-3, BERT and so on. Tesla took the lead in launching the platform of Transformers model in image vision.
Domestic enterprises are also close behind:
Millie Zhixing has announced that the autonomous driving cognitive model has been officially upgraded to DriveGPT. Baidu said that it will use the big model to improve the perception of autonomous driving and apply the big model to data mining. Huawei has also announced its participation in the big model competition, and its self-developed Pangu will be launched soon.
As a leading artificial intelligence company in the industry, Shang Tang braved the wind and waves in the field of large models. In the past year or two, it has fully applied the capabilities of the big model in more than 20 scenarios in various business lines, including intelligent driving.
Behind Shang Tang's "SenseNova" big model system is the deep accumulation in the research and development of big models. Shang Tang has its own full-stack large-scale model research and development system, including basic training for large-scale models and various system optimizations during implementation.
For example, OmniObject3D, a multimodal data set recently released by Shang Tang for real perception, reconstruction and generation, contains 190 categories and 6000 objects, and the data quality is very high.
For another example, in 20 19, Shang Tang released a large-scale visualization model with parameters of 10 billion for the first time. By 2022, the scale of parameters has reached 32 billion, which is the largest visualization model in the world so far.
In addition, Shang Tang continues to demonstrate its large-scale model capability in the field of intelligent driving. The BEV sensor algorithm developed by 202 1 won the Waymo Challenge with absolute advantage. The transformer structure of BEV Former 202 1 is still the most influential BEV work in the industry. UniAD developed this year is the industry's first end-to-end autopilot solution that integrates sensing and decision-making.
The other end of technical strength is the progress of mass production. Shang Tang also gave his own formula for mass production of intelligent driving:
Autopilot technical capability = scene data x data acquisition efficiency x data utilization efficiency? = scene data x data acquisition efficiency x advanced algorithm x advanced computing power.
The advanced algorithm model will not only improve the driving scene data resources through cross-industry data aggregation, but also improve the data acquisition efficiency through data closed-loop development mode and automatic data labeling, and will greatly improve the perception accuracy and perception richness, thus doubling the data utilization efficiency.
Relying on the original AI algorithm and model accumulation, Shang Tang's leading BEV sensing algorithm promotes the first batch of mass production applications in China, and adopts domain adaptation algorithm to effectively solve the cross-domain generalization problem. Shang Tang's self-driving GOP perception system has reduced the labor cost of target data collection by 94%, realized low-cost vehicle-side model development, and has been put into mass production application.
Important infrastructure in the era of smart cars.
With the evolution of electronic and electrical architecture technology from distributed to centralized, large computing chips have become the physical basis for realizing new electronic and electrical architecture.
In recent years, the computing power of car-side chips has advanced by leaps and bounds. For example, the computing power of Atlan single chip in NVIDIA planning exceeds 1000TOPS, and that of THOR single chip exceeds 2,000tops, which will greatly enhance the perception and decision-making ability of bicycles.
In the cloud, the generalized application of AGI in automatic driving, network connection and other scenes will require vehicles with higher computing power than exponential level—from data labeling to model training, from scene simulation to algorithm iteration.
Computing power will be the new infrastructure in the era of smart cars.
In this context, in recent years, mainstream enterprises have started the exploration of two-line parallelism, independently developing computing platforms on the vehicle side and establishing supercomputing centers in the cloud. After entering the era of large-scale model, with the introduction of multi-mode, the amount of data will also increase on a large scale, which will inevitably lead to a sharp increase in the demand for computing power by AGI.
It can be seen that NVIDIA's vehicle-side cloud synchronization layout will provide an end-to-end full-stack AI accelerated computing solution, and Tesla also released Dojo, a self-developed cloud supercomputer center, as early as August 20021.
According to recent reports, elon musk will also set up an artificial intelligence company to compete with OpenAI. It has purchased thousands of NVIDIA GPUs and has been recruiting artificial intelligence researchers and engineers.
At home, Geely, Weilai, Tesla, Mo Hao Zhixing, Tucki and other enterprises have also followed suit to lay out cloud computing clusters, and invested huge sums of money to enhance the computing power reserve of Zhijia.
For Shang Tang, if the big model will be the superstructure supporting intelligent driving, then the large computing power is the digital pedestal.
Xu Li, chairman and CEO of Shangtang Technology, said that at present, the demand for basic computing power and infrastructure of large models is very strong, and the demand for parallel efficiency of basic computing power is also very high, but the infrastructure that is really easy to use is actually very scarce.
For this reason, it took Shang Tang five years to build SenseCore, an industry-leading AI equipment, and completed the deployment of 27,000 GPUs, realizing the computing power output capability of 5.0 exa FLOPS. It is one of the largest intelligent computing platforms in Asia at present, which can support the simultaneous training of 20 super-large models and hundreds of billions of parameters.
AiDC Artificial Intelligence Computing Center, located in Lingang, Shanghai, will provide computing support for closed-loop data storage, marking, desensitization, simulation training, algorithm iteration and deployment of smart cars, open up the whole process of data-driven algorithm production, accelerate the production and continuous iteration of advanced intelligent driving technology AI model, and promote mass production.
On the basis of AIDC, AI Big Device will also provide a series of services to support the production of large models:
Processing the automatic data labeling required by large models will improve the efficiency of intelligent labeling by a hundred times; The deployment of large model reasoning improves the reasoning efficiency by more than 100%; Large-scale model parallel training, single cluster up to 4000 cards parallel, dense model with parameters exceeding 500 billion, trillion-level parameters; Incremental training of large model reduces the incremental fine-tuning cost by 90%; Open source models and large-scale models train developer tools to improve development efficiency on a large scale. Computing facilities of this scale are unmatched even by Tesla in the same period, which will certainly promote the efficient closed loop of large models.
03, "big model+big computing power" promotes the whole process of smart car industry.
The automobile industry is facing a once-in-a-century great change. Although Shang Tang put forward the strategic layout of "big model+big computing power" to promote the development of AGI, in fact, this concept has long been recognized at the industry level.
Based on the three core capabilities of perception, decision-making and control, and AI cloud, Shang Tang's "big model+big computing power" has realized the production of the trinity product system of driving, cabin and cloud;
In addition to the full-stack capability in the field of intelligent driving and the mass production solution of parking navigation, "big model+big computing power" is also helping Shang Tang to build a cross-scene ecology of intelligent cockpit.
During the auto show, the Jueying Future showcase, which is deeply integrated with Shang Tang's "SenseNova" vehicle system, was upgraded and unveiled. We also boarded the language model "Shang Tang Consultation" and the AIGC platform "Shang Tang SenseMirage", and reconstructed the interaction between people and vehicles through multi-point integration to create the third space.
Taking "consultation" as an example, as a natural language processing model with hundreds of billions of parameters, it uses a lot of data for training and fully considers the Chinese context, showing excellent multi-round dialogue and long text understanding ability.
Shang Tang also demonstrated a number of innovative applications of automobile scenes supported by the language big model, such as automatically extracting key information as an "email assistant" during driving and automatically generating meeting minutes as a "meeting assistant", which greatly saved users' time and energy in handling work during driving and brought rich imagination space for future travel application scenarios.
In addition, with the development, production and application of artificial intelligence model as the core, one-stop
This article comes from the author of car number, the heart of the car, and the copyright belongs to the author. Please contact the author in any form. The content only represents the author's point of view and has nothing to do with the car reform.