Recently, the 2023 Kunpeng Ascend Developer Summit was held in Songshan Lake, Dongguan, with the theme of "creating the future and enjoying the extraordinary". The Kunpeng Ascend Developer Summit is a technical event for developers in the ICT field. It aims to create a platform for ecological partners and developers to learn and communicate, to help developers gain a deeper understanding of Kunpeng and Ascend full-stack technologies, and to accelerate industry technologies, products and solutions. Program innovation.
Impulse Online, as a certified partner of Kunpeng and Ascend, was invited to participate in the 2023 Kunpeng Ascend Developer Summit, and Zhou Yueqian, partner and product general manager of Impulse, gave a special session on Kunpeng BoostKit confidential computing solutions entitled "AI Native Privacy Computing Platform Empowerment" "Sharing, Openness and Prosperity of AIGC", and completed the signing ceremony of the product solution of the trusted AIGC all-in-one machine created by Impulse and Kunpeng on the spot, aiming to jointly promote the development of the computing industry and build an open and win-win ecology.
According to the "2023 China AIGC Industry Development Research Report" released by iiMedia Data Center, it is pointed out that in the next five years, the market capacity of China's AIGC will continue to grow on a large scale, and it is expected to reach a market size of 270 billion by 2028. The report also pointed out that In 2023, the ranking of industries empowered by China's AIGC, in addition to popular fields such as virtual humans, live broadcast e-commerce, and games, also has great room for promoting the development of traditional industries, such as finance and government affairs. A wide range of AIGC models, including text generation, image generation, video generation, virtual human construction, etc. The above models can create paintings or copywriting based on user-specified attributes through simple text input.
The three core elements supporting the operation of AIGC are algorithms, data and computing power. Algorithms provide computing logic, data is the fuel for model training and reasoning, and computing power is the engine that starts the model. The use of an AIGC model usually requires large-scale data to support training and tuning, as well as computing resources that match its performance requirements. Therefore, accurate model algorithms, high-quality data sources, and safe and reliable computing resources The power market is the basic condition for the operation of the AIGC model.
Combining the three core elements of AIGC mentioned above, there are also related pain points when building AIGC application products.
01
data plane
When large enterprises use the AIGC model, they need to use their own internal data to fine-tune the model, and also need to use external data to expand the size and diversity of the training set, thereby improving the accuracy of the model. In the process of internal and external data circulation and use, it is difficult to confirm the right of data, and there are also compliance risks in the process of data acquisition.
02
Model level
Enterprises have a high threshold for building their own AIGC models, and need to have customized model algorithm development capabilities, model training and reasoning capabilities, etc.
03
Computing level
The computing power resources in the current computing power market cannot accurately lock inappropriate computing behaviors. There is a lack of security technology to ensure the security of computing power. There is a risk of personal privacy leakage in user models and data during the computing process. Secondly, computing power resources lack incentives. , There is also a lack of effective coordination mechanism between computing power nodes and computing tasks, which will lead to low utilization of computing power resources. At present, the domestic computing power market is not perfect enough.
Combined with the problems at the data, computing power, and algorithm levels, Impulse Online, as a leader in privacy computing infrastructure services combining software and hardware, has joined hands with Kunpeng to create a product solution for trusted AIGC all-in-one machines.
Core technical barriers
“
1
Compatible with Kunpeng TEE OS: Impulse trusted AIGC platform is compatible with Kunpeng hardware + Kunpeng TEE OS, provides standard privacy computing OS upwards, and realizes performance tuning of privacy computing algorithms based on Kunpeng BoostKit, integrates Kunpeng remote authentication and trusted metrics, and integrates rich privacy Calculation operator.
“
2
Heterogeneous hardware + TEE: The split learning framework uses GPU to accelerate deep learning in TEE. The CUDA agent of TEE provides a seamless algorithm migration experience upwards. The joint privacy computing of TEE+GPU is being accelerated.
“
3
Off-chain TEE + on-chain zero-knowledge proof: Realize the complementarity of credibility and security, and provide credible verification and auditing of AI training and reasoning processes based on privacy-protected data, algorithms, and computing power proofs.
“
4
Authenticable data fingerprint technology: Authenticable data fingerprints protect AIGC training data and output property ownership, and traceless watermark technology ensures that data quality will not be damaged.
The distributed data network and computing power network based on privacy computing can protect the privacy and rights of training data providers, provide safe output means for high-quality industry and personal data, and protect the privacy of data based on Kunpeng TrustZone to meet compliance requirements. According to regulatory requirements, data ownership is guaranteed through watermarking technology and stored on the chain; it can provide high-quality, credible, and cost-effective massive AI computing power for computing power demanders, and provide performance tuning and intelligent task scheduling based on Kunpeng BoostKit. Improve the resource utilization of idle computing power, and support diversified GPU acceleration capabilities to meet the computing power requirements of different scenarios.
At present, the platform has launched dozens of mainstream graph-generated graphs and Vincent graphs AIGC models; multiple levels of trusted AI computing power meet different scenarios, and supports one-click initiation of AIGC model training or reasoning tasks; comprehensive coverage of training data crowdfunding network .
In the future, Impulse Online will continue to deeply integrate Kunpeng series products, build a larger AIGC distributed computing power network based on Kunpeng ecology, empower the digital transformation of various industries, and promote the prosperity and development of the AIGC industry.