This article is reproduced from OpenHarmony TSC official WeChat public account " Summit Review Issue 4 | Building a Trusted Execution Environment in Heterogeneous Computing Scenarios "
Speakers | Jin Yier
Review and arrangement | Liao Tao
Typesetting proofreading | Li Pingping
Guest profile
Jin Yier, Chief Scientist of Huawei Trusted Computing, Co-Chairman of IEEE Hardware Security and Trustworthy Committee, Member of OpenHarmony Technical Steering Committee Security and Confidential Computing TSG, Honorary Professor of University of Florida. He graduated from Yale University in 2012 with a Ph.D. in Electrical Engineering. Wrote the book "Integrated Circuit Security" and published more than 200 papers in internationally renowned journals and conferences. He is the co-founder of the Asian Hardware Security Annual Conference and has also participated in several academic conferences and academic conferences on information security and integrated circuits. magazine organization and is currently a Distinguished Lecturer for the IEEE Committee on Design Automation (CEDA).
Article content source
The First Open Atom Open Source Foundation OpenHarmony Technology Summit - Security and Confidential Computing Sub-Forum
video review
Video link:
嵌入代码：<iframe src="//player.bilibili.com/player.html?aid=994532812&bvid=BV1Fs4y1S7ZT&cid=1082741929&page=1" scrolling="no" border="0" frameborder="no" framespacing="0" allowfullscreen="true"> </iframe>
Contents
Heterogeneity refers to the combination of CPU, DSP, GPU, ASIC, FPGA and other computing power units with different process architectures, different instruction sets, and different functions to form a hybrid computing system, so that it has more powerful and efficient functions. How to build a trusted execution environment in a heterogeneous computing scenario? Professor Jin Yier, Chief Scientist of Huawei Trusted Computing and Co-Chairman of the IEEE Hardware Security and Trust Special Committee, put forward some thoughts at the first OpenHarmony Technology Summit.
Jin Yier first started from the phenomenon of the slowdown of Moore's Law. Moore's Law has developed rapidly from 1975 to 2020, which has greatly increased the density of integrated transistors in chips and promoted the rapid operation of the "flywheel" of the semiconductor business model, but it has gradually slowed down; from 2020 to 2030, in Under the guidance of TSMC's advanced chip manufacturing process technology development, the slowdown of Moore's Law has become more prominent. It seems that the dividend of Moore's Law will "come to an end." In addition, laws such as Dennard Scaling and Amdahl's Law are gradually slowing down or becoming invalid, which indicates that the computing structure will change. How will the computing structure change? At present, there are mainly three types of views: 1. Architecture changes will occur in the past ten years, that is, heterogeneity; 2. Materials will be updated in the past 20 years; 3. In the longer term, the computing model will change. In short, heterogeneity will become the mainstream in the short term.
Remarks: The picture comes from public information
At the same time, artificial intelligence (AI) promotes the optimization and upgrading of industrial structure, and the trend of industrial intelligence is about to subvert traditional industries. In order to make better use of AI, it is first necessary to provide it with a large amount of computing power in heterogeneous scenarios. In the context of the increasing amount of computing data, the performance tax phenomenon of data centers has gradually become prominent, and the bottleneck of computing power has shifted from data computing to AI. for data transfer. When the CPU performs data calculations, a large amount of time is not doing the calculations it is good at, but moving data. Based on this, Nvidia proposed a plan to transfer data handling from CPU to DPU. In the evolution of the future data center, the CPU is used to support general computing, the GPU supports computing acceleration, and the DPU supports data movement and data processing in the data center. .
Remarks: The picture comes from public information
How should security strategies develop in the era of gradual changes in computing architecture? The areas that CPUs are not good at are gradually being transferred to professional hardware, resulting in computing gradually moving away from the CPU, while the confidential computing environment still revolves around the CPU. What if the CPU is missing, what about the confidential computing environment? There are two main directions for system security construction, one is the security of the system itself, and the other is the security capabilities provided by the system. The security of the system itself is similar to CET, CFI protection, etc.; the security capability provided by the system, that is, the confidential computing environment, etc.
The following figure shows the scenarios of trusted computing and heterogeneous computing, reflecting the "parallel" and "independent" development of heterogeneous computing and trusted computing. In the iteration of heterogeneous computing scenarios, from being CPU-centric, to CPUs being just ordinary PUs, to the possibility that there may be no CPUs in the future. However, confidential computing is very "loyal" and always revolves around the CPU, even if the latest ARM CCA technology is still around the CPU. So how to build a trusted execution environment in a heterogeneous computing scenario? Distributed confidential computing is one way of thinking, and applying the technology on the CPU to other XPUs is another way of thinking, but the most critical question is how to use it. This problem is very obvious in AI security. AI models are all on the GPU, and TEE can’t be used even if it wants to, because its protection boundary has not been expanded.
In academia, Graviton first proposed the idea of building a large trusted execution environment. By modifying the GPU, the concept of the trusted execution environment of the CPU can be used on the GPU; HIX proposed that it is not necessary to change the GPU, but by changing the CPU, It is also feasible to strengthen the I/O path and extend TEE to GPU devices; HETEE has proposed a method to improve the interconnection, which is also a cross-generational work; in addition, some scholars have proposed observation methods to construct a Large trusted execution environment.
In general, it mainly includes 3 clear routes, that is, either change the GPU, or change the CPU itself, or change the interconnection. In the development of computing in the future, the CPU is no longer the center. How should the CPU-centered security concept be expanded? In view of the premise that the CPU will continue to exist in the short term, the current solution still takes CPU expansion as the mainstream. The innovation of the interconnection bus will also bring about fundamental changes to the architecture of the data center. And if the interconnection bus can support the expansion of the trusted environment, a trusted execution environment similar to the original data center can be built.
Going back to embedded systems, we find that embedded systems are going faster. The SOC on mobile phones has achieved that all modules are concentrated on one SOC a long time ago. After opening it, I found that the expansion of the trusted environment of the SOC is faster and earlier than that of the data center, but because it is inside the mobile phone, no one has discovered it. Whether Qualcomm's chips, Apple's chips, or Huawei's chips, if you look carefully, they all have a small TEE environment. In fact, the CPU has been packaged in it long ago, but these technologies have not fed back to the entire computing system. system.
Remarks: The picture comes from public information
Professor Jin Yier finally mentioned whether the traditional TEE expansion and the TEE expansion on the mobile phone will be two lines, or will eventually merge, and which direction the traditional TEE expansion will go, whether to change the GPU or Change the CPU, or change the Internet? Looking forward to future cooperation and discussion between academia and industry.