Tenstorrent Lian Weihan: RISC-V and AI are a perfect fusion

Recently, Lian Weihan, chief architect of Tenstorrent, shared the combination of RISC-V architecture and AI innovation at the "2024 World Artificial Intelligence Conference" - "Intelligence from the Core" RISC-V and Generative AI Forum hosted by VeriSilicon, taking Tenstorrent as an example, he shared the combination of RISC-V architecture and AI innovation. He said that open source, especially RISC-V, plays a key role in promoting the development of AI computing power.

Tenstorrent is led by legendary CPU architect Jim Keller as CEO. In February this year, it signed a multi-level AI development cooperation agreement with LSTC to jointly build a 2nm AI accelerator chip. LSTC is the abbreviation of "Technology Research Group, State-of-the-Art Semiconductor Technology Center". It is a semi-official organization in Japan, led by the advanced foundry company Rapidus and established in 2022. LSTC's members include the RIKEN Institute of Physical and Chemical Research, the National Institute of Advanced Industrial Science and Technology, and several Japanese universities. LSTC's goal is to promote the design and manufacture of artificial intelligence chips.

Computing Challenges in the AI ​​Era

As AI technology develops rapidly, computing power has become the core force supporting this change. Lian Weihan pointed out that although AI algorithms and applications have made great progress, the problem of computing power is still the key bottleneck restricting its further development. Only by solving the problem of computing power can the potential of AI be fully released. Chen Weihan said: "AI's demand for computing power is almost "unprecedented and unparalleled", and Tenstorrent also firmly believes that the demand for computing power must occur everywhere."

The computing power should be determined based on cost, power consumption, and hardware requirements. Tenstorrent focuses on the computing power requirements of data centers and the cloud, providing chip manufacturers with suitable chiplets.

In addition, Tenstorrent is also developing more AI-related products such as chips and software. Currently, companies including Samsung, LG, and LSTC are cooperating with it.

The key to Tenstorrent is its processor core, Tensix Cores. Each Tensix Core contains five RISC-V processors, an array math unit for tensor operations, a SIMD unit for vector operations, hardware for accelerating network packet operations and compression/decompression, and up to 1.5 MB of SRAM.

The rise of open source architecture

As an open source instruction set architecture, RISC-V plays an important role in the development of AI computing power due to its flexibility and scalability. Chen Weihan said that RISC-V has better scalability and can support all AI needs from edge to cloud with one instruction set, unlike Arm which has different instruction sets. In addition, RISC-V's scalability allows it to be expanded according to application needs. In terms of efficiency, since it does not need to be compatible with past applications, it achieves maximum efficiency. In terms of stability, after years of development and cooperation in the industrial chain, RISC-V has been proven to be a commercializable instruction set.

AI requires a more flexible approach

Chen Weihan particularly emphasized that there are many bottlenecks in the speed of AI processing. Sometimes it is necessary to wait for data to arrive, and sometimes explosive data processing is required. At the same time, power consumption must also be paid attention to. It is precisely because of the scalability of RISC-V that it can have just the right flexibility to execute AI computing power.

In addition, he also said that high-performance CPUs have great potential in AI. Not all AI is suitable for matrix operations and GPU training. In fact, some applications can be calculated on high-performance CPUs, and Chiplet is very suitable for combining high-performance CPUs and AI.

It’s just the beginning

Chen Weihan is very optimistic about the development of RISC-V in the field of AI because of its strong scalability, expandability, high efficiency, and diversified IP suppliers. He concluded: "It is just the beginning. As an open source project, RISC-V has unlimited imagination and development potential."