The latest developments in artificial intelligence chips — KellyOnTech Chip Series
This is the fifth episode of the chip series. In this episode, let’s talk about the latest developments in artificial intelligence chips.
What is the prospect of the artificial intelligence chip market?
According to Allied Market Research report, the artificial intelligence chips market was valued at USD 6,638 million in 2018 and is expected to reach USD 91,185 million by 2025, growing at a CAGR of 45.2% from 2019 to 2025.
The growth of the global AI chip market is mainly driven by the increasing demand for smart homes, the development of smart cities, and the emergence of quantum computing. In the near future, more and more countries will adopt AI chips to develop smarter robots.
Video version: What is the prospect of the artificial intelligence chip market?
What is an artificial intelligence chip?
According to Wikipedia, AI accelerators are a type of specialized hardware accelerators or computer systems designed to accelerate artificial intelligence and machine learning applications, including artificial neural networks and computer vision.
Typical applications include robotics algorithms, the Internet of Things, and other data-intensive or sensor-driven tasks. They are typically multi-core designs, often focused on low-precision arithmetic, novel data flow architectures, or in-memory computing power.
As of 2018, the typical AI integrated circuit chip contains billions of MOSFET (metal–oxide–semiconductor-field-effect) transistors. Many vendor-specific terms exist for such devices, and it is an emerging technology without a dominant design.
Video version: What is an AI chip?
Which company is the global leader in AI chips?
It is Cambricon Technology Corporation, established in 2016. It is engaged in the design and sales of artificial intelligence core chips in various cloud servers, edge computing devices, and terminal devices, providing customers with a wealth of chip products and system software solutions.
In May 2018, Cambricon released its third-generation AI chip for edge devices with the help of TSMC’s (Taiwan Semiconductor Manufacturing Company) 7nm process. Moreover, the company aims to have 1 billion smart devices that are expected to use its AI processors.
On January 21, 2021, Cambricon’s SiYuan 290 smart chip and accelerator card and the Xuansi 1000 smart accelerator were officially unveiled for the first time after mass production. The SiYuan 290 smart chip is Cambricon’s first training chip, using TSMC’s 7nm advanced process, integrating 46 billion transistors, supporting MLUv02 extended architecture, and fully supporting AI training, inference or hybrid AI computing acceleration tasks.
We know that GPUs have led the artificial intelligence chip market since 2018, used to process images and solve most parallel computing problems, such as Bitcoin mining. But in the future ASIC (application-specific integrated circuit) dedicated integrated circuits, also called ASIC custom chips, will overtake GPUs in terms of revenue.
What is an ASIC?
It is a dedicated integrated circuit is an integrated circuit chip that is customized for a specific application, not for general purpose use. For example, a chip designed to run in a digital recorder or efficient bitcoin miner is an ASIC.
Video version: Which company is the global leader in AI chips?
Google TPU
The most famous ASIC custom chip is the familiar Google TPU (Tensor Processing Units). At this year’s I/O Developers Conference, Google announced its fourth-generation TPU artificial intelligence chip. Google said its speed is twice that of the previous version launched in 2018.
As Google CEO Sundar Pichai pointed out, these chips will then be combined with 4,096 v4 TPUs in pods. A single Pod provides more than 1 exaflop of computing power. Is that not much of an idea? Google claims that if there are about 10 million laptops all working together in the world, then it can be “almost” comparable to a single pod in terms of power.
Which artificial intelligence chip startup is worth watching?
Which artificial intelligence chip startup is worth watching? — Tenstorrent
I will introduce an AI chip startup-Tenstorrent, touted by the industry as having the most promising architecture on track to become the next generation of computing giants.
Founded in 2016 in Toronto, Canada, Tenstorrent is the world’s first chip company to design a conditional execution architecture for facilitating scalable deep learning. Its flagship product is GraysKull.
Tenstorrent takes an approach that dynamically eliminates unnecessary calculations, thus breaking the direct link between the growth of model size and the calculation/memory bandwidth requirements. Conditional calculations make it possible to adapt the model’s inference and training to the precise input proposed, such as adjusting the NLP model calculations to the exact length of the presented text, and the dynamic trimming of the model based on input features. Grayskull provides significant baseline performance improvements in today’s most widely used machine learning models, such as BERT, Resnet-50 and other machines, while using conditional execution to open up order-of-magnitude gains during conditional execution of the current model and future executions optimized for this method.
In May this year, Tenstorrent raised $200 million in a new round of financing from investment institutions led by Fidelity Management, including Eclipse Ventures, Epic CG and Moore Capital. The deal is valued at US$1 billion and also makes Tenstorrent a unicorn company for artificial intelligence chips. I introduced another artificial intelligence chip unicorn company-SambaNova Systems previously, take a look if you’re interested.
The Founding Team of Tenstorrent
Ljubisa Bajic, co-founder and CEO of Tenstorrent, has been designing chips at AMD for more than a decade. Bajic also worked at Nvidia.
When designing AMD’s now very successful Zen architecture in 2014, he met Jim Keller, a star chip designer invited by AMD. Keller has designed breakthrough chips at Apple and Tesla.
Bajic also successfully persuaded Keller to resign as Intel’s senior vice president in June 2020 and join Tenstorrent as president and chief technology officer. With this industry powerhouse on board, Tenstorrent has set a goal: to develop chips for training machine learning programs, surpassing the current global design.
The company’s strategy is to use software to better allocate the use of computing power on each chip. Gossip here, in fact, Bajic founded Tenstorrent with the encouragement of Keller who provided seed funds for the company. Tenstorrent also focuses on developing a different way to link its many chips so that they can share data faster than standard products. Instead of connecting to each other via external networking technologies such as Ethernet, these chips can communicate more directly with each other.
AI Chip Market Outlook
According to IDC, a leading global market intelligence firm, the dedicated AI chip market reached nearly $15 billion in hardware last year, led by Nvidia, and is expected to grow by 20 percent annually to $31 billion by 2024.
Recently, Baidu launched the second-generation artificial intelligence chip Kunlun 2 for autonomous driving and other fields.
A prototype of the “robot car” was also released, highlighting its ambitions in autonomous driving.
I will continue to follow the latest developments in AI chips, and for those who want to compare the two AI chip unicorns, feel free to check out the SambaNova Systems video.