For your better reading and interactive experience, in order for you to see more content in time, click "Follow" We update wonderful stories and share different story moments for you every day
A high-profile controversy is making headlines in the global technology landscape: traditional graphics processing units (GPUs) are showing their limitations in artificial intelligence (AI) applications. The discovery first caught a lot of attention among innovators in Canada and the U.S., who are working to develop new semiconductor technologies that can replace traditional GPUs. The goal of these companies is not only to create more efficient AI computing platforms, but also to directly challenge the industry giant Nvidia in the United States, which accounts for about 80% of the GPU market.
New explorations in the field of semiconductors
GPUs have long been widely used in the gaming sector, and they are designed primarily to process high-definition images. However, with the rise of generative AI, the power consumption of GPUs in this new field has become increasingly prominent. This challenge has led to the exploration and development of new semiconductor technologies, especially those designed specifically for generative AI. The goal of these startups is to develop semiconductor technologies that can run AI applications efficiently while significantly reducing energy consumption.
New trends in transnational cooperation
An important shift is quietly taking place: Junyoshi Koike, president of Japan's RAPIDUS, announced in November that they would be partnering with Canada's TENSTORRENT, marking their first step in tapping into the new AI semiconductor market. The focus of this collaboration is on the development and production of new semiconductors for AI computing. THE ADDITION OF JIM KELLER, CEO OF TENSTORRENT, A SEMICONDUCTOR ENGINEER WITH A REPUTATION IN SILICON VALLEY, ADDS MORE ANTICIPATION AND POSSIBILITIES TO THIS COLLABORATION.
A legend in the semiconductor world
Jim Keller, who has worked for well-known American companies such as Apple, Tesla and AMD, has participated in the design of many major semiconductors. His reputation in Silicon Valley is legendary, and he is widely regarded as a "legend among semiconductor engineers" in the industry. Junyoshi Koike spoke highly of Keller's joining, believing that it was of great significance to the industry as a whole.
A new chapter in technological innovation
RAPIDUS PLANS TO START PRODUCTION OF AI-SEMICONDUCTORS DESIGNED BY TENSTORRENT IN 2025 AT ITS NEW FACTORY IN CHITOSE CITY, HOKKAIDO. The move aims to mimic the successful partnership between NVIDIA and TSMC in the past, and aims to make a technological leap forward based on the division of labor. FOR RAPIDUS AND TENSTORRENT, THIS IS NOT ONLY A TECHNICAL COOPERATION, BUT ALSO AN IMPORTANT STEP IN OPENING A NEW ERA OF SEMICONDUCTOR TECHNOLOGY.
Limitations of GPUs
GPUs have long been known for their ability to excel at parallel computing of large amounts of data. However, they were originally designed for game image processing and not optimized for AI computing. This leads to a significant weakness in the data center servers that are the foundation of generative AI. Among them, the biggest problem is the inefficiency of electricity. In the design of GPUs, the separation of combinator and memory results in high energy consumption during data exchange, which is not necessary for the operation itself.
New solutions for power efficiency
IN THIS CHALLENGE, TENSTORRENT IS DEVELOPING A NEW TYPE OF AI SEMICONDUCTOR THAT REDUCES THE WASTE OF POWER BY PLACING COMBINATORS AND MEMORY CLOSER TOGETHER, SHORTENING THE DISTANCE THAT DATA CAN TRAVEL. Keller figuratively compares AI operations to a very natural way of dealing with AI operations where the direct dialogue between combinators and the direct transfer of the result of the computation to the next computation.
Adoption of open source
IN TERMS OF THE "INSTRUCTION SET" FOR THE OPERATION OF SEMICONDUCTORS, NVIDIA ADOPTED THE TECHNOLOGY OF THE BRITISH ARM, WHILE TENSTORRENT CHOSE THE OPEN-SOURCE "RISC-V". The aim is to control licensing costs and to flexibly change the design of semiconductors according to customer needs. Keller points out that multi-million yen GPUs are too expensive for companies looking to develop multi-million yen robots. AS A RESULT, TENSTORRENT IS WORKING TO DEVELOP A NEW GENERATION OF AI SEMICONDUCTORS THAT CAN COMPETE WITH NVIDIA IN TERMS OF PRICE.
New trends in inferential computing
With the development of generative AI, developers are devoting a lot of computing resources to improving the accuracy of the "Xi" stage. IN THE FUTURE, WITH THE POPULARITY OF SERVICES SUCH AS CHATGPT, AI MAY CONSUME MORE POWER FOR "INFERENCE" OPERATIONS THAT OUTPUT ANSWERS TO QUESTIONS. THIS TREND HAS SPAWNED A GROUP OF STARTUPS FOCUSED ON IMPROVING GPU ISSUES, SUCH AS D-MATRIX AND RAIN AI.
Emerging companies in memory computing technology
FOUNDED IN 2019 BY INTEL'S CEO SID SHETH, D-MATRIX FOCUSES ON THE USE OF AN "IN-MEMORY COMPUTING" TECHNOLOGY THAT ENABLES MEMORY TO PERFORM COMPUTING FUNCTIONS TO REDUCE THE POWER CONSUMPTION OF DATA CENTER SERVERS IN INFERENCE OPERATIONS. JUDGING FROM THE AI SEMICONDUCTOR "CORSAIR", WHICH D-MATRIX PLANS TO SUPPLY IN 2024, THE AMOUNT OF COMPUTING AND SPEED FAR EXCEED NVIDIA'S MAIN GPU.
New developments in the Chinese market
Nvidia has long had an important presence in the Chinese market, and China has been one of its largest markets. However, with the export restrictions of high-end GPU chips in the United States, Nvidia had to launch a special version of the product with reduced performance for the Chinese market. This strategy has sparked widespread discussion, with many Chinese companies turning to domestic GPU chips or choosing AI cloud services from major manufacturers to reduce their reliance on Nvidia products.
The rise of AI chips in China
China's AI chip industry is rapidly emerging, and many local companies such as Alibaba's Pingtou are developing dedicated AI processors to replace Nvidia's AI chips. This trend shows that the demand for AI chips in the Chinese market is gradually shifting to domestic solutions and reducing dependence on foreign technology. This not only reflects China's strategic shift towards technological self-sufficiency, but also reflects the trend of increasing diversification of the global technology market.