ChatGPT uses Nvidia GPUs exclusively. For now.
Artificial intelligence algorithms are incredibly complex, needing the enormous power of supercomputers to operate. OpenAI’s ChatGPT runs on a Microsoft-built supercomputer powered by 10,000 Nvidia GPU’s. It’s expensive, and power-hungry, and the company is aiming to improve performance and reduce reliance on a single chip supplier. The result may be a new in-house semiconductor design unit at the company, or possibly an acquisition of an existing chipmaker. OpenAI has also announced that they are open to a collaborative effort with existing suppliers, including Nvidia.
Access all episodes of This Week in Engineering on engineering.com TV along with all of our other series.
* * *
Episode Transcript:
Integrated circuits, particularly micro-processors and memory, are the building blocks of almost every modern technology. Much of modern computing technology has its origins in basic general-purpose microprocessors like the venerable 8080, with code defining the application.
As more sophisticated needs developed in industries such as aerospace and automotive, special-purpose semiconductors evolved to handle the unique needs of industries where software simply wasn’t enough. Traditional, mass-market chip makers have never been interested in a very expensive design of special-purpose devices with relatively small production runs, and as a result, a new generation of “fabless” IC developers have evolved rapidly in the last decade.
Fabless production is a process made more complex by the Biden administration’s new sanctions against the transfer of integrated circuit technology to China. For some users in industries such as automotive, circuit density and processing speed may not be critical, and there are a wide range of manufacturing options for custom chips in the 20 nm range. But for the rapidly evolving AI industry, however, very dense, very fast processors will be needed.
Reuters has recently reported that ChatGPT owner OpenAI is coping with a shortage of high-performance ICs with multiple strategies, from tighter collaboration with current suppliers like Nvidia, to in-house development of their own designs. OpenAI currently uses a Microsoft-built supercomputer that uses semiconductors originally designed for graphics purposes: Nvidia GPUs of which 10,000 are required to operate the OpenAI system.
Reliance on Nvidia as sole source supplier is obviously problematic, as are the high costs of the units and the considerable infrastructure energy costs needed to run them. A new generation of AI-specific semiconductors are clearly needed, and OpenAI has expressed interest in alternate suppliers other than Nvidia, setting up an in-house design team and, interestingly, acquisition of an existing semiconductor maker.
The timing for the latter option is interesting. A large amount of U.S. federal money is available through the Chips Act for domestic semiconductor production, and although there is a global shortage of the very specialized equipment needed to produce advanced chips in the 7 nm range, export restrictions on this technology to China should in the short term alleviate that problem. There may never be a better time for a deep-pocketed new technology company to take control of its hardware needs than right now.
The ability of OpenAI to scale rapidly depends on semiconductor availability, and as artificial intelligence becomes ubiquitous globally, it’s also a possible secondary market as a hardware supplier to other AI service providers. Fab or fabless, the artificial intelligence industry may produce the next Microsoft, or Apple.
