Article reprint source: AIGC
Original source: Pinwan
Image source: Generated by Unbounded AI
The latest entrant into the independent chip-making trend is OpenAI. Reuters quoted sources as saying that OpenAI is planning to develop dedicated AI chips.
Currently, this plan is still in its early stages and has not made any substantial progress. After all, the cost and difficulty of developing a self-developed AI chip is no less than developing a large language model like GPT. It requires a continuous investment of billions of dollars over a period of several years to achieve. Especially as a startup company, OpenAI has no experience in developing hardware before, which means that this investment may not necessarily succeed in the end.
But the fact that OpenAI is planning to develop its own chips is not surprising to the market. Silicon Valley and Wall Street are closely watching the progress of this artificial intelligence unicorn. One of the focuses is whether OpenAI's financial situation and revenue scale can match its vision after launching the groundbreaking chatGPT product?
There are several bottlenecks that prevent OpenAI from expanding its scale, and chips are one of them. Currently, OpenAI has a huge amount of computing power provided by Nvidia graphics cards directly and indirectly through cooperation with Microsoft. The partnership between OpenAI and Nvidia can be traced back to its early days - Huang Renxun donated the world's first DGX supercomputer to OpenAI for research use.
But this kind of cooperation comes at a price, especially Nvidia's exclusive position in the field of artificial intelligence computing power has affected OpenAI's research and development progress. CEO Altman has said more than once that OpenAI's computing power is tight, which not only makes the operating costs of large models rise sharply, but also affects the development of new models.
Once OpenAI has self-developed chip technology, it will significantly benefit its artificial intelligence development process, including significantly reducing costs, more direct and precise control of research and development progress, and getting rid of various direct or indirect intervention controls by Nvidia and Microsoft. OpenAI will not Less computing power is provided through Microsoft's Azure platform.
In fact, getting rid of Nvidia's computing bottleneck is the common demand of all big model players, and Microsoft is no exception. After years of rumors, it is reported that Microsoft's self-developed artificial intelligence chip named Athena will be launched at the annual developer conference in Seattle next month.
At present, people know very little about Athena, and it is not even certain whether the chip will be named Athena in the end. It is only known that this chip is designed for the development of large artificial intelligence models and data centers. People once thought that this chip was jointly developed by Microsoft and AMD, but this news was later denied by Microsoft.
It is reported that this chip has entered the small-scale testing stage, and engineers from Microsoft and OpenAI have been trying out this chip to verify its performance.
Microsoft's Athena Project will be implemented from 2019. Compared with Google's TPU and Amazon Cloud's self-developed chip development plan, Microsoft's progress seems to be a little behind. The outside world is concerned about how much performance improvement Microsoft's chip solution will have, and more importantly, how this will affect the cooperation between Microsoft and OpenAI.
The cooperation between OpenAI and Microsoft was initially interpreted as a strong alliance, but this judgment was soon proved to be overly optimistic. The two parties compete in areas such as model development and customer resources, presenting a complex nested relationship of "fighting and pulling", and now this relationship has come to the field of self-developed chips.
Facts have proved that OpenAI does not want to rely entirely on resources provided by Microsoft, and it is a significant move to try to develop AI chips independently.
Now the question is back to square one. If OpenAI has decided to develop its own chips, what is its most likely next move? Many people have turned their attention to OpenAI's investment portfolio. After all, CEO Altman was previously the boss of the famous incubator Y Combinator, so directly acquiring a chip company might be a good choice.
Atomic Semi has come into people's attention. It has received investment from OpenAI Startup Fund and is currently valued at over 100 million US dollars. The introduction on the official website points out that the company's vision is to build "anything". Its two founders, Sam Zeloof and Jim Keller, the former became an Internet celebrity for rubbing chips by hand in his garage, and the latter has a very glamorous semiconductor industry resume and led the development of AMD's earliest Zen architecture, which AMD used to achieve a Jedi counterattack in the processor field.
In fact, Jim Keller, who has worked for AMD, Intel, Broadcom and Apple, is also the CEO of another artificial intelligence chip startup Tenstorrent. This company, which has raised more than $300 million in seven rounds of financing, is directly targeting the field of large-model training and reasoning. The chip uses the RISC-V architecture. Just in early October, Tenstorrent announced a partnership with Samsung, which will manufacture chips for it.
Therefore, overall, considering Sam Altman's influence and connections in the venture capital circle, OpenAI still has many options to choose from if it wants to make the first move in developing its own chips. One of the advantages of starting from scratch is that it can get rid of the burden of architecture and compatibility, especially it can launch highly targeted products, such as the acceleration chip based on the RISC-V architecture, which seems to be a good choice.
Of course, OpenAI has another more stable way to promote the pace of self-developed chips, which is to cooperate with other large manufacturers. However, Microsoft's Athena Project has clearly shown that OpenAI seems to prefer to be completely self-reliant, which is in line with the company's consistent style.
Another recent rumor can support this point: there are rumors that Sam Altman, Masayoshi Son and former Apple chief designer Jony Ive are planning to build new "AI hardware", which is obviously a field that OpenAI has never been involved in before. It can be seen that OpenAI's overall strategy is based on powerful large-scale model capabilities and independent development in all related links - this is clearly an idea for building a giant enterprise.
The trend of self-developed chips continues to rise. Now all large cloud computing companies are vigorously promoting the development of self-developed chips. The most important reason is the saying "The world has suffered from Nvidia for a long time."
The monopoly position is actually secondary. The key is that Nvidia's own production capacity is not enough, and it is still implementing a "rationing" system similar to luxury goods, which of course caused dissatisfaction among large companies. Especially as Nvidia has made a lot of money in this wave of artificial intelligence, Huang Renxun has taken the initiative to invest in a number of related start-ups. These companies have one thing in common: they all firmly embrace Nvidia's hardware ecosystem. These green factory direct descendants will be strong competitors in the cloud market in the future. Of course, large companies will not let this situation grow.
Self-developed chips are the only way to go, and it is not completely impossible to break Nvidia's monopoly. Self-developed chips can better customize relevant performance according to their own needs. This may not be obvious enough for cloud vendors, but for artificial intelligence companies like OpenAI, the benefits are huge. It does not need to consider the complex needs of various users in the data center (such as bandwidth), but only needs to focus on improving the efficiency of parallel computing. For many chip startups, this wave of self-developed chips is also a huge opportunity, especially chips based on the RISC-V architecture. In this regard, China, as one of the largest players in this open source architecture, even has a first-mover advantage.
Although no one can really challenge Nvidia's leadership at present, the entire artificial intelligence chip market is clearly on the eve of drastic fluctuations, and changes are coming.
