OpenAI, one of the top streams in the big model world, may become a new member of the independent core manufacturing trend.
Recently, according to media reports, OpenAI is exploring the manufacture of its own artificial intelligence chips and has begun evaluating a potential acquisition target. Since at least last year, the company has discussed various options to address the shortage of expensive AI chips that OpenAI relies on, according to people familiar with the matter. Those options include making your own AI chips, working more closely with other chipmakers, including Nvidia, and diversifying suppliers outside of Nvidia.
Among the positions open for recruitment on the official website of OpenAI, hardware engineer related positions have recently appeared to evaluate and co-design AI hardware. In September, OpenAI also recruited Andrew Tulloch, a leader in the field of artificial intelligence compilers, from Meta, which seems to confirm the investment in self-developed chips.
Bitter heroic for a long time?
OpenAI layout of self-developed chips, the market does not seem to be surprised.
In the past half a year when large models have been popular, if data is the raw material for the production of large models, then computing power is the productivity of large models. Previously, OpenAI released a set of data that the growth rate of computing power required for large model training remained at a rate of 3-4 months/times, far exceeding Moore's Law by 18-24 months/times.
Powerful computing power means faster data processing speed and more powerful large model service capabilities. At the same time, as more and more enterprises enter the big model, resulting in a sharp increase in demand for high-end GPUs such as A100 and H100, NVIDIA and its manufacturing partner TSMC are struggling to meet the supply.
According to the Financial Times, in August, Baidu, ByteDance, Tencent and Alibaba ordered a total of $5 billion worth of chips from Nvidia. The outbreak of external demand has led to an exceptionally tight supply in the market. According to Brian Venturo, co-founder and CTO of CoreWeave, in the first quarter of this year, it was still easy to get an NVIDIA GPU, but from April, the delivery time will have to wait until Q1 or even Q2 of 2024.
OpenAI CEO Sam Altman has also repeatedly complained about the shortage of computing power, a market dominated by Nvidia, which controls more than 80 percent of the world's chip market best suited to run AI applications.
At a recent London hearing, Altman said the shortage of computing power kept API availability below par and frankly acknowledged that the shortage of computer chips could prevent ChatGPT from rolling out a larger "context window." The context window determines how responsive the model is and what data is used for a single prompt.
Financial commentator Zhang Xuefeng said, "OpenAI's self-developed chips mean that they want to independently develop and produce customized chips that meet their AI technology needs." Self-developed chips can enable OpenAI to better optimize the collaborative work of algorithms and hardware, and improve the performance and efficiency of artificial intelligence systems. At the same time, self-developed chips can also reduce dependence on suppliers, reduce costs and better meet the unique needs of OpenAI. ”
In fact, talking about computing power is always inseparable from the cost issue. Analyst Stacy Rasgon has estimated that ChatGPT costs about 4 cents per query. If ChatGPT queries grow to one-tenth the size of Google search, it would require about $48.1 billion worth of GPUs and $16 billion a year of chips to keep running.
According to foreign media reports, OpenAI's revenue in 2022 will be $28 million, with a loss of $540 million, and the main reason for the loss is the cost of computing power. In addition, ChatGPT mobile revenue increased to $4.6 million in September, but growth weakness appeared. According to the latest data from market intelligence firm Appfigures, its revenue growth has begun to slow to 31% in July, 39% in August and 20% in September.
It's unclear whether OpenAI will move forward with plans to customize chips. However, Zhang Xuefeng pointed out that OpenAI self-developed chips still face challenges in different aspects: First, technology, self-developed chips require a high degree of technical strength and expertise, and OpenAI may require more time and resources to improve research and development capabilities and related technologies. OpenAI may need to evaluate its strategic benefits and weigh the long-term benefits brought by self-developed chips. Third, the supply chain, the AI chip market is highly competitive, and the supply chain shortage problem still exists, which may cause OpenAI to postpone its self-developed chip plan to wait for a better market opportunity.
Collective self-developed chips
After OpenAI was exposed to self-developed chips, the chip story of its partner Microsoft also had a "preview".
On October 6, The Information, citing people familiar with the matter, said that Microsoft plans to launch its first chip designed for artificial intelligence at next month's annual developer conference to reduce costs and reduce dependence on Nvidia. Microsoft chips are used in data center servers, designed to train software such as large language models (LLM), while supporting inference, powering all the AI software behind ChatGPT.
Microsoft's chip research and development work has begun since 2019, but at present, everyone knows very little about relevant information. According to The Information, the chip, code-named "Athena," will be built using TSMC's 5nm process. If successfully put into production, the cost per chip is expected to be reduced by one-third.
In addition to Microsoft, other large technology companies have also begun to enter the era of self-developed chips. As early as May 18 this year, Facebook's parent company Meta disclosed the details of its data center project to support AI work, mentioning that it has built a custom chip, MTIA for short, to speed up the training of generative AI models. This is the first time Meta has launched an AI custom chip. Meta said MTIA is part of a "family" of chips that accelerate AI training and inference workloads.
It is worth noting that Google and Amazon have already started their own chip plans. As early as 2013, Google secretly developed a chip focused on AI machine learning algorithms and used it in internal cloud computing data centers to replace Nvidia's GPUs. In May 2016, this self-developed chip was unveiled, that is, TPU.
Amazon is the first cloud manufacturer to get involved in self-developed chips, launching its self-developed AI inference chip Inferentia in 2018; Earlier this year, Inferentia 2, built for artificial intelligence, was released three times faster compute performance, a quarter of total accelerator memory, a quarter of throughput, and a tenth of latency.
Tracy Woo, a senior cloud analyst at Forrester Research, said the AI boom is putting more pressure on cloud providers to develop their own chips: You can buy them from Nvidia, but when you look at the Big Macs like Google and Amazon, they have the capital to design their own chips.
At present, Microsoft, Google and other large manufacturers have come down to develop their own chips, angel investor and senior artificial intelligence expert Guo Tao believes that the core reason is to improve computing power, reduce dependence on external suppliers, enhance competitive advantages and promote derivative design.
The "Golden Triangle" is not solid?
As soon as the news of OpenAI and Microsoft's self-developed chips broke, one of the reasons pointed to "reducing dependence on NVIDIA".
The story of NVIDIA, OpenAI and Microsoft starts in 2016. OpenAI, which was founded only one year ago, met Nvidia CEO Jensen Huang and harvested the lightweight small supercomputing DGX-1, which can be completed in one month for a year's calculation. Three years later, OpenAI and Microsoft joined hands and received a $100,000 investment.
At that time, the money contributed, the effort contributed, and the infrastructure that contributed the infrastructure supported the research and development of the OpenAI large model. Until the end of last year, the birth of ChatGPT caused a sensation all over the world. The three-company collaboration was put in the spotlight and was dubbed a "strong alliance" in the industry, but over time, the biggest winner seems to have been Nvidia.
At present, NVIDIA occupies 82% of the global data center AI acceleration market, and monopolizes the global Al training market with 95% market share, and the company's market value exceeds $1 trillion. At the earnings conference at the end of August, NVIDIA CFO Colette Kres expected demand for NVIDIA data center GPUs to continue into next year, "As we shorten lead times and increase capacity, supply will continue to increase in the coming quarters." ”
Faced with the challenges of high cost and chip shortage, OpenAI and Microsoft have begun to develop their own chips. As soon as the news came out, in addition to caring about chip performance in the market, many people in the industry questioned the relationship between the three: whether the three will continue to maintain a good cooperative relationship, or there will be a situation of "tearing faces".
In addition, Google, Microsoft, OpenAI, etc., as important customers of NVIDIA, will undoubtedly compete with NVIDIA after the formal application of chips that have not been developed, so there is also a voice in the industry "whether it will shake NVIDIA's status".
"At present, the NVIDIA artificial intelligence chip market is in a monopoly position, but this situation may change as other companies also begin to develop their own chips." However, Guo Tao judged that NVIDIA has rich experience and technical advantages in the field of artificial intelligence, and the chip ecosystem is relatively mature, so it may still maintain its "hegemonic" status.
In Zhang Xuefeng's view, NVIDIA, as a leading company in the AI chip market, will still maintain a competitive advantage in the AI chip war, but competitors can better meet the needs of specific fields and applications through self-developed chips, which is expected to promote market diversification, and its 'hegemony' position may face certain impacts. He said: "As more enterprises, including large manufacturers and startups, realize the importance of self-developed chips, the AI chip battle may present a situation of multiple participants' self-developed chips in the future, and the further development of the market will depend on the comprehensive impact of technological innovation, application demand and market competitiveness." ”
Comprehensive from CaiLian News, Wall Street News, Phoenix Network Technology, Interface News