The power of the machine reports.
Edit: sia
Not everyone in the industry buys the account of this crazy number.Sam Altman has been making headlines lately, and the reasons behind it are somewhat exaggerated.
The smart man, who had never designed a chip, thought the semiconductor industry needed him and was pushing for a project aimed at increasing global chip manufacturing capacity. To do so, he needs to raise between $5 trillion and $7 trillion and is in talks with different investors, including the UAE**.
An OpenAI spokesperson said, "OpenAI had productive discussions about increasing the global infrastructure and **chain of chips, energy, and data centers, which are critical for AI and related industries. Given the importance of national priorities, we will continue to keep the U.S.** informed and look forward to sharing more details later. ”
According to the report, Altman's sky-high figures may include not only the construction of the fabs themselves, but also the construction of a completely new infrastructure around them, including power plants, etc.
Considering that Altman is a newcomer to the chip industry, everyone is also curious about the number he calculated from **. How exaggerated is $5 trillion to $7 trillion?
The GDP of the United States is about $23 trillion, which is equivalent to one-third of the GDP of the United States. In the case of the chip industry, this figure is about 4 times Nvidia's current market capitalization and well exceeds the global semiconductor industry valuation, which had sales of $527 billion last year and is expected to reach the $1 trillion mark by 2030.
Recently, Jim Keller, the Silicon Valley chip god, the "silicon fairy", Tenstorrent CTO, also expressed his "opinion" on Altman's astounding ambitions on X, stressing that the same work, he can accomplish for less than a trillion dollars.
Okami emphasizes "less" $1 trillion.
Earlier, when Altman asked why not raise the amount raised from $7 trillion to $8 trillion, the "Silicon Immortals" replied, "I can do it with less than $1 trillion."
Altman's chip plan is essentially to completely expand the semiconductor ** chain in order to solve the problem of chip supply shortage in the next three to five years, but it may also lead to problems such as foundry overcapacity and chip depreciation. Jim Keller argues that the focus is not on making more chips, but on the complexity of processors and simplifying the hardware chain (to reduce the cost of AI servers and other devices). "Start with a place to eliminate the profit stack," Keller writes. In the process of delivering products to end users, in order to obtain more gross profits, the participants in the ** chain are "adding layers of weight", and in Keller's view, it is possible to remove two or three layers of "stacking links". Next, in order to make the chip work faster, it is necessary to improve the matching of hardware and software resources. Of course, this is more difficult to do, but it is not impossible. 
He believes that through some improvements to the ** chain and the improvement of hardware and software resources, it is more conducive to solving the chip problem. Of course, this is also a very difficult task.
In fact, Jim Keller's TensTorrent is a "problem solver". They have a very ambitious roadmap to rapidly improve existing AI-based chip architectures, including RISC-V-based high-performance CPU chiplets and advanced AI accelerator chiplets to provide powerful solutions for machine learning.
Among them, the CPU is the main play of TensTorrent - in AI computing, the CPU plays a very, very important role, especially in terms of training. The CPU consumes more than 50% of the time and power consumption during AI training in the data center, including the pre- and post-processing of the data by the CPU.
Jim Keller, who is also an angel investor in the company, said, Tenstorrent's design is "the most promising architecture."
The demand for AI performance is growing at a rapid pace, and only time will tell if Tenstorrent and others will catch up with it in the foreseeable future.
The company's ambitious roadmap.
Altman's idea was not well received by industry insiders, with many believing that his idea was not feasible. Just recently, NVIDIA CEO Jensen Huang said that the AI-powered data center market will scale to $2 trillion in the next five years, emphasizing the fact that increasing capacity is only one aspect, and that architectural change is also important.
It doesn't take much investment to build an alternative semiconductor chain for AI chips. Instead, the industry needs to continue to innovate GPU architectures to improve performance — in fact, Huang claims that NVIDIA has improved AI performance by a factor of 1 million over the past decade.
The performance of the chip architecture will increase at the same time, and you can't assume that more computers will be bought. "You also have to assume that computing is going to get faster, so overall, you're not going to need as many chips," Huang said. ”
NVIDIA GPUs are evolving very fast in terms of AI and high-performance computing (HPC) performance. In 2018, the NVIDIA V100 data center GPU had only 125 TFLOPS of half-precision compute performance, but the H200 delivered 1,979 FP16 TFLOPS. In the future, computers will also be able to accomplish tasks at an even faster pace.
The shortage of AI chips will eventually be resolved, thanks in part to architectural innovations that have eliminated billions of dollars from spending billions of dollars on data centers for companies that want to use AI on-premises.
Reference Links.