Original title: the total cost of a Taiwan Semiconductor Manufacturing Co Ltd 5nm chip exceeds 2900 yuan?
On Tuesday, Lei Feng reported that the research and development of Taiwan Semiconductor Manufacturing Co Ltd's 2nm process is ahead of schedule, and the industry is optimistic that the rate of risk trial production will be as high as 90% in the second half of 2023. This is both good news and bad news for the industry. The good news is that chips such as AI and mobile phones can achieve better energy efficiency and make more innovations with advanced manufacturing processes. The bad thing is that the cost of advanced process chips puts great pressure on wafer-free chip design companies.
A report entitled "AI Chips: What They Are and Why They Matter", written by two authors of the Center for Security and emerging Technologies (CSET) at the Walsh School of Foreign Affairs at Georgetown University, estimated with the help of the model that Taiwan Semiconductor Manufacturing Co Ltd's fee per 5nm wafer could be about $17000, nearly twice that of 7nm.
The report also estimates that each 300mm-diameter wafer can typically produce 71.4 5nm chips, which brings the wafer-less company's manufacturing cost to $238 (about 1642 yuan) per wafer.
Not only not as good as, through the investigation of semiconductor industry and AI chip design, the author estimated not only the manufacturing cost of 5nm chip at US $238 through the model, but also the design cost of US $108 per chip and the packaging and testing cost of US $80 per chip.
As a result, the total cost paid by chip designers for each 5nm chip will be as high as $426 (about 2939 yuan).
Advanced process of "Noble" after 5nm
The data given by International Business Strategies (IBS), a market research institution, show that the cost of chips rises rapidly after 28nm. The cost of the 28nm process is $62.9 million, but with 7nm and 5nm, the original chip has skyrocketed, and 5nm will rise to $476 million. Samsung says the cost of its 3nm GAA could exceed $500m.
The total cost of designing a 5nm chip will be as high as nearly $500m. What is the average cost per chip? CSET's model in the report is analogous to the NVIDIA Corp P100 GPU, which is based on Taiwan Semiconductor Manufacturing Co Ltd's 16nm node and contains 15.3 billion transistors with a bare chip area of 610mm, equivalent to a transistor density of 25 MTr/mm2。
As a result, silicon wafers with a diameter of 300 mm can produce 71.4 chips of 610 square millimeters.
Assume that the chip area of the 5nm GPU is 610mm square and the transistor density is higher than the P100 GPU, reaching 90.7 billion transistors. In the following table, the node transistor density between Taiwan Semiconductor Manufacturing Co Ltd 90 and 5nm is estimated by the model. For nodes in the range of 90 to 7nm, the model uses a hypothetical GPU with the same specification, including transistors except transistor density, and the hypothetical 5nm GPU is associated with the hypothetical node.
CSET's cost model uses a fab-free perspective, including the cost, materials, labor, manufacturing R & D and profits of building the plant. After the chip is manufactured, it will be outsourced to the Chip Test and Packaging (ATP) company.
Of course, the use of more advanced process node chip design companies will also have a corresponding increase in costs. In the end, the sum of the chip design cost and the ATP cost is equal to the total production cost, and it is concluded that the total cost paid for each 5nm chip will be as high as $426.
The reason why the cost of advanced process chips is increasing, it can not be ignored that the cost of semiconductor manufacturing equipment increases by 11% every year, and the design cost of each chip increases by 24%, which is higher than the growth rate of 7% in the semiconductor market.
In 2018, Taiwan Semiconductor Manufacturing Co Ltd CEO Wei Zhijia jokingly said that Taiwan Semiconductor Manufacturing Co Ltd expected to invest 25 billion US dollars in 5nm, and you will know what the price will be in the future!
Moreover, with the increase of the complexity of semiconductors, the demand for high-end talents is also increasing, which further pushes up the cost of advanced process chips. The effective number of researchers, dividing semiconductor research and development spending by the salaries of highly skilled researchers, increased 18-fold from 1971 to 2015, according to the report.
In other words, Moore's Law continues to increase a lot of investment and talent.
Lei Feng net(official account: Lei Feng net)It was previously reported that in order to support advanced processes, Taiwan Semiconductor Manufacturing Co Ltd's R & D staff has tripled in 10 years. In 2017, there were nearly 6200 R & D personnel, nearly twice as many as in 2008. These 6200 people are only engaged in research and development, not production.
How does the high cost of advanced processes affect the development of semiconductors and AI?
The semiconductor market is growing at a rate of more than 3% of the world economy. At present, the semiconductor industry accounts for 0.5% of the world's economic output. For the semiconductor industry, advanced processes and high-performance chips drive the progress of the industry, and transistors become the key.
The reduction of transistor size reduces the power consumption of each transistor, and the peak performance utilization of CPU doubles every 1.57 years until 2000. Since then, as the size of the transistor has declined and slowed, efficiency has doubled every 2.6 years, equivalent to a 30% increase in efficiency each year.
According to the report, Taiwan Semiconductor Manufacturing Co Ltd claimed that the speed increase and power consumption reduction brought about by node progress changed in a constant proportion from 90 nm to 5 nm, but the trend tended to be stable. Both Samsung indicators have a downward trend between 14 nm and 5 nm, but there is a lack of data at nodes greater than 14 nm.
However, due to the continuous increase in semiconductor equipment, R & D and other costs, a large number of wafer foundry factories are unable to participate in the production and competition of advanced processes. For example, GlobalFoundries does not produce chips under 14 nm.
The following table shows the time of mass production of each process node and the number of foundry factories. It can be seen that as the process moves forward, the number of fabs becomes smaller and smaller. At present, only TSMC, Samsung and Intel Corp are left in the advanced process foundry.
Although there are fewer and fewer contract manufacturers, the industry's demand for advanced processes has not diminished. AI chip has a great demand for advanced process. The most advanced AI chip is faster than the most advanced CPU, and has a higher AI efficiency algorithm.The efficiency of AI chips is a thousand times that of CPU, which is equivalent to the improvement of CPU26 years driven by Moore's Law.
This requires the continuous improvement of the transistor, and the improvement of the transistor continues, but the progress is slow. Thanks to the invention of FinFET transistor, Intel Corp launched the commercial 22nm FinFET in 2011, and the industry also pushed the semiconductor process forward from 22nm to 5nm based on FinFET.
When it comes to 2nm, Taiwan Semiconductor Manufacturing Co Ltd and Intel Corp all use GAA (Gate-all-around, surround gate) or GAAFET to maintain the performance improvement of advanced processes. Wei Zhijia revealed that for each generation of Taiwan Semiconductor Manufacturing Co Ltd process, the product speed and efficiency of customers will be increased by 30%-40%, and the power consumption can be reduced by 20%. 30%.
However, the high cost and the reduction of performance improvement make AI spend more time and money on computing, which becomes the bottleneck of AI development.
The report points out that the cost of training in AI laboratories is very high, with training costs estimated at between $50 million and $100m each for AlphaGo, AlphaGo Zero, AlphaZero and AlphaStar models.
It is worth noting that CSET is currently paying attention to the impact of advances in AI and advanced computing. According to the report, the US government is considering how to control AI technology, but because AI software, algorithms and data sets are not ideal control targets, hardware is the focus.
In the future, how many AI chips can afford to use advanced processes? How will the development of AI in the world be restricted?
Follow the official account of "Core Infrastructure" and reply to "5nm" to get the original English report.
Note: the chart in this article is from the report.Lei Feng net
The original article of Lei Feng net is forbidden to reprint without authorization. For details, please see the reprint instructions.