Feature articles |
By Nick Flaherty
Elon Musk has ambitious plans for the semiconductor industry. To secure the supply chain for AI chips for his companies, Tesla, SpaceX and xAI, he is proposing a series of fabs, the Terafab, that will provide a terawatt of processing performance a year in the next decade.
Musk has a history with this kind of approach that can disrupt an industry. Back in 2015 he teamed with Panasonic to set up a battery factory in Nevada, creating the concept of a battery ‘gigafactory’ in the process. He has taken the technology and expanded with battery gigafactories in Texas, Shanghai and Berlin. Now every battery plant for electric vehicles is a gigafactory.
The initial $25bn chip plans centre arounds building a semiconductor fab in Texas, although the Tesla gigafactory in Austin doesn’t have enough room or power supply for a fab.
One key question was who the semiconductor technology partner would be, and that turns out to be Intel. The choice was limited, with TSMC in Taiwan, Samsung, nominally in South Korea as the only other technology suppliers able to provide the technology for sub-2nm chip-making technology. While both have US fabs, with Samsung in Texas and TSMC in Arizona, the technology is controlled from other countries.
So Intel, which is now part-owned by the US government, is a suitable candidate. It also has key packaging technology with its Foveros and EMIB and is working on a glass substrate technology that will help to reduce the AI power consumption.
The other possible technology supplier would be IBM, which is supplying 2nm technology to Rapidus in Japan. The Rapidus deal, detailed in eeNews Europe over last three years, demonstrates what is possible with sufficient political will and financial backing. The company is now delivering prototype chips with the IBM process and is planning the same kind of end-to-end facility with design, fabrication and advanced packaging, that Musk is proposing for the Terafab.
Musk’s deal with Intel gives some insight into the process technology that will be used. This is important as the process development kit (PDK) is an integral part of the chip design process. Intel’s backside power delivery and RibbonFET are key elements for the 14A process that would be suitable in the timeframes for a new fab, ie two to three years.
However the key limitation is the equipment. Intel has been a strong proponent of leading edge lithography with extreme UV (EUV) light and optics with a high numerical aperture (NA) with equipment developed and produced by ASML in the Netherlands. The lead time for this leading edge equipment, where each TWINSCAN EXE:5200 costs $350m, is several years. Just ten systems are planned for delivery in 2027 to Intel, Samsung and sk Hynix. It is possible that Terafab could take Intel’s order to speed up the facilitation of the fab, which would be a significant boost to Intel’ balance sheet.
This technology is seen as critical to the US national security, with a high NA system planned for the Albany research centre in New York state.
At the same time there are similar long lead times for key equipment from Applied Materials, Lam Research and Tokyo Electron. All these companies provide equipment for material processing at the leading edge, and Musk is reported to be calling around to speed up the timeframe.
Rapidus highlights the challenge. It installed ASML’s NXE:3800E EUV machine, the first in Japan, in December 2024 at their IIM-1 foundry in Chitose, Hokkaido, for the 2nm pilot line, with the process fully proven by IBM. But volume production is still not expected until 2027.
Texas is a good place to set up a fab. There is a strong ecosystem of suppliers, including GlobalWafer for the supply of 300mm silicon wafers, to supply the competition. Texas is home to fabs from Texas Instruments, Samsung and NXP, but all of these are producing legacy process technologies. That has a different set of suppliers to the leading edge of 14A and below that Terafab will require. However the prospect of a leading edge fab cluster in the state with the backing of Musk will encourage the ecosystem to expand.
The initial aim of the Terafab is to deliver chips for both AI training and inference. The current DoJo D1 AI chip consists of 25 chiplets built by on a 7nm process by TSMC in Taiwan for training the AI system used in Tesla cars and trucks. Including SpaceX in the deal implies that the design of the next generation chips will the focussed on AI chips that can operate in orbit to address the power challenges.
Musk sees a fab with more automation and limited intervention, even using humanoid robots. This will address one of the other challenges around skills. The semiconductor industry globally is facing a significant shortage of skilled staff that would limit the ability to scale up production. The process is already highly automated, with critical processes in highly controlled environment. The people are needed for controlling the equipment, a process that can also be automated.
The number of fabs needed to hit the terawatt target is one question as this determines how much space and power would be required at the site. This depends on the mix of chips for training and inference. The current DoJo D1 chips consume 400W, while the latest Rubin GPUs from NVIDIA are hitting the 1kW power envelope for an individual chip the size of the reticle used in the lithography equipment.
There are a number of unknown factors in hitting that terawatt target. The system architecture will determine the power requirements and the number of chips required. A chiplet-based design on a glass substrate with high bandwidth memory will have different yields from reticle-sized chips. The recently announced deal with AI startup Cursor will also impact on the architecture of the chips.
The number of wafers that can be processed will be determined by the equipment, particularly the high NA lithography, and that will then determine the amount of AI processing delivered by each fab. Numbers such as 100 million to 200 million chips per year are being suggested, which implies the chiplet approach but also implies multiple fabs.
So Terafab has a number of challenges. But the choice of Intel has a strong echo with the past. Much of Intel’s success was based on the idea of ‘copy exactly’. When a new process was proven in Portland, Oregon, it was copied exactly in other fabs, from Chandler, Arizona, to Ireland to Israel.
This approach will serve Musk well for multiple fabs to meet the volume target, and with the Nevada gigafactory he demonstrated the ability to drive through the commissioning of end-to-end production processes. Using this approach will yield significant benefits for the rest of the industry to boost output within the current constraints of space, skills and power.
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。