Japan chipmaker Rapidus opens arm in Nvidia's backyard from Nikkei Asia

[hskuo]

PALO ALTO, California/TOKYO -- Japanese chipmaking venture Rapidus has established a subsidiary in the heart of California's Silicon Valley, aiming to draw artificial intelligence companies as customers.

Rapidus Design Solutions is in the city of Santa Clara, home to big-name chip companies Nvidia and Intel.

Rapidus President Atsuyoshi Koike appeared at Thursday's news conference announcing the unit with Mukesh Khare, general manager of IBM Semiconductors. IBM and Rapidus are collaborating on technology development.

"Initially, clients from Silicon Valley will make up a considerable portion of the share," Koike told reporters. "It's critically important to engage in development together with startups."

Rapidus has named Henri Richard, a former marketing executive at Advanced Micro Devices and IBM, as president of Rapidus Design Solutions. The unit will soon begin full-fledged sales activities in Silicon Valley, making use of his experience and connections.

Nvidia had a market share of 80% or so in AI chips in 2022, according to British research firm Omdia. Nvidia focuses on design while entrusting production to Taiwan Semiconductor Manufacturing Co., the world's top chip foundry.

Rapidus has only slim chances if it tries challenging TSMC head-on. But the Japanese company stands to benefit from startups and other new players developing AI chips.

Sam Altman, CEO of ChatGPT developer OpenAI, is reportedly exploring semiconductor development. More competitors to Nvidia would lead to opportunities to offer contract manufacturing services.

Rapidus is collaborating with Canadian chip design company Tenstorrent to develop AI semiconductors. Tenstorrent has already announced Samsung Electronics as an AI chip foundry. Rapidus is forming a closer relationship with Tenstorrent with an eye toward future contract production.

Rapidus plans to accept small-lot orders along with larger ones and assist on quick production turnarounds. It aims to reduce delivery times to less than half of competitors' by using AI-assisted analysis to streamline manufacturing.

Wafers will be processed one by one in short periods. If a defect is found, Rapidus can immediately make corrections at the design stage. This will differ from the conventional approach of processing large volumes of wafers at once over a span of hours.

American technology giants like Microsoft and Amazon now develop semiconductors in-house. At an event hosted by Intel in February, Microsoft announced plans to outsource production of such chips to Intel.

Competition to manufacture AI chips on contract is heating up as they gain in importance as critical goods. AI chips will account for 16% of the overall semiconductor market in 2027, up from 10% in 2023, Statista data shows.

Rapidus aims to mass-produce advanced 2-nanometer chips from 2027. Japan's Ministry of Economy, Trade and Industry is providing Rapidus with the equivalent of billions of dollars in support toward this goal. Building a large sales network will be key to making the investment pay off.

Japan chipmaker Rapidus opens arm in Nvidia's backyard

Silicon Valley location looks to capture AI semiconductor demand among startups

asia.nikkei.com

------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Interesting strategy. Good luck to Rapidus.

 

[Daniel Nenni]

my concern:

On December 13, 2022, (Rapidus) announced a partnership with IBM . At this point, the company decided to purchase a license for manufacturing technology for GAA transistor process (2 nm generation process) products, for which mass production technology had not yet been established . [6] Although IBM had successfully developed a 2 nm product the previous year, it was not producing semiconductors on a large scale and was looking for someone to license its technology . [7] In terms of product development that takes advantage of miniaturization technology, South Korea's Samsung Electronics will begin mass production using the GAA transistor process (3 nm generation process) in 2022, and Taiwan's largest company , Taiwan Semiconductor Manufacturing Co., Ltd. (TSMC) will begin mass production of GAA transistors in 2025. The company has announced mass production of its process (2 nm generation process), and has decided to follow the leading companies . [8]

Rapidus - Wikipedia

ja.wikipedia.org

 

[MKWVentures]

didnt rapidus plan for low volume fab in 2025?

 

[Fred Chen]

didnt rapidus plan for low volume fab in 2025?

Yes they hope to do single-wafer processing: https://www.eetimes.com/rapidus-ceo-chasing-single-wafer-processing-dream/

 

[MKWVentures]

"“My idea is single-wafer processing,” Koike said. “It means the fab does not have any stock, so I can manufacture very precisely to control the cycle time.”"

Dream is the right word (single wafer processing, no inventory). I think I worked on a project like this years ago. It doesn't typically work as the setup times at tools are problematic. Things may have changes but I have not heard anyone say so.

I will be interested to see if this Rapidus partnership with IBM is realized. They should have a 2nm tooled out fab running in 2027? so shell in 2025 with all equipment ordered? is there a reel showing factory progress to date. They committed a pilot line in 2025

 

[Daniel Nenni]

What's left of IBM semiconductor is a research facility. They do bleeding edge process technology research for patents and licensing. IBM does not do high volume manufacturing. I would not bet on this joint venture actually competing with TSMC. Maybe something for niche low volume technologies.

 

[hskuo]

What's left of IBM semiconductor is a research facility. They do bleeding process technology research for patents and licensing. IBM does not do high volume manufacturing. I would not bet on this joint venture actually competing with TSMC. Maybe something for niche low volume technologies.

I am surprised by the low volume order. If Rapidus still use tradition lithography processes, then the mask set cost will be more than $20M per product. If the wafer price is ~$30k and with very very good yield (D0=0.1) for AI chips (large die, could be ~40 good die per wafer), then the die cost will be ~$1k. If output 1k wafers only, there will be ~40k die produced in the best case. The mask cost NRE will be $500/per die extra cost add-on. How much dies or wafers output needed to cover NRE and break-even? It is hard to be economically viable in small volume.

 

[Mooredaddy]

I am surprised by the low volume order. If Rapidus still use tradition lithography processes, then the mask set cost will be more than $20M per product. If the wafer price is ~$30k and with very very good yield (D0=0.1) for AI chips (large die, could be ~40 good die per wafer), then the die cost will be ~$1k. If output 1k wafers only, there will be ~40k die produced in the best case. The mask cost NRE will be $500/per die extra cost add-on. How much dies or wafers output needed to cover NRE and break-even? It is hard to be economically viable in small volume.

This was my immediate reaction. Its like the complete opposite of the TSMC model (economies of scale and massive data to scale yield). I'm trying to figure out in my brain how Rapidus will be economically viable? IBM is hardly a flex also. The image that comes to mind is of a artisian workshop for leading edge semis, but that just doesn't make sense economically. Very strange

 

[hist78]

This was my immediate reaction. Its like the complete opposite of the TSMC model (economies of scale and massive data to scale yield). I'm trying to figure out in my brain how Rapidus will be economically viable? IBM is hardly a flex also. The image that comes to mind is of a artisian workshop for leading edge semis, but that just doesn't make sense economically. Very strange

According to the comments from the Rapidus founder, it needs US$54 billion to start 2nm manufacturing. The investment, scale, customers, target markets, and manufacturing process may be a mismatch.

https://www.reuters.com/technology/japanese-chip-venture-rapidus-needs-54-bln-begin-production-says-chairman-2023-02-02/

 

[MKWVentures]

As I have said before, The Rapidus model does not make sense. Single wafer doesn't work (trust me, I have looked at it). Building a 2nm process in partnership with IBM to compete with TSMC in 2027 does not work. The financial investment model does not work

There may be something to Rapidus but it is different than what is reported. I am guessing a 16nm fab emerges in 2030 with 10 Japanese customers or something like that.

My recommendation for any new country (India) or company (Rapidus): Do an old technology, develop your company, partner with a successful company who does fab work. TSMC in Japan will be wildly successful. A Rapidus TSMC joint venture would we successful. A Rapidus UMC JV would even be successful.

 

[hist78]

As I have said before, The Rapidus model does not make sense. Single wafer doesn't work (trust me, I have looked at it). Building a 2nm process in partnership with IBM to compete with TSMC in 2027 does not work. The financial investment model does not work

There may be something to Rapidus but it is different than what is reported. I am guessing a 16nm fab emerges in 2030 with 10 Japanese customers or something like that.

My recommendation for any new country (India) or company (Rapidus): Do an old technology, develop your company, partner with a successful company who does fab work. TSMC in Japan will be wildly successful. A Rapidus TSMC joint venture would we successful. A Rapidus UMC JV would even be successful.

"A Rapidus TSMC joint venture would we successful. "

The problem is that Rapidus doesn't have too much value to offer in such joint venture. Such as:

1. Market access. TSMC and its longtime partners/customers (Toyota, Denso, Renesas, Sony, Fujitsu, etc.) already have the access, relationship, and volume.
2. Capital. It's not an issue at all. TSMC, MITI (Ministry of International Trade and Industry), and the TSMC's favorable borrowing position will solve it easily. On the other hand, most Rapidus' funding are coming from Japanese government.
3. Technology and IPs. Rapidus, as a new company by itself and licensing technology from IBM, has little to put onto the table.

 

[MKWVentures]

"A Rapidus TSMC joint venture would we successful. "

The problem is that Rapidus doesn't have too much value to offer in such joint venture. Such as:

1. Market access. TSMC and its longtime partners/customers (Toyota, Denso, Renesas, Sony, Fujitsu, etc.) already have the access, relationship, and volume.
2. Capital. It's not an issue at all. TSMC, MITI (Ministry of International Trade and Industry), and the TSMC's favorable borrowing position will solve it easily. On the other hand, most Rapidus' funding are coming from Japanese government.
3. Technology and IPs. Rapidus, as a new company by itself and licensing technology from IBM, has little to put onto the table.

Agreed. The deal would involved Rapidus providing cash, TSMC would create a deal where they control it and cannot lose money. Rapidus gets to claim they are a semiconductor Fab company. I have been involved with a couple of these agreements over the years.

But lets go back to original question and my comment.... How will Rapidus be successful given their announced plans?

 

[hist78]

Agreed. The deal would involved Rapidus providing cash, TSMC would create a deal where they control it and cannot lose money. Rapidus gets to claim they are a semiconductor Fab company. I have been involved with a couple of these agreements over the years.

But lets go back to original question and my comment.... How will Rapidus be successful given their announced plans?

If Rapidus is positioning itself as a national research and development project instead of a for profit startup, then it "will" achieve its own defined goals successfully. Although it can be as good as a gimmick.

 

[Jert]

Rapidus's intended differentiation is already clearly stated in their name "Rapid"us.

I don't think they or anyone else crazy enough to have any chance to start from scratch and catch up to TSMC in pricing and volume. Btw already two crazy giants Samsung & Intel chasing that impossible dream so why would Rapidus even consider trying? Would be the dumbest idea ever.

But I suppose there are plenty non-consumer products industries that do not require chips in billion-count volumes but also want advanced technology while cycle time and availability to get the chips is much more critical. Perhaps they (defense? aerospace? etc.) are willing to pay much much more than the typical billin-mobile-phone volume $1K per die cost/price?

 

[MKWVentures]

I don't think they or anyone else crazy enough to have any chance to start from scratch and catch up to TSMC in pricing and volume. Btw already two crazy giants Samsung & Intel chasing that impossible dream so why would Rapidus even consider trying? Would be the dumbest idea ever.

I think your inputs are correct and have lowered my confidence that Rapidus will be successful.

Their official plan is to supply high volume 2nm 2027. Lets see how that progresses

 

[hskuo]

Rapidus's intended differentiation is already clearly stated in their name "Rapid"us.

I don't think they or anyone else crazy enough to have any chance to start from scratch and catch up to TSMC in pricing and volume. Btw already two crazy giants Samsung & Intel chasing that impossible dream so why would Rapidus even consider trying? Would be the dumbest idea ever.

But I suppose there are plenty non-consumer products industries that do not require chips in billion-count volumes but also want advanced technology while cycle time and availability to get the chips is much more critical. Perhaps they (defense? aerospace? etc.) are willing to pay much much more than the typical billin-mobile-phone volume $1K per die cost/price?

No matter how many dies will be produced, the cost of mask set ( ~$20M) needs to be paid for the first try which might not work and need re-spin. If the small volume order means ~10 wafers, which will be ~400 good dies for large die and mature production, then the cost will be ($20M+$300k)/400 =$51K/die. If they go chiplet 3D design, it could be cheaper with higher yield in smaller die size, but can be still in ~$20-30k die cost range with " small volume" for very very good situation. The above good yield situation definitely will not fit Rapidus in the near future and the cost could be 5-10x higher at early stage. I would expect very very few or no private company which is very rich and can afford it.

 

[Mooredaddy]

No matter how many dies will be produced, the cost of mask set ( ~$20M) needs to be paid for the first try which might not work and need re-spin. If the small volume order means ~10 wafers, which will be ~400 good dies for large die and mature production, then the cost will be ($20M+$300k)/400 =$51K/die. If they go chiplet 3D design, it could be cheaper with higher yield in smaller die size, but can be still in ~$20-30k die cost range with " small volume" for very very good situation. The above good yield situation definitely will not fit Rapidus in the near future and the cost could be 5-10x higher at early stage. I would expect very very few or no private company which is very rich and can afford it.

Exactly. The economics of being leading edge don’t make sense unless you are producing a sh*t ton of wafers. Besides, how are you ever going to get to economical yields if you are running so few wafers. The upfront fixed cost is the same.