Tesla Moves to TSMC with Autonomous Chips

[Daniel Nenni]

It is reported that Tesla can lead the industry in building fully self-driving electric vehicles, and TSMC is a key player. It manufactures new FSD chips for Tesla, making the "brain" of electric vehicles smarter.

TSMC has never commented on single customers and market rumors. However, at the previous TSMC North American Technology Forum, Tesla's senior executives appeared to share the successful case of cooperation with TSMC, which has convinced the industry of the relationship between the two companies' cooperation in advanced chips.

According to industry insiders, Tesla’s old FSD is produced by Samsung’s 14nm process, and will be upgraded to Samsung’s 7nm process, and it is considering upgrading to 5nm. Due to design upgrades and comprehensive consideration of mass production quality and production scale expansion, it is determined to switch to TSMC's 5nm family (including 4nm) process production, and Samsung will produce old chips and memory for partial support.

Recently, a lot of foreign and foreign data have shown that Tesla’s new FSD chip is gradually being introduced into Model Y and other car models. The data of a few Model Y sold in Texas shows that it can support the latest FSD HW4 by the end of May this year. In addition, according to the U.S. Tesla sales information website, a small number of Model X models sold in Miami support the latest FSD HW4.

Musk previously said on an investor call that the cost of upgrading from HW3 to HW4 is significant. As Musk announced that Tesla is expected to achieve fully automatic driving later this year at the earliest, the industry is optimistic that Tesla's fully self-driving cars will inevitably trigger a new wave of buying momentum, injecting strong momentum into TSMC's automotive business.

TSMC is also quite confident in the introduction of its own advanced manufacturing process into automotive applications. Zhang Xiaoqiang, Senior Vice President of Business Development, recently revealed that automotive customers are eager to promote 3nm process technology. Therefore, TSMC launched the N3AE solution for customers to design and use, which will help Customers reduce time to market by two to three years.

特斯拉 CEO 馬斯克喊實現全自駕 台積成大功臣 | 產業熱點 | 產業 | 經濟日報

特斯拉執行長馬斯克昨（6）日表示，特斯拉最快今年稍晚有望實現全自動駕駛。業界分析，特斯拉領先業界達成全自駕電動車界新里程...

money.udn.com

 

[dlip410]

I wonder how the adoption of this new technology will effect the life cycle of a Tesla. Is there a robust, automotive grade, 5nm ME already available. Cybersecurity may also be a concern. How vulnerable is this new SOTA tech from external attacks.

 

[milesgehm]

We hear this every year "Tesla is expected to achieve fully automatic driving later this year"

 

[blueone]

We hear this every year "Tesla is expected to achieve fully automatic driving later this year"

Even if they do, I can't visualize self-driving Level 4 or 5 software interacting with human drivers on a large-scale basis. Especially when there are large numbers of humans and self-driving vehicles on the road simultaneously. Humans violate traffic laws a lot. Self-driving cars won't ever be allowed to. It's a recipe for road rage from human drivers.

We were just driving on I-25 in Colorado over the weekend, including the Denver area. The speed limit is 75mph for most of it. 75mph is nowhere near the average speed. I didn't do a formal sample, but 90mph felt more like it, and weaving in and out of lanes and tailgating were common. Some of this behavior was in the numerous miles of construction zones. We didn't see even one police officer in about 400 miles of driving on that road. I tried to imagine self-driving cars in that mix, obeying traffic laws. I just can't.

Perhaps with dedicated lanes, but I doubt that concept will fly.

ADAS I believe in, FSD... I'm a deep skeptic.

 

[Xebec]

Intel is the big loser with Tesla it seems.

Dojo supercomputer? Nvidia cards.
Infotainment for 2M vehicles/year? AMD
FSD and other custom chip fabrication? Not Intel

This is another exponential growth story Intel is missing on..

 

[blueone]

Intel is the big loser with Tesla it seems.

Definitely not aligned with Tesla.

Dojo supercomputer? Nvidia cards.

I'd bet on Nvidia too, for now, but supercomputing is an evolutionary thing.

Infotainment for 2M vehicles/year? AMD

Until RISC-V unseats them both.

FSD and other custom chip fabrication? Not Intel

This is another exponential growth story Intel is missing on..

As I posted, I'm an FSD non-believer. Evolving and advancing ADAS is the future, IMO, and Mobileye is probably the best strategy in ADAS.

 

[hist78]

We hear this every year "Tesla is expected to achieve fully automatic driving later this year"

Don't overlook the "at the earliest" part of Mr. Musk's statement.

"As Musk announced that Tesla is expected to achieve fully automatic driving later this year at the earliest"

 

[hist78]

Intel is the big loser with Tesla it seems.

Dojo supercomputer? Nvidia cards.
Infotainment for 2M vehicles/year? AMD
FSD and other custom chip fabrication? Not Intel

This is another exponential growth story Intel is missing on..

In April it was reported/rumored that Samsung won the contract to manufacture chips for Intel's Moboileye.

IFS wasn't selected to be the foundry partner for its sister company Moboileye. Isn't it strange?

 

[nghanayem]

In April it was reported/rumored that Samsung won the contract to manufacture chips for Intel's Moboileye.

IFS wasn't selected to be the foundry partner for its sister company Moboileye. Isn't it strange?

Like most acquisitions, I don't think they were ever on intel nodes, but I could be wrong. At least with the Q4 being on a "28nm FDSOI" process it is safe to say it sure as hell wasn't on any TSMC, Intel, or UMC process.

 

[Xebec]

As I posted, I'm an FSD non-believer. Evolving and advancing ADAS is the future, IMO, and Mobileye is probably the best strategy in ADAS.

Fwiw the backup plan to FSD may just be AGI...

Whether it pans out or not - Right now every Tesla ships with two FSD chips built on an advanced node, and it looks like they'll keep growing those shipments for several more years. Thats yet another large amount of non Intel silicon on a growing market.

(Also these same chips do ADAS from what I can tell. It's a term I'm not fully familiar with - it seems to just be "safety features not including full autonomy").

 

[blueone]

Fwiw the backup plan to FSD may just be AGI...

Whether it pans out or not - Right now every Tesla ships with two FSD chips built on an advanced node, and it looks like they'll keep growing those shipments for several more years. Thats yet another large amount of non Intel silicon on a growing market.

In Intel terms, Tesla's volumes are not "a large amount of silicon". Tesla sales are on track to be two million units worldwide for 2023, which means four million chips. Compare that to CPU volumes, even in Intel's diminished state.

(Also these same chips do ADAS from what I can tell. It's a term I'm not fully familiar with - it seems to just be "safety features not including full autonomy").

Exactly. ADAS includes features like automatic emergency braking, pedestrian detection, lane discipline warnings and correction, drowsiness detection, adaptive cruise control, parking assist and automation, among a longer list of features. Based on what I've read, ADAS features are improving at a good pace, but they are simple to implement compared with FSD. I'm thinking if ADAS is like launching an earth-orbit satellite, Level 5 FSD is like a manned mission to Mars.

 

[Xebec]

In Intel terms, Tesla's volumes are not "a large amount of silicon". Tesla sales are on track to be two million units worldwide for 2023, which means four million chips. Compare that to CPU volumes, even in Intel's diminished state.

Exactly. ADAS includes features like automatic emergency braking, pedestrian detection, lane discipline warnings and correction, drowsiness detection, adaptive cruise control, parking assist and automation, among a longer list of features. Based on what I've read, ADAS features are improving at a good pace, but they are simple to implement compared with FSD. I'm thinking if ADAS is like launching an earth-orbit satellite, Level 5 FSD is like a manned mission to Mars.

Understand - 4 million chips @ 260mm2 each isn’t the biggest order of chips (though it’s not tiny), but I was just pointing out these are early days and it appears to be another exponential curve (in general) that Intel isn’t getting much action on. Tesla looks to be compounding 40-50% annually and is talking about selling their chips to other OEMs now. Intel missed the iPhone, and quite a a few other markets that turned out to be painful 5 years later. FSD (or ADAS) chips could be an area of opportunity that doesn’t seem headed for Intel foundry today. I hope that changes..

Got it re: ADAS. I think the same chips doing FSD are effectively doing ADAS though, at least on Tesla vehicles - from a silicon perspective, do you think we’ll see different ADAS silicon and is Intel (Foundry) positioned to take that market?

(A bit off topic - but on the FSD side of things. I’ve owned a Tesla Model 3 for 5 years, but have been a strong skeptic on FSD being real this decade. However, the Lex Fridman interview of John Carmack has me softening a bit as he had some very strong arguments on why FSD could happen by 2030 as there are several paths to get there, including AGI. The AGI argument he has was that he posits “looking back we probably have all of the code now, just not together and not in the right sequence”. It’s a bit of a reach but he is someone I think knows the space pretty well - there’s a lot more supporting argument than I remember here. If you want a good interview on software development, AI, history of computers/software, etc — the Lex Fridman John Carmack interview podcast is worth the 2.5-3 hour listen).

 

[blueone]

Understand - 4 million chips @ 260mm2 each isn’t the biggest order of chips (though it’s not tiny), but I was just pointing out these are early days and it appears to be another exponential curve (in general) that Intel isn’t getting much action on. Tesla looks to be compounding 40-50% annually and is talking about selling their chips to other OEMs now. Intel missed the iPhone, and quite a a few other markets that turned out to be painful 5 years later. FSD (or ADAS) chips could be an area of opportunity that doesn’t seem headed for Intel foundry today. I hope that changes..

I think IFS is still working to position itself within Intel (to design groups) and to potential foundry customers. I have not personally seen evidence of progress in either realm yet that makes me hopeful a clear direction has been set.

Got it re: ADAS. I think the same chips doing FSD are effectively doing ADAS though, at least on Tesla vehicles - from a silicon perspective, do you think we’ll see different ADAS silicon and is Intel (Foundry) positioned to take that market?

To be honest, I'm not sure yet what IFS is positioned to really do.

(A bit off topic - but on the FSD side of things. I’ve owned a Tesla Model 3 for 5 years, but have been a strong skeptic on FSD being real this decade. However, the Lex Fridman interview of John Carmack has me softening a bit as he had some very strong arguments on why FSD could happen by 2030 as there are several paths to get there, including AGI. The AGI argument he has was that he posits “looking back we probably have all of the code now, just not together and not in the right sequence”. It’s a bit of a reach but he is someone I think knows the space pretty well - there’s a lot more supporting argument than I remember here. If you want a good interview on software development, AI, history of computers/software, etc — the Lex Fridman John Carmack interview podcast is worth the 2.5-3 hour listen).

I'm assuming you're speaking of Artificial General Intelligence. Correct? I have to admit, I'm personally very skeptical of AGI development, though I am not an expert in the field. Just my perhaps completely incorrect view, I think the human-like capabilities of LLMs are making the entire AI/ML field look more capable than it really is.