News TSMC to Start Dedicating New Capacity to Auto Chips First

[Admin]

According to the Taiwan Economic Ministry, TSMC will start prioritizing auto chips as it adds new capacity.

TSMC to Start Dedicating New Capacity to Auto Chips First : Read more

 

[bigdragon]

That's not right. The automotive manufacturers should have to play by the same rules as other companies. Auto companies shouldn't be able to cut the line and get their production up faster. Being shortsighted and pinching pennies has consequences. It's not TSMC's fault when Just In Time (I like to call it Not In Time) bites these manufacturers hard for making a forecasting mistake.

 

[caseym54]

What in God's name do automobile manufacturers need 7nm parts for? They should be perfectly happy with what Intel can make.

 

[cryoburner]

The article states that the auto chips will be given priority for any new capacity that opens up, but that doesn't mean they'll be using a majority of it, just that they won't be limited by the companies that are using a much larger share of capacity for things like CPUs, GPUs and mobile chips. If the auto industry only accounted for around 3% of TSMC's sales last year, I wouldn't expect that to increase dramatically this year. But they can provide those companies with some assurance that they won't be limited by larger customers squeezing them out of any increased production they require. They're probably doing that to avoid potentially losing them as customers. Due to their relatively low share of the production capacity though, I wouldn't expect this to have much effect on the other, larger customers though. Companies like AMD and Nvidia will still be getting the manufacturing capacity they requested, just if a little more happens to open up somewhere, the auto industry may be given first dibs on it.

 

[hotaru251]

That's not right. The automotive manufacturers should have to play by the same rules as other companies. Auto companies shouldn't be able to cut the line and get their production up faster. Being shortsighted and pinching pennies has consequences. It's not TSMC's fault when Just In Time (I like to call it Not In Time) bites these manufacturers hard for making a forecasting mistake.

business is about profit.

they pick who they think is most profitable.

 

[Chrys]

What in God's name do automobile manufacturers need 7nm parts for? They should be perfectly happy with what Intel can make.

High performance specialty chips for Lidar, Cameras (Backup, Side), and Image Processing

business is about profit.

they pick who they think is most profitable.

TSMC is being pressured by the government. I don't think their pricing varies much(possibly at all) anymore after they got rid of all discounts and increased their prices.

 

[Chung Leong]

What in God's name do automobile manufacturers need 7nm parts for? They should be perfectly happy with what Intel can make.

They don't. Embedded chips are usually fabbed on outdated nodes. That's where the shortage is and capacity will only shrink with time. Car-makers can't switch to new chips overnight so TSMC is trying to get other customers to do so.

 

[jkhoward]

That's not right. The automotive manufacturers should have to play by the same rules as other companies. Auto companies shouldn't be able to cut the line and get their production up faster. Being shortsighted and pinching pennies has consequences. It's not TSMC's fault when Just In Time (I like to call it Not In Time) bites these manufacturers hard for making a forecasting mistake.

I agree. I’d wager that computers are far more impactful on the world then fancy cars. We’ve been building cars without chips since the 70’s. If there’s truly a shortage in chip production then the auto companies need to work around the shortages like everyone else. Shoot, my new car with its fancy chips has loads of software issues like the freaking touch screen not working or detecting a collision when I’m going down the road lol.

 

[InvalidError]

That's not right. The automotive manufacturers should have to play by the same rules as other companies.

There are two sets of rules: regular orders where you get allocated a production window in the back-order queue for a nominal price and rush orders that get squeezed into whatever spots open up and go to the highest bidder which may include compensation to clients for shuffling their delivery windows to accommodate the rush order.

Someone in the automotive industry probably put in a chunky bid for TSMC's slack. so TSMC is scrambling to make room on its production calendar.

 

[Endymio]

Anyone (including the article's author) who reads this announcement as being a statement that auto chips are going to get the bulk of new nodes like 3nm or even 5nm is blowing smoke. They'll get some of the new capacity expansion at older nodes only.

 

[caseym54]

High performance specialty chips for Lidar, Cameras (Backup, Side), and Image Processing

Still, you don't need 7, or even 10nm to do those things. If they are going to use older lines then why TSMC? There are places with perfectly fine 14, 22 and 32nm processes that would be fine for most things. Even Lidar processing can be done at 14nm, and cameras for visual use need nothing of the sort.

 

[InvalidError]

Still, you don't need 7, or even 10nm to do those things.

For all of those things taken separately as their own independent driver assist/warning thing in cars with no self-driving capabilities whatsoever, sure. For newer higher-end models designed to possibly achieve full self-driving though, processing power requirements are orders of magnitude higher.

 

[Endymio]

For newer higher-end models designed to possibly achieve full self-driving though, processing power requirements are orders of magnitude higher.

Tartan Racing in 2007 won the DARPA autonomous vehicle challenge with a fully self-driving vehicle built with chips fabbed on 90nm - 250nm nodes.

Obviously a production model needs smaller, cheaper, and at least somewhat faster chips -- but the high cost of cutting-edge nodes rules out their widespread use in consumer vehicles. Automakers will certainly lag two or three generations behind the most advanced process technology.

 

[InvalidError]

Tartan Racing in 2007 won the DARPA autonomous vehicle challenge with a fully self-driving vehicle built with chips fabbed on 90nm - 250nm nodes.

Driving on a mostly barren test track with a mountain of equipment towering over the roof which gives it a huge FoV advantage is a whole different challenge from self-driving in a packed urban environment using sensors integrated in the vehicle body that severely limit each sensor's FoV. You need several times as much processing power to stitch together inputs from multiple extra sensors at a higher frequency and shorten reaction time.

In August 2020, Tesla announced that its 7nm FSD 4.0 chip will be over 3X more powerful than its 3.0 hardware which is itself 20X more powerful than its Nvidia-powered FSD (1/21th the time per frame) and would enter engineering sample production on TSMC 7nm in late 2020, volume production in late 2021 and be in its new cars in 2022.

When you charge $6000+ for FSD, you can afford the cost of a pair of ~250sqmm 7nm FSD chips and even pay a premium to jump the queue.

 

[Endymio]

Driving on a mostly barren test track ...is a whole different challenge from self-driving in a packed urban environment...

From the DARPA website: "The DARPA Urban Challenge was held on November 3, 2007, at the former George AFB in Victorville, Calif. ... this event required teams to build an autonomous vehicle capable of driving in traffic, performing complex maneuvers such as merging, passing, parking, and negotiating intersections...."

Your points about FOV and the limitations of the test environment are valid, but overstated.

In August 2020, Tesla announced that its 7nm FSD 4.0 chip will [enter] volume production in late 2021

By Q4 2021, TSMCs N7 node will be 3.5 years old, and even N5 will be nearly two years old. As I said, automakers are not going to use the most advanced nodes for these chips.

Also, according to Tesla's own statements, the driving impetus for the shrink is as much power-related as performance. The last thing you want in an electric car is chips sucking down several hundred watts.

 

[InvalidError]

Also, according to Tesla's own statements, the driving impetus for the shrink is as much power-related as performance. The last thing you want in an electric car is chips sucking down several hundred watts.

Which would be one more reason to more aggressively push for a smaller lower-power process instead of sticking to generations-old stuff, get more stuff done more reliably, quickly and accurately within the same 100W FSD computer budget.

 

[cryoburner]

From the DARPA website: "The DARPA Urban Challenge was held on November 3, 2007, at the former George AFB in Victorville, Calif. ... this event required teams to build an autonomous vehicle capable of driving in traffic, performing complex maneuvers such as merging, passing, parking, and negotiating intersections...."

I think there's still a bit of a difference between having one vehicle driving down the road under careful supervision through an urban course with predefined waypoints at a 14mph average speed (30mph top speed), and having thousands of autonomous vehicles driving around to arbitrary locations within a wider range of environments, at higher speeds, while their "driver" is doing their nails and playing Flappy Bird. : P

The requirements for avoiding a serious accident in that competition were undoubtedly a lot lower, and didn't need to really deal with things like adverse whether conditions and night-driving.

 

[Endymio]

Which would be one more reason to more aggressively push for a smaller lower-power process instead of sticking to generations-old stuff, get more stuff done more reliably, quickly and accurately within the same 100W FSD computer budget.

Yes -- except that cell phone socs have a far smaller power budget, and thus outbid everyone else for those cutting-edge processes.

Your own example of Tesla makes the point for me. They're going into production with a 3.5 year old node, and they'll likely stick with that node for at least 18 months, by which time it'll be five years old -- not simply elderly, but octogenarian, in the litho world.

The requirements for avoiding a serious accident in that competition were undoubtedly

Undoubtedly. The car was also using a single-core 2006-era cpu. And not only have chips advanced dramatically in the last 15 years, but so have the sensors and behavioral software.