News Samsung inks $16.5 billion Tesla AI chip deal — Elon Musk says Samsung will produce new A16 chips: 'The strategic importance of this is hard to ove...

[Admin]

Samsung has secured a $16.5 billion deal to produce Tesla's future AI6 self-driving SoC at its Texas fab, a major strategic win for Samsung Foundry.

Samsung inks $16.5 billion Tesla AI chip deal — Elon Musk says Samsung will produce new A16 chips: 'The strategic importance of this is hard to ove... : Read more

 

[Notton]

Interesting, so they chose Samsung over Intel Foundries?
Isn't Samsung famous for failing at <5nm nodes, being unable to find customers outside of crypto ASICs, and also an underperforming and overheating Exynos?
Well, at least they shouldn't have to worry about overheating with a car sized device.

 

[bit_user]

Samsung agreed to allow Tesla to assist in maximizing manufacturing efficiency,
This is a critical point, as I will walk the line personally to accelerate the pace of progress. And the fab is conveniently located not far from my house.

Uh oh. I hope he doesn't DOGE it up. That's the last thing Samsung needs, right now.

I was going to say it's a shame Intel didn't get this deal, but now I'm not so sure.

 

[bit_user]

Well, at least they shouldn't have to worry about overheating with a car sized device.

Given that it's going in an electric car, efficiency is definitely important. Poor efficiency means a heavier and more expensive thermal solution, possibly wasting even more power on active cooling. All of the energy used to power and cool the electronics also means reduced range.

 

[Notton]

Given that it's going in an electric car, efficiency is definitely important. Poor efficiency means a heavier and more expensive thermal solution, possibly wasting even more power on active cooling. All of the energy used to power and cool the electronics also means reduced range.

Well, at least you don't have to worry about the range shortening in Winter because it doubles as a cabin heater.

 

[watzupken]

It is funny that Intel foundry is not very popular for some reason.

 

[bit_user]

Well, at least you don't have to worry about the range shortening in Winter because it doubles as a cabin heater.

Depends on where it's located and exactly how it's cooled. I wouldn't assume the waste heat goes into the cabin, but you might be right.

 

[Eximo]

Some quick math for you.

Let's say a 30W TDP.

Give the vehicle a nice 300 mile range with the lower end battery of 72kWh. (310 miles/500km, really) This gives the nominal Wh/Mile at 240W/per mile. (Typical Model 3 long range)

At 50 miles per hour we can run for 6 hours. That would put our chip power costs at 180Wh. Or a shocking loss of 3/4 of a mile, or just over a kilometer.

A slight breeze or running the headlights would be worse.

It all adds up, but I wouldn't worry too much about it.

 

[bit_user]

Let's say a 30W TDP.

You're going to drive a car using basically a 30 W GPU? I'd say more like 300 W, at least. Then, you need enough power to cool the thing on a hot day. So, probably at least another 25W for fans, because a giant radiator would add weight and that hurts efficiency.

Nvidia's AGX Thor chip, which is supposedly shipping this year, can do 500 TFLOPS of fp16, which is slightly more than a RTX 5090. Furthermore, they support a NVLink-type interconnect which enables you to use a pair. So, we're definitely talking about power consumption well into the hundreds of Watts, possibly up to about a kW (for a pair).

https://en.wikipedia.org/wiki/Nvidia_Drive#Blackwell_based
​

 

[Eximo]

Granted, that was more for the typical onboard computer. Tesla claims that the chip in the Cybertruck uses 72W, and the total system for FSD uses about 300W. I assume that means all the cameras as well. Though those would be running all the time anyway.

Still, that would take 7.5 miles off the range of a Model 3, for the Cybertruck, it would be more of an impact because the efficiency is worse. But the battery is bigger. Just add a little battery capacity to the design, increase efficiencies here and there and you would be none the wiser.

Though, having been in some of these cars with FSD, they also drive very conservatively and probably make up the difference between it and a human driver. Honestly, people make too much of range limitations. It only matters when there aren't sufficient charging stations, and generally, I have not found any issues getting between charging stations when traveling. And this was before my car could use Tesla chargers. 95% of the time you are charging at home. And 99% of the time traveling to places you have been before and know roughly how much charge you need to get there and back.

 

[thisisaname]

Neither Tesla nor Samsung Foundry disclosed details about the upcoming AI6 project, but considering the fact that Tesla tends to update its main SoCs every three to four years and the AI5 processor is expected to land in Tesla cars sometime in 2027, one can expect AI6 to arrive in 2029 at the earliest.

I would think 2030 at the earliest.

 

[thestryker]

I was going to say it's a shame Intel didn't get this deal, but now I'm not so sure.

Yeah when I saw a headline about this last night I thought "interesting Intel didn't cut whatever deal was necessary" until I read more. This seems like a reflection on how dire things are for Samsung Foundry.

 

[S58_is_the_goat]

Interesting, so they chose Samsung over Intel Foundries?
Isn't Samsung famous for failing at <5nm nodes,

What node is intel on now? 10nm?

 

[Notton]

What node is intel on now? 10nm?

Intel Foundries has 18A, which is 1.8nm-class and it should have entered high volume production by now. We probably won't see the final product until 2026.

Which is a lot faster than what Tesla is asking, a 2nm-class by 2029.

 

[John Nemesh]

You're going to drive a car using basically a 30 W GPU? I'd say more like 300 W, at least. Then, you need enough power to cool the thing on a hot day. So, probably at least another 25W for fans, because a giant radiator would add weight and that hurts efficiency.

Nvidia's AGX Thor chip, which is supposedly shipping this year, can do 500 TFLOPS of fp16, which is slightly more than a RTX 5090. Furthermore, they support a NVLink-type interconnect which enables you to use a pair. So, we're definitely talking about power consumption well into the hundreds of Watts, possibly up to about a kW (for a pair).

​

https://en.wikipedia.org/wiki/Nvidia_Drive#Blackwell_based​

​

1) Tesla cars have about 73-80Kwh batteries in them, so power usage, as a percentage of your battery, is going to be minimal.

2) cooling is a non issue, they have VERY sophisticated heat pumps in the vehicles, which is used to regulate the temp on their computers. Excess heat is diverted to the cabin or to other components like the batteries....or just vented out.

 

[S58_is_the_goat]

Intel Foundries has 18A, which is 1.8nm-class and it should have entered high volume production by now. We probably won't see the final product until 2026.

Which is a lot faster than what Tesla is asking, a 2nm-class by 2029.

Until the new ceo cancels it, i wouldn't get my hopes up until its actually out in full production.

 

[ezst036]

Well, at least you don't have to worry about the range shortening in Winter because it doubles as a cabin heater.

Now if only we could find a way to make CPUs smart enough to not heat up in the summer at all and only heat up in the winter. That would be great.

 

[bit_user]

Now if only we could find a way to make CPUs smart enough to not heat up in the summer at all and only heat up in the winter. That would be great.

You can underclock it in the summer. For a self-driving chip, that means top speed will be limited to 40 mph in warm weather.
: D

 

[hax0red1]

Cooling the chip could be easily integrated into the glycol/water inverter, motor, battery cooling. During the winter you would get a boost in heater as well.

 

[phead128]

It is funny that Intel foundry is not very popular for some reason.

The reason why Intel foundry has no significant external customers is because Reuters spread fake news, biased foreign propaganda, and greedy stock market manipulators. Intel will be fine because TSMC/Samsung is capacity limited, US govt will bail out Intel, Intel got high-NA EUVL, Taiwan nuclear WW3, and Pat G. said yields were good at this stage of development.

Did I capture every cope correctly? The same people who believe Samsung low yields claims as plain fact , are the same people do every mental gymnastics possible to deny yield challenges at Intel. You can't reason with fanbois.

 

[phead128]

Intel Foundries has 18A, which is 1.8nm-class and it should have entered high volume production by now. We probably won't see the final product until 2026.

Which is a lot faster than what Tesla is asking, a 2nm-class by 2029.

Here's a clue: Intel foundry will be spinned off as a separate entity or sold for parts by 2029 at this rate. Customers don't just chase the latest nodes for sake of process leadership, they also need a reliable proven partner that actual exists in couple years and are not insolvent or US govt welfare queens (which says a lot for Musk to avoid, he knows a loser when he sees one).

 

[bit_user]

The reason why Intel foundry has no significant external customers is because Reuters spread fake news,

OMG, are you seriously blaming Intel's failures on the fake news? Is that why Arrow Lake sucks, too??

Nobody in a position to seriously consider using their node would care about that. They would definitely at least talk to Intel, themselves.

The same people who believe Samsung low yields claims as plain fact , are the same people do every mental gymnastics possible to deny yield challenges at Intel. You can't reason with fanbois.

Samsung's rumors of low yields are in line with their business performance and public statements.

Customers don't just chase the latest nodes for sake of process leadership, they also need a reliable proven partner that actual exists in couple years

Well, there's definitely a trust deficit, given how Intel cancelled 20A and scaled back external offerings of 18A. With all of the layoffs, there's reason for concern over whether they'll hit their milestones on 14A or could go ahead and cancel it outright. If you're looking to get chips fabricated, you don't want a phone call mere months before the product ramp is scheduled to happen, telling you that you need to find a new fab partner and setting you back to square 1.

which says a lot for Musk to avoid, he knows a loser when he sees one).

TBH, I think it could be as much or more about cost. We've been hearing about how TSMC thinks High-NA is too expensive, which obviously could be just them blowing smoke. However, there might be something to it, and maybe it's disadvantaging 14A on price.