AMD Reportedly Shifts from Samsung to TSMC's 4nm Arizona Facility

[Nomad76]

AMD has reportedly decided to move its 4 nm chip orders from Samsung to TSMC's facilities in Arizona, United States. This is a significant loss for Samsung, as AMD had been working closely with Samsung on the SF4X process for EPYC server processors, Ryzen APUs, and Radeon graphics cards. It's worth noting that in May 2023, AMD announced the exact opposite—moving from TSMC to Samsung for some of its 4 nm CPUs. This partnership was an important part of AMD's strategy to use multiple manufacturers; however now appears to be failing. Reports indicate that AMD's decision comes from concerns about the stability and consistency of Samsung's manufacturing process, leading AMD to cancel plans for mass production of graphics chips using Samsung's 4 nm technology.

These changes in manufacturing strengthen TSMC's leading position in the chip manufacturing industry, even though relying on a single manufacturer creates geopolitical risks. AMD continues to deepen its relationship with TSMC, committing to use TSMC's advanced 2 nm technology for its upcoming "Venice" processors, which have completed testing at TSMC's Arizona facility and remain on schedule for release in 2026. Meanwhile, Samsung Electronics has reportedly achieved better-than-expected results in testing its SF2 (2 nm) process, with initial yields above 30%. The company plans to stabilize this process in the second half of 2025 to begin mass production of the Exynos 2600 mobile chip. Building on this progress, Samsung is reportedly close to securing an agreement to manufacture Qualcomm's Snapdragon 8 Elite 2 using its new 2 nm process.



View at TechPowerUp Main Site | Source

 

[TheinsanegamerN]

So reading between the lines, Samsung STILL cant get their process node working right.

 

[Quicks]

Samsung should just stick to memory and leave the rest to the big boys. Then again they have troubles with HBM memory as well.

 

[Denver]

Does it make sense to start development on a node and then cancel it years later? I don't think so.

4nm LPP+ is a mature and stable node, if it wasn't Samsung wouldn't be producing tens of millions of Exynos units for its Galaxy products. Either the initial rumor was false, or it was true and TSMC only offered a discount to keep AMD orders.

 

[TheinsanegamerN]

Does it make sense to start development on a node and then cancel it years later? I don't think so.

4nm LPP+ is a mature and stable node, if it wasn't Samsung wouldn't be producing tens of millions of Exynos units for its Galaxy products. Either the initial rumor was false, or it was true and TSMC only offered a discount to keep AMD orders.

Or the yields are terrible, and Exynos mobile chips are small enough to be *profitable* in a vertical monopoly sense, but anyone outside of samsung cant make money on it.

 

[londiste]

The way it is written is unclear. AMD was planning to manufacture CPUs and GPUs on Samsung 4nm. Report says they have chosen not to proceed on that with GPUs. No word on CPUs. The size difference there is significant so that will affect decisions.

In any case, we should be cheering for underdogs here. TSMC monopoly is not a good thing.

 

[Denver]

Or the yields are terrible, and Exynos mobile chips are small enough to be *profitable* in a vertical monopoly sense, but anyone outside of samsung cant make money on it.

Samsung's 4 nm process node achieves yield rates of 75 percent

Samsung's Foundry division has had a rough time holding onto customers over the past few years, mostly due to thermal and yield issues with its modern process...

www.techspot.com

https://www.notebookcheck.net/Exynos-2400-2200-2100-die-shots-highlight-the-chips-evolution-over-the-years.832760.0.html

They hit 75% years ago, today it should be much more stable. Exynos 2400(137.44 mm2) is twice the size of Zen5's CCDs, and also larger than the IO die in 6nm and much larger than the IO die would be in 4nm LPP.

Another possible reason is that Samsung's fabs in the US don't have the same capacity or yields as the one in South Korea.

 

[TheinsanegamerN]

They hit 75% years ago, today it should be much more stable.

Statistics are meaningless without context. 75% yield on what? On the exynos 2400? 2200? 2100? a 10mm2 test die?

Also, what power use was that yield? Was that yield seen on dies that pull 60+ watt, or ones that pull 10?

Exynos 2400(137.44 mm2) is twice the size of Zen5's CCDs, and also larger than the IO die in 6nm and much larger than the IO die would be in 4nm LPP.

You got me there.

Another possible reason is that Samsung's fabs in the US don't have the same capacity or yields as the one in South Korea.

Samsung has no 4nm fabs in the US, the CHIPS funding for the new 4nm facility was only approved a year ago, both to build the new Taylor plant and upgrade the Austin plant. Samsung's existing 4nm production lines are located at Hwaseong and Pyeongtaek in South Korea. Giheung is scheduled to be upgraded.

https://image.semiconductor.samsung.com/content/samsung/p6/semiconductor-us/sas/chips-and-science-act_240416/samsung-factsheet-chips-final.pdf

Manufacturing sites | Manufacturing | Samsung Semiconductor Global

Samsung Foundry has world-class facilities and is expanding worldwide to provide better service. Discover their manufacturing sites in Korea and the United States.

semiconductor.samsung.com

It would seem very silly to me that Samsung would finally get their act together, produce a great 4nm node, then totally fall asleep at the wheel and not bother building out significant capacity, especially when cutting edge nodes are so profitable.

Counter to this point: Samsung already has issues with slowing demand and excess fab capacity:

https://www.eetimes.com/samsung-slows-opening-of-texas-fab-despite-chips-stimulus/#:~:text=%E2%80%9CAnother%20worry%20for%20them%20is,backlog%20for%20the%20first%20quarter.%E2%80%9D

They slowed production of 4 and 5nm by over 50%

Exclusive: Samsung lifts foundry shutdown, Pyeongtaek fab to run at full capacity by June

Exclusive: Samsung lifts foundry shutdown, Pyeongtaek fab to run at full capacity by June Samsung Electronics restarts production as Exynos, crypto chip demand rises

www.chosun.com

So.....something isnt adding up. they have all this excess capacity and no yield problems, yet AMD continues to overpay for TSMC?

My guess would be that yield of "75%" only applies to their smallest low power chips, and trying to build the substantially more power hungry zen 5 CCDs on it tanked yields to the point that its not considered viable. It works for Samsung, between being lower power and not having to pay themselves as a middle man, similar to intel, but makes no financial sense for outside customers.

 

[alawadhi3000]

Does it make sense to start development on a node and then cancel it years later? I don't think so.

4nm LPP+ is a mature and stable node, if it wasn't Samsung wouldn't be producing tens of millions of Exynos units for its Galaxy products. Either the initial rumor was false, or it was true and TSMC only offered a discount to keep AMD orders.

The decision could be political, maybe its just a rumor, who knows.

The node that AMD is interested in is 4HPC not 4LPP+, it just started volume production this year so yields are probably not that good compared with previous 4 nm nodes like 4LPP+.

Samsung should just stick to memory and leave the rest to the big boys. Then again they have troubles with HBM memory as well.

Well if you have short memory then I'll remind you that a few years back Samsung had the best node, before that it was Intel. Things never stay the same.

 

[remixedcat]

Or the yields are terrible, and Exynos mobile chips are small enough to be *profitable* in a vertical monopoly sense, but anyone outside of samsung cant make money on it.

I've heard the opposite where the yeilds were excellent from Samsung's arizona facility... hmm... this has some other reason me thinks...

 

[blinnbanir]

Hmm TSMC Arizona is in the US. Tariffs anyone?