AMD to Shift Some of its 4 nm CPU Silicon-fabrication to Samsung from TSMC

[btarunr]

AMD has reportedly signed up with Samsung Electronics to shift some of its 4 nm processor silicon fabrication from TSMC. The apex Taiwan-based foundry is reportedly operating at capacity for its 4 nm-class nodes, with customers such as Apple and Qualcomm sourcing 4 nm mobile SoCs on the node, leaving AMD with limited allocation and/or bargaining power with TSMC. The company relies on 4 nm for its Ryzen 7040 series "Phoenix" mobile processors, and is in the process of adapting its design for Samsung's 4 nm-class nodes (of which there are five types for AMD to choose from).

Switching to Samsung probably gives AMD more scalability, particularly given that "Phoenix" has missed its release timeline, leaving AMD with the 5 nm + 6 nm Ryzen 7045 series "Dragon Range" MCM in the premium segments, and older 6 nm 7035 series "Rembrandt-R" in the mainstream and ultraportable segments, but nothing "apt" to compete against Intel "Raptor Lake-U" and "Raptor Lake-P." AMD has a limited window in which to ramp up "Phoenix," as Intel readies "Meteor Lake" for a 2H-2023 debut, with a focus on mobile variants.



View at TechPowerUp Main Site | Source

 

[dgianstefani]

I wonder if we'll see different performing variants of the same SKU, or whether specific models will be made on different nodes.

 

[dir_d]

Taking a wild guess but the 5 nm + 6nm approach is not going well so they went to Samsung for a unified 4 nm part.

 

[Denver]

This should be very expensive and unfeasible, unless the volume is so absurd that it pays off...there would also be loss of efficiency with the change.

 

[sLowEnd]

This reminds me of when Apple used to dual source from TSMC and Samsung for their A9 chip.

iPhone 6s: Samsung And TSMC A9 SoCs Tested

The A9 SoC inside the iPhone 6s comes from two different vendors (Samsung and TSMC) using two different FinFET processes (14nm and 16nm, respectively). We test both versions to see if there's any power or performance differences.

www.tomshardware.com

Hopefully the differences between the two sources end up minimal for consumers.

 

[Count von Schwalbe]

Taking a wild guess but the 5 nm + 6nm approach is not going well so they went to Samsung for a unified 4 nm part.

The Phoenix APUs are monolithic.

 

[yeeeeman]

so they'll basically make phoenix on a 7nm TSMC like process. Good luck with it amd. We know for a fact what this results in, from qualcomm. But maybe you can do some magic and remove the crapiness from samsung fab process.

 

[Denver]

so they'll basically make phoenix on a 7nm TSMC like process. Good luck with it amd. We know for a fact what this results in, from qualcomm. But maybe you can do some magic and remove the crapiness from samsung fab process.

Samsung's 4nm LPP looks very similar to TSMC's 5nm. It must be about 10% worse than TSMC 4nm.

 

[sethmatrix7]

When Korea may actually be safer than Taiwan..

 

[Nanochip]

So how does Samsung “4nm” compare with TSMC 5nm, and the upcoming intel “4” ?

 

[Daven]

Meanwhile over at Intel…

Pat: “World say hello to Intel 7 available now for fab customers.”

World: …cricket sounds…

 

[Nanochip]

Meanwhile over at Intel…

Pat: “World say hello to Intel 7 available now for fab customers.”

World: …cricket sounds…

Lol lies. Intel will use 18A for its foundry customers and ARM at least has signed up. Intel’s foundry ambitions may well be a flop but ARM signing up is arguably major. For many reasons.

 

[dir_d]

The Phoenix APUs are monolithic.

Then it must just come down to pricing, must have a better deal at Samsung.

 

[Fouquin]

Taking a wild guess but the 5 nm + 6nm approach is not going well so they went to Samsung for a unified 4 nm part.

The two are different product segments. Dragon Range (Ryzen 7045 series) is desktop Zen 4 MCMs repackaged on BGA; so 5nm CCDs and 6nm IOD. Phoenix (Ryzen 7040 series) is a monolithic single-die Zen 4 APU on 4nm. The former is working well as a chiplet design should. The latter is having problems getting wafer allocation because AMD is not the primary customer for 4nm wafers at TSMC, and they're playing fourth fiddle to the top dogs. Doesn't have anything to do with pricing and everything to do with not getting enough chips made because TSMC literally cannot give them more right now.

 

[Count von Schwalbe]

Doesn't have anything to do with pricing and everything to do with not getting enough chips made because TSMC literally cannot give them more right now.

I would think that they could get the wafers, given a high enough premium over asking, but it wouldn't make much sense for these parts.

Pity they can't move the low-end RDNA3 parts there, instead of TSMC 6nm.

 

[Fouquin]

I would think that they could get the wafers, given a high enough premium over asking, but it wouldn't make much sense for these parts.

Pity they can't move the low-end RDNA3 parts there, instead of TSMC 6nm.

If they could get the wafers it would have been in a supply contract months ago. I don't believe TSMC offers flexible supply on leading edge nodes, you need to be locked into that contract well in advance. AMD outbidding companies like Apple is laughable at best.

 

[Daven]

Lol lies. Intel will use 18A for its foundry customers and ARM at least has signed up. Intel’s foundry ambitions may well be a flop but ARM signing up is arguably major. For many reasons.

You do realize that ARM isn’t actually a foundry customer. They are not fabbing even a single chip.

 

[LabRat 891]

I wonder what little differences there'll be between them?

IIRC, the K7 (Slot A / Socket 462/A) Athlon XP era had multiple fabs making "the same chip", but they physically were different packages, and had different OCing qualities. I think I've ran across this w/ older nVidia GPUs as well (diff fabs, on diff processes, making the same gen uArch dies; with different dimensional and clocking qualities)

 

[mechtech]

so they'll basically make phoenix on a 7nm TSMC like process. Good luck with it amd. We know for a fact what this results in, from qualcomm. But maybe you can do some magic and remove the crapiness from samsung fab process.

All depends what they are fabbing. A lot of people like their ssd and bdie ram.

 

[TechLurker]

I wonder if this is also in advance of Samsung maybe deploying more AMD IP in some of their product stack down the line. Basically, AMD using one design to "pipe clean" and familiarize themselves with Samsung's nodes, and then port over some designs that will work well within Samsung's fabs as quantity and quality matures.

I recall that besides RDNA IP, Samsung might be working with the Xilinx half to help improve their Exynos after repeated failures, and there was some speculation of maybe involving AMD on a joint chip design to basically redesign Exynos from the ground up.

 

[qcmadness]

But there are just reports that tsmc's 5nm / 4nm process is under-utilized.

TSMC may see 5nm capacity utilization drop below 70% in 2Q23

TSMC's capacity utilization for 5/4nm process nodes, which will fall to about 75% in the first quarter of 2023, may drop below 70% in the second quarter, according to industry sources.

www.digitimes.com

 

[R0H1T]

There shouldn't be much difference given they won't be clocking to 6GHz or come insane level, most of them would be tuned for best efficiency so there shouldn't be too much difference between the chips unless 4nm Samsung is really really bad.

Laptop chips also downclock incredibly quickly so there's that.

 

[Wirko]

If they could get the wafers it would have been in a supply contract months ago. I don't believe TSMC offers flexible supply on leading edge nodes, you need to be locked into that contract well in advance. AMD outbidding companies like Apple is laughable at best.

This also means that AMD has been aware for many months that they wouldn't be able to get enough wafers from TSMC. If they are smart, they designed the Phoenix chip with two foundries in mind so they need not waste an entire year now, or more, to port the design to Samsung's foundry.
(To be clear, I'm not saying that a single design can simply be copied to another stack of floppies and mailed to Korea; I'm saying AMD has had enough time to prepare.)

 

[TheinsanegamerN]

But there are just reports that tsmc's 5nm / 4nm process is under-utilized.

TSMC may see 5nm capacity utilization drop below 70% in 2Q23

TSMC's capacity utilization for 5/4nm process nodes, which will fall to about 75% in the first quarter of 2023, may drop below 70% in the second quarter, according to industry sources.

www.digitimes.com

It maybe less about capacity and more about samsung offering a notably better price. With their profits in the toilet, losing 3.4 billion last quarter, they will likely be *incentivising* customers, and AMD will want tot ake advantage of that to raise margins on things like mobile chips.

 

[kondamin]

If that pushes prices down i guess it’s a good thing.

if the can push idle power to an absolute minimum they would make nice stb and nas processors even if they can’t reach super high clocks.