News AMD to use Samsung's 3nm tech as it looks to dual-source future chips: report

[Admin]

AMD is reportedly interested to use Samsung's 3nm-class node with gate-all-around transistors.

AMD to use Samsung's 3nm tech as it looks to dual-source future chips: report : Read more

 

[artk2219]

This is what they used to do with Global foundries, prior to them getting stuck in a process rut, the smallest node they currently offer is 12nm. Granted thats still perfectly fine for pretty much anything that isn't the bleeding edge, but they're not going to be making the latest CPU's, GPU's, or phones.

 

[Pierce2623]

This is what they used to do with Global foundries, prior to them getting stuck in a process rut, the smallest node they currently offer is 12nm. Granted thats still perfectly fine for pretty much anything that isn't the bleeding edge, but they're not going to be making the latest CPU's, GPU's, or phones.

Yeah and AMD planned to continue to buy new wafers from GloFo for minor chips like IO dies until they canceled their 7nm and decided they weren’t going to work on any advanced nodes at all anymore. AMD IO dies would’ve been on GloFo 7nm for Zen 4 until GloFo messed it all up with their lack of R&D. As it stands now the Zen 3 IO die is the newest chip they buy from GloFo and I’m sure you already know that it’s on the 12nm process you mentioned.

 

[Notton]

Samsung 3nm GAAFET was supposed to come out in 2023.
It was supposedly being produced in 2022 June.
In fact, there were rumors of AMD looking at Samsung 3nm GAAFET all the way back in 2021 Nov.
It's 2024 May.

So... here is hoping AMD passes on the cost savings to customers, because I don't see why anyone would use Samsung 3nm, unless it was a huge discount deal.

 

[ThomasKinsley]

AMD worked with Samsung to improve graphics in their Exynos chips. Now they are choosing them for fab work. It seems like AMD and Samsung are forming a partnership much like Nvidia and MediaTek. Interesting turn of events in the GPU industry.

 

[usertests]

I just want to see AMD make a dirt cheap APU over at Samsung, such as Sonoma Valley. Then shove them into the channel like Alder Lake-N. Did I mention CHEAP?

The other options would be more interesting for analysis, such as dual sourcing chiplets or making a GPU die.

 

[spiketheaardvark]

Is this because Samsung has what they want or is it because because Nvidia and Apple have bought up all of the available supply from TSMC?

I'd be curious to know what the profit per mm2 of a desktop CPU is compared to a top end datacenter "GPU". Nvidia is selling every data center GPU as fast as they can make them and I bet they can afford to pay a premium for the wafer supply on those things.

 

[watzupken]

I guess the reasons are,
1. Diversification - In case war breaks out in Taiwan or any natural disaster
2. Getting too squeezy over at TSMC - With major companies like Apple, Nvidia, Qualcomm and Intel vying for allocation. Which leads to,
3. Cost - Its not cheap with all the demands.

 

[renz496]

Is this because Samsung has what they want or is it because because Nvidia and Apple have bought up all of the available supply from TSMC?

I'd be curious to know what the profit per mm2 of a desktop CPU is compared to a top end datacenter "GPU". Nvidia is selling every data center GPU as fast as they can make them and I bet they can afford to pay a premium for the wafer supply on those things.

the former? isn't that intel also enter the fray to get TSMC 3nm?

 

[usertests]

Is this because Samsung has what they want or is it because because Nvidia and Apple have bought up all of the available supply from TSMC?

AMD does not have much trouble getting wafers from TSMC, but they could almost always use more and a number of their own products are competing for wafers, from high-end AI chips, to desktop CPUs/GPUs, to laptop APUs, to the console chips, etc.

What Samsung could offer is something worse than what TSMC has, but signficantly cheaper wafers that are not being shared between dozens of different products. It would increase the volume that AMD can sell, and possibly allow them to address the low-end better. We can see similar things being done over at Intel, which can use TSMC in addition to its own nodes.

 

[suryasans]

I believe this Samsung node will be used to produce AMD's low end version of AMD Ryzen AI with ultra low power profiles. It's the most logical choice because AMD can bundle Samsung memory chips (LP-DDR5 and nVME NAND) to its platforms that will be sold to the OEMs.

 

[Notton]

the former? isn't that intel also enter the fray to get TSMC 3nm?

No, Intel's former CEO pre-ordered a bunch of TSMC allocation.
Intel was obligated to make something under the contract and decided to use it for their GPU.

 

[renz496]

No, Intel's former CEO pre-ordered a bunch of TSMC allocation.
Intel was obligated to make something under the contract and decided to use it for their GPU.

nah i was talking about 3nm specifically. well even without intel AMD will still need to contest with the big boys like apple/nvidia/qualcomm for those bleeding edge nodes. last i heard 80% of 3nm capacity still go towards apple. that number most likely not going down unless apple fully move to TSMC next major node. i think that's why nvidia blackwell sticks with 5nm.

 

[NinoPino]

So... here is hoping AMD passes on the cost savings to customers, because I don't see why anyone would use Samsung 3nm, unless it was a huge discount deal.

Until released we cannot say if it is a good or bad product.
I hope for a surprise.

 

[NinoPino]

AMD worked with Samsung to improve graphics in their Exynos chips. Now they are choosing them for fab work. It seems like AMD and Samsung are forming a partnership much like Nvidia and MediaTek. Interesting turn of events in the GPU industry.

But the first collaboration round for Exynos was not very effective.

 

[NinoPino]

No, Intel's former CEO pre-ordered a bunch of TSMC allocation.
Intel was obligated to make something under the contract and decided to use it for their GPU.

Are you saying that wasn't Gelsinger to choose to use TSMC ?

 

[TerryLaze]

Isn't it extremely involved to port a CPU design over to a different process?
ARM had to work with intel for quite a while to make it possible to produce arm designs at intel FAB, at least they made a big deal out of it making it look like it's pretty hard.
I have a feeling that if this was about the actual cpu cores samsung would have made a big deal out of this as well, why shouldn't they advertise that they can do something like that?!

 

[Notton]

Are you saying that wasn't Gelsinger to choose to use TSMC ?

It may have been Gelsinger. Intel bought out TSMC 3nm years ago, back in 2021 Aug.
But Gelsinger joined Intel as CEO in 2021 Feb.
That timeline from joining to inking in a contract sounds hectic.

 

[williamcll]

My guess is that it would be used to make the I/O chip as well as some lower end processors (embedded, refreshes of last gen mobile processors)

 

[Pierce2623]

I just want to see AMD make a dirt cheap APU over at Samsung, such as Sonoma Valley. Then shove them into the channel like Alder Lake-N. Did I mention CHEAP?

The other options would be more interesting for analysis, such as dual sourcing chiplets or making a GPU die.

Just make a cheap 8600g on Samsung 3nm and sell it for $150. That would be perfect. 8CU of rdna3 is mostly enough to saturate the bandwidth anyways.

 

[usertests]

Just make a cheap 8600g on Samsung 3nm and sell it for $150. That would be perfect. 8CU of rdna3 is mostly enough to saturate the bandwidth anyways.

I want AMD to offer an alternative to the Alder Lake-N mini PCs, which often cost $150-200 for the total system including storage and RAM.

Mendocino was the obvious counter to Alder Lake-N with a very similar performance profile, but AMD couldn't or wouldn't make it in a high enough volume to matter.

Sonoma Valley is a drop-in replacement for Mendocino based on leaked docs, and other leaks are pointing to it using Zen 5c cores. Now that we know AMD's claimed 16% IPC uplift for Zen 5, you can calculate that moving from Zen 2 to Zen 5c would result in +56% IPC, which is massive.

If Samsung could make a lot of Sonoma Valley APUs for AMD, they could challenge Intel's dominance at the low-end, even if Intel switched from Gracemont to Skymont E-cores. It would have a significantly worse iGPU than a 8600G but still be a good cheap computer.

 

[Pierce2623]

I want AMD to offer an alternative to the Alder Lake-N mini PCs, which often cost $150-200 for the total system including storage and RAM.

Mendocino was the obvious counter to Alder Lake-N with a very similar performance profile, but AMD couldn't or wouldn't make it in a high enough volume to matter.

Sonoma Valley is a drop-in replacement for Mendocino based on leaked docs, and other leaks are pointing to it using Zen 5c cores. Now that we know AMD's claimed 16% IPC uplift for Zen 5, you can calculate that moving from Zen 2 to Zen 5c would result in +56% IPC, which is massive.

If Samsung could make a lot of Sonoma Valley APUs for AMD, they could challenge Intel's dominance at the low-end, even if Intel switched from Gracemont to Skymont E-cores. It would have a significantly worse iGPU than a 8600G but still be a good cheap computer.

I’ve actually got a Mendocino handheld and it’s pretty damn good for emulation but really craps the bed with any modern 3D rendering. It runs at a slightly higher power envelope than Alder Lake N but it also beats Alder Lske N in single or multithreaded loads, but yeah overall the performance per watt will be close to Alder Lake N.

 

[usertests]

I’ve actually got a Mendocino handheld and it’s pretty damn good for emulation but really craps the bed with any modern 3D rendering. It runs at a slightly higher power envelope than Alder Lake N but it also beats Alder Lske N in single or multithreaded loads, but yeah overall the performance per watt will be close to Alder Lake N.

I remember before Mendocino details were official, when we thought it would be a carbon copy of the custom APU used in Steam Deck, with 8 CUs. Then we learned it had 2 CUs and 64-bit memory which they call "dual-channel". Then came the $399-699 "mainstream notebook" pricing slide. RIP.

I think Mendocino always gets the single-threaded win, but Alder Lake-N gets the multi-threaded win when it has 8 cores, so that's part of why I see the two dies as very similar. In practice, 8-core N300/N305 are much less common than the quad-cores, and the 2-core Athlon Gold/Silver versions of Mendocino are very scarce.