Question [AT] AMD amended WSA with GloFo

[jpiniero]

AnandTech Forums: Technology, Hardware, Software, and Deals

Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

www.anandtech.com

WSA ends at the end of 2024. Says AMD is obligated to buy 1.6 billion worth of wafers from GloFo until then.

Hope you like Picasso and Dali.

 

[Hitman928]

AnandTech Forums: Technology, Hardware, Software, and Deals

Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

www.anandtech.com

WSA ends at the end of 2024. Says AMD is obligated to buy 1.6 billion worth of wafers from GloFo until then.

Hope you like Picasso and Dali.

I think (from AMD's side) it's more about making sure they have guaranteed wafers for their LTS products (Naples, and IO dies Zen2+ generations). They'll probably have some chips produced that are for cheap markets as well, but the main focus is not on the budget stuff.

 

[gdansk]

Zen(+) is still reasonably good for Chromebook tier devices and embedded.
The X570 chipset is made at GF. And they will need a lot more IO chiplets if Epyc keeps ramping up as they have.
All those over 3 years could be enough.

 

[Leeea]

Chipsets, the entire product stack
IO dies
FPGAs
various other embedded products, including signage processors and possibly a Geode replacement. Keep in mind Geode was discontinued 2019 and it does not seem like anyone is offering a replacement.

 

[moinmoin]

The Ryzen Embedded and Epyc Embedded line launched early 2018 and have a guaranteed availability for ten years, so up to 2028. So even if IODs, chipsets and APUs shouldn't cover all of the $1.6 billion already AMD could build a neat warehouse with these embedded chips (though I do expect IODs for all Zen2/3 Ryzen and Epyc chips to cover the vast majority of it as is).

@Leeea FPGAs could be a good point. Is Xilinx already a customer of GloFo? I think I read Xilinx almost exclusively uses TSMC though.

Edit: Good notice by user sgeocla in the comments:

this new deal means that AMD won't have to pay penalties for Xillinx products that are on TSMC's older nodes. Under the old WSA AMD could not freely use TSMC modes older than 7nm.

 

[Gideon]

Why are everybody recommending buildnig FPGAs on terribly outdated processes?

AFAIK these are one of the first things going to newer nodes and their product prices also usually make using leading-edge nodes a non-issue.

IMO using GloFo for FPGAs would be a terribe idea, though I guess the very lowest parts of the stack could transition there if really needed. Still not sure that Xilinx has the volume to help with this WSA.

Chipsets I/O dies and chromebook material (Dali, Picasso) will surely be included but even accouting for all that, the amount of wafers is still gigantic. Besides the amendment starts from 2022 (which means this year is not included).

Perhaps some very-lowest end RDNA2/RDNA3 GPUs could be made on GloFo? Most people don't care about power-consumption on that level, and GPUs are usually pretty straightforward to port. I'm not quite convinced the current product stack is sufficient till up to and including 2024 (, I sure as hell hope nobody buys a Picasso Chromebook in 2024),

Over all all I can say is that Hector Rui(n)z surely caused AMD a lot of pain due to this

 

[jpiniero]

I sure as hell hope nobody buys a Picasso Chromebook in 2024),

Intel is selling a ton of Chromebooks with the 2017 Celeron (Atom) N4000 right now although Intel is discontinuing it so OEMs will have to move to Goldmont Plus fully eventually. The way things are going Picasso will be slow but not out of place in 2024.

 

[Olikan]

Why are everybody recommending buildnig FPGAs on terribly outdated processes?

AFAIK these are one of the first things going to newer nodes and their product prices also usually make using leading-edge nodes a non-issue.

IMO using GloFo for FPGAs would be a terribe idea, though I guess the very lowest parts of the stack could transition there if really needed. Still not sure that Xilinx has the volume to help with this WSA.

Chipsets I/O dies and chromebook material (Dali, Picasso) will surely be included but even accouting for all that, the amount of wafers is still gigantic. Besides the amendment starts from 2022 (which means this year is not included).

Perhaps some very-lowest end RDNA2/RDNA3 GPUs could be made on GloFo? Most people don't care about power-consumption on that level, and GPUs are usually pretty straightforward to port. I'm not quite convinced the current product stack is sufficient till up to and including 2024 (, I sure as hell hope nobody buys a Picasso Chromebook in 2024),

Over all all I can say is that Hector Rui(n)z surely caused AMD a lot of pain due to this

Xilling still sells alot of FPGAs on older nodes, even 28nm ones

Also, GPU I/O dies could be produced at glofo

 

[jpiniero]

Perhaps some very-lowest end RDNA2/RDNA3 GPUs could be made on GloFo? Most people don't care about power-consumption on that level, and GPUs are usually pretty straightforward to port. I'm not quite convinced the current product stack is sufficient till up to and including 2024

There are Navi products intended for the low end coming out. AMD could re-release the 570 and 580 for right now but I dunno if they would be able to source GDDR5.

 

[moinmoin]

Regarding the cost of the WSA for the remaining 3 years, 1.6 billion for all three years is actually quite low, around a half to a third of the years for which we have numbers.

Unfortunately AMD's latest Form 10-K for 2020 didn't give new numbers, but the one for 2019 mentioned following costs:
2019 (through May 15, 2019) $0.5 bil (so possibly $1.1 bil for the whole year)
2018 $1.6 bil
2017 $1.1 bil

Since the link in the old post doesn't work anymore, here's a listing of all Form 10-K documents: https://ir.amd.com/sec-filings/filter/annual-filings

The advantage for AMD is that it can treat the WSA as fixed cost regardless of what it actually does with it. I'm sure they'll find a way to make the best use/most profit possible with it.

 

[NTMBK]

Are Microsoft and Sony still making last gen consoles? Microsoft has the Series S for the low end, but Sony don't have an equivalent. I suspect they will keep churning out PS4s for a while longer.

EDIT: PS3 production lasted over 3 years after the launch of the PS4.

 

[sandorski]

Does GF have any new Fabs planned? They need to do something eventually.

 

[eek2121]

I suspect they are doing this for Xilinx. However, I wonder why they opened it up? Are they planning on using TSMC or Samsung for 14nm/12nm as well?

EDIT: Also note that current leaks have the IO die for Zen 4 parts on TSMC 6nm.

 

[eek2121]

There are Navi products intended for the low end coming out. AMD could re-release the 570 and 580 for right now but I dunno if they would be able to source GDDR5.

They could always refresh Polaris. Shoot, they could potentially use GDDR6 or even DDR5.

 

[jpiniero]

Edit: Good notice by user sgeocla in the comments:

Pretty sure the pay to play on a different node was only for 7 nm products, and that was removed when GloFo 7 ended. They were (are?) fabbing the PS4 and Pro at TSMC 16.

 

[moinmoin]

Pretty sure the pay to play on a different node was only for 7 nm products, and that was removed when GloFo 7 ended. They were (are?) fabbing the PS4 and Pro at TSMC 16.

As well as all Xbox One chips. And I'm pretty sure Radeon wasn't all done on GloFo through all the years either. So you may be right.

 

[Gideon]

They could always refresh Polaris. Shoot, they could potentially use GDDR6 or even DDR5.

You can't just use GDDR6, this means a new memory controller. If they're already doing that they might as well do more (as then a already a year+ design and validation time is commited).

They could for instance backport and dual source Navi 24. Not sure any of it is a great idea though, as it eats up very valuable design and validation resources that could postpone more important products.

Also not sure if Navi can run on 12nm at all, never mind the die size and power consumption. Even if they can they needed to decide on this long ago or it wouldn't be on time.

 

[scineram]

Is half a million wafers a lot?

 

[Mopetar]

They could always refresh Polaris. Shoot, they could potentially use GDDR6 or even DDR5.

They'd need a redesign on the memory controller, but honestly with the way the GPU market is right now and the fact that memory bandwidth is what limits ETH hash rate, they could turn those billions spent on GF wafers into tens of billions of additional revenue.

 

[Leeea]

The fourth application!:
eDRAM equivalent

Remember the Broadwell i7 CPUs from Intel? They used a separate 128 MB eDRAM cache chiplet next to the CPU itself:
https://www.anandtech.com/show/1619...ective-review-in-2020-is-edram-still-worth-it

The RX6000 series is already running around with an infinity cache. We know the next series for GPUs out of AMD will be chiplet. We know AMD CPUs are already chiplet. What if AMD just uses GoFlo's 12nm to turn out high density SRAM cache chiplets?

 

[Tuna-Fish]

Why are everybody recommending buildnig FPGAs on terribly outdated processes?

AFAIK these are one of the first things going to newer nodes and their product prices also usually make using leading-edge nodes a non-issue.

IMO using GloFo for FPGAs would be a terribe idea, though I guess the very lowest parts of the stack could transition there if really needed.

Correct, moving any FPGA products to GloFo doesn't make sense. The reason they still want to make FPGAs with other old processes is that a lot of FPGA are sold into low-volume products where the manufacturer really wants exactly the same chips available for a decade or more. This means that Xilinx has to be able to use large processes of foundries other than GloFo, but it makes absolutely no sense to develop new designs on their processes.

 

[scannall]

They'd need a redesign on the memory controller, but honestly with the way the GPU market is right now and the fact that memory bandwidth is what limits ETH hash rate, they could turn those billions spent on GF wafers into tens of billions of additional revenue.

The problem there is that the substrate shortage is still bad. There isn't enough to go around since a substrate plant burned down.

 

[moinmoin]

The problem there is that the substrate shortage is still bad. There isn't enough to go around since a substrate plant burned down.

In the last earnings call Su stated that substrate is an under-invested area and that AMD is already investing in substrate capacity dedicated to AMD.

We continue -- on the substrate side, in particular, I think there has been under-investment in the industry. And so we've taken the opportunity to invest in some substrate capacity dedicated to AMD, and that will be something that we continue to do going forward.

 

[eek2121]

They'd need a redesign on the memory controller, but honestly with the way the GPU market is right now and the fact that memory bandwidth is what limits ETH hash rate, they could turn those billions spent on GF wafers into tens of billions of additional revenue.

Of course they would, but it would probably be worth it. They also may have already done it and never released it.

The problem there is that the substrate shortage is still bad. There isn't enough to go around since a substrate plant burned down.

As the poster above mentions, AMD has made a large investment in this area. They might have some shortage, but it likely isn’t a big deal.

 

[scannall]

Of course they would, but it would probably be worth it. They also may have already done it and never released it.


As the poster above mentions, AMD has made a large investment in this area. They might have some shortage, but it likely isn’t a big deal.

They have made an investment sure, and a big one. But the facilities are not yet online. Supposedly it will be running in Q4.