News AMD's semi-custom division to produce ARM CPUs

[DrMrLordX]

AMD: We Stand Ready to Make Arm Chips

The ARMy grows

www.tomshardware.com

Looks like customers are reaching out to AMD to produce ARM-based products through their semi-custom division. That's . . . interesting. Why aren't they going to Ampere instead? Is Ampere saying "no"?

 

[eek2121]

AMD: We Stand Ready to Make Arm Chips

The ARMy grows

www.tomshardware.com

Looks like customers are reaching out to AMD to produce ARM-based products through their semi-custom division. That's . . . interesting. Why aren't they going to Ampere instead? Is Ampere saying "no"?

Maybe because AMD has more experience creating CPUs, even ARM CPUs, than Ampere?

Also, don’t be surprised if you see AMD jump into the mobile ARM game pretty soon.

 

[DrMrLordX]

Maybe because AMD has more experience creating CPUs, even ARM CPUs, than Ampere?

Not really. AMD hasn't produced a known ARM chip since K12. Ampere's Altra lineup is pretty legit.

 

[JasonLD]

They can just take ARM reference CPU and combine it with their own RDNA GPUs. (Would be just like Samsung's next Exynos lol) I doubt they will do more than that though,.

 

[NTMBK]

We already know about the AMD-Samsung partnership for a phone SoC, so this is hardly news. If they have done the work to bring their GPU architecture down to phone power levels, they will want to reuse that work in other projects.

 

[moinmoin]

My guess is that all of AMD's IPs beyond the cores are internally fully interchangeable between x86 and ARM since the development on K12, and the only reason AMD itself sells no ARM-based product is purely a market decision. Semi custom is the gate customers can use to get access to all the uncore, packaging and interconnect IPs AMD introduced since the launch of Zen, so some requests for ARM solutions are bound to happen at some point. Also chances are that planning for cooperation is ongoing for a long time already, predating Ampere's more recently announced achievements.

 

[Asterox]

Not really. AMD hasn't produced a known ARM chip since K12. Ampere's Altra lineup is pretty legit.

Officially not, but behind closed doors is obviously very interesting.

Page 5 - What’s the fate of K12/ARM at AMD?

Page 5 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

https://twitter.com/x/status/1332601734565416962

 

[eek2121]

Not really. AMD hasn't produced a known ARM chip since K12. Ampere's Altra lineup is pretty legit.

Every Ryzen processor has an ARM core inside of it: https://en.m.wikipedia.org/wiki/AMD_Platform_Security_Processor

 

[amd6502]

Officially not, but behind closed doors is obviously very interesting.

Page 5 - What’s the fate of K12/ARM at AMD?

Page 5 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

https://twitter.com/x/status/1332601734565416962

Wow, that's a little more than just simple big.little layout:

2x Big cores (Coretex X1/A78+)

2x frequency-limited Big cores (A78)

4x Little cores for low power and background threads


This looks pretty well suited for a broad range, from Telephones+phablets to Tablets + 2in1 + Chromebooks and to Automotive computing and other embedded cases.

 

[uzzi38]

Maybe because AMD has more experience creating CPUs, even ARM CPUs, than Ampere?

Also, don’t be surprised if you see AMD jump into the mobile ARM game pretty soon.

I'm going to disagree here - I don't think you should expect AMD to hop into that game any time soon. Mobile RDNA basically screams they have no intentions of doing so - rather it seems like AMD would prefer it if existing ARM silicon developers use Radeon IP instead.

 

[uzzi38]

My guess is that all of AMD's IPs beyond the cores are internally fully interchangeable between x86 and ARM since the development on K12, and the only reason AMD itself sells no ARM-based product is purely a market decision. Semi custom is the gate customers can use to get access to all the uncore, packaging and interconnect IPs AMD introduced since the launch of Zen, so some requests for ARM solutions are bound to happen at some point. Also chances are that planning for cooperation is ongoing for a long time already, predating Ampere's more recently announced achievements.

Not sure if they'd be fully interchangeable - they'd almost certainly have to do some work to bring the old K12 IP up to date, but they definitely have the groundwork laid for creating Zen based CPUs with ARM front-ends if nothing else.

You're absolutely right about the market decision bit though.

 

[DisEnchantment]

I'm going to disagree here - I don't think you should expect AMD to hop into that game any time soon. Mobile RDNA basically screams they have no intentions of doing so - rather it seems like AMD would prefer it if existing ARM silicon developers use Radeon IP instead.

Indeed
No Baseband IP of any kind
No Audio DSP or basically any DSP (they could shop around for sensor, camera, audio and other DSP blocks from synopsys/cadence), on top of this needs a framework a la Hexagon SDK
No AOP
Without these Mobile SoC will not happen. Or it will fail miserably (Hello Tegra)

 

[NTMBK]

Not sure if they'd be fully interchangeable - they'd almost certainly have to do some work to bring the old K12 IP up to date, but they definitely have the groundwork laid for creating Zen based CPUs with ARM front-ends if nothing else.

You're absolutely right about the market decision bit though.

Why would they need to make custom cores? They could do semicustom with X1/X2 cores.

 

[leoneazzurro]

Well it all depends on who are the customers who are interested in ARM from AMD: if they are server-oriented (not needing a lot of technology used in the mobile field) or console makers (which usually have their own IP to implement and their own SDKs) an ARM product would be quite feasible, while I agree than in mobile the situation is much more complicated. But, it's also true that the acquisition of Xilinx and their IP could open (not immediately but in a foreseeable future) for some interesting evolution in fields not yet explored by AMD itself.

 

[uzzi38]

Why would they need to make custom cores? They could do semicustom with X1/X2 cores.

Oh yes they absolutely could.

But why would they? X1 is extremely similar in power, performance and area to Zen 3, while being on 5LPP whilst the latter is on N7.

 

[DrMrLordX]

We already know about the AMD-Samsung partnership for a phone SoC, so this is hardly news. If they have done the work to bring their GPU architecture down to phone power levels, they will want to reuse that work in other projects.

That would just be a Samsung design, though. It would be weird for anyone to source that design from AMD (and not Samsung).

Every Ryzen processor has an ARM core inside of it: https://en.m.wikipedia.org/wiki/AMD_Platform_Security_Processor

Irrelevant. You can't use it to execute aarch64 code or really anything else meaningful. Do I really have to be that specific in saying that AMD doesn't produce any known ARM CPUs/SoCs?

 

[NTMBK]

Oh yes they absolutely could.

But why would they? X1 is extremely similar in power, performance and area to Zen 3, while being on 5LPP whilst the latter is on N7.

Because they would need to invest a LOT of R&D in making a custom core, with no guarantee that it would significantly beat ARM's designs. (Just look at what happened with Samsung's Mongoose designs.) Whereas they could integrate an ARM design with their own GPU IP pretty easily, and sell a semicustom chip to e.g. Nintendo.

 

[uzzi38]

Because they would need to invest a LOT of R&D in making a custom core, with no guarantee that it would significantly beat ARM's designs. (Just look at what happened with Samsung's Mongoose designs.) Whereas they could integrate an ARM design with their own GPU IP pretty easily, and sell a semicustom chip to e.g. Nintendo.

As hinted at elsewhere in this thread, AMD already had a very efficient ARM core design.

K12. It was effectively OG Zen with an ARM frontend (AKA had the Zen backend), and was capable of similar performance per core (and per clock).

 

[moinmoin]

Not sure if they'd be fully interchangeable - they'd almost certainly have to do some work to bring the old K12 IP up to date, but they definitely have the groundwork laid for creating Zen based CPUs with ARM front-ends if nothing else.

Note that I explicitly wrote only about uncore, interconnect and packaging IPs (could have mentioned graphics as well), not cores. While with K12 AMD certainly worked on ARM cores, I don't think custom ARM cores are going to fly as part of the semi custom business, updated or not, certainly not among the first couple contracts.

The main issue here is transparency. The ARM cores by Arm itself are all known entities, whereas whatever custom design AMD might have is completely unknown.

So new semi custom business customers are much more likely to want specific known Arm IP core designs, maybe combined with own IPs, and go to AMD to bundle all that with and using AMD IPs.

 

[NTMBK]

As hinted at elsewhere in this thread, AMD already had a very efficient ARM core design.

K12. It was effectively OG Zen with an ARM frontend (AKA had the Zen backend), and was capable of similar performance per core (and per clock).

All we know is that they were "sister" designs, and that K12 had a "wider engine". We don't know how far through design, validation and implementation K12 was. We don't know how much it differed from Zen. We don't know how much extra it cost to design both, instead of just Zen. We don't know how good the final design would be, or how much die area it would take compared to an ARM design. There's been a lot of speculation, but not a lot of hard facts.

And it was a 14nm, ARMv8 design, contemporary with Zen 1. AMD would need to do a lot of work to bring it to 5nm, implement ARMv9, update the vector units to support SVE2, and improve the design with the last 5 years of learning to keep the design competitive. Whereas they could just license a core from ARM and be done.

 

[eek2121]

That would just be a Samsung design, though. It would be weird for anyone to source that design from AMD (and not Samsung).



Irrelevant. You can't use it to execute aarch64 code or really anything else meaningful. Do I really have to be that specific in saying that AMD doesn't produce any known ARM CPUs/SoCs?

It's not irrelevant at all. It's an ARM core and it executes ARM code. As others have also stated, AMD has ARM R&D work going on as well. AMD has been toying with ARM for years. The reason we don't see an end user product is lack of demand and desire by OEMs to support.

 

[uzzi38]

All we know is that they were "sister" designs, and that K12 had a "wider engine". We don't know how far through design, validation and implementation K12 was. We don't know how much it differed from Zen. We don't know how much extra it cost to design both, instead of just Zen. We don't know how good the final design would be, or how much die area it would take compared to an ARM design. There's been a lot of speculation, but not a lot of hard facts.

And it was a 14nm, ARMv8 design, contemporary with Zen 1. AMD would need to do a lot of work to bring it to 5nm, implement ARMv9, update the vector units to support SVE2, and improve the design with the last 5 years of learning to keep the design competitive. Whereas they could just license a core from ARM and be done.

Oh, I guess it wasn't all that well known then. I'm totally serious about what it is, and it didn't tape-out on 14nm but rather a different node.

Aaaand that's my input on K12 done for the day

 

[DrMrLordX]

It's not irrelevant at all. It's an ARM core and it executes ARM code.

LOL that you actually are making such an argument. Trust Zone says nothing positive or negative about AMD's ability to deliver performant ARM solutions. If Via/Zhaoxin starts using Trust Zone-like ARM cores, does that mean they too will be able to create a competitor to the M1 or the Altra? No! It would only mean that they've used a tiny, generally-non-performant ARM core as a security processor. That's it!

 

[NTMBK]

Oh, I guess it wasn't all that well known then. I'm totally serious about what it is, and it didn't tape-out on 14nm but rather a different node.

Aaaand that's my input on K12 done for the day

Huh, sorry, didn't know you had insider info on K12! My apologies.

 

[turtile]

AMD: We Stand Ready to Make Arm Chips

The ARMy grows

www.tomshardware.com

Looks like customers are reaching out to AMD to produce ARM-based products through their semi-custom division. That's . . . interesting. Why aren't they going to Ampere instead? Is Ampere saying "no"?

I think AMD has a lot of advantages over Ampere. For one, AMD is a much larger company with more resources to help partners. The same advantage that Intel has over AMD.

AMD has superior packaging technology compared to Ampere. I assume that AMD's infinity fabric (consider the next gen) will save on product costs, and increase the max frequency of the highest sku due to chiplet binning. AMD already has 3D stacking for cache and I'm sure we will see more with Zen 4.

AMD can share the same socket and platform with the already established platform. This will save on costs since it is massed produced. In addition, AMD can integrate its CDNA chips and Xilinx chips.