What’s the fate of K12/ARM at AMD?

[nicalandia]

For that matter Jim Keller recently said K12 was "stupidly cancelled".

Jim Kelly is an Awesome CPU Designer, but clearly not Management oriented(not at the level of Dr. Lisa Su). he should have known that a company on the verge of Bankruptcy required to consolidate it's effort on making a very competitive X86 core to fight against it's sole competitor(Intel), instead of trying to make yet another ARM CPU that would have to fight at least a dozen other competitors.

 

[moinmoin]

We already had that very discussion earlier in this thread (mostly in agreement with your stance).

The interesting thing is with Xilinx and Pensando AMD now actually has multiple products that build on ARM, so the big question is: Where to go from there? Does AMD still have staff with significant knowledge in ARM from the K12 days, or will it all be dominated by new Xilinx and Pensando staff? Will the combined company keep using stock ARM cores, or will the K12 approach eventually be resurrected, consolidating the backend for both x86 and ARM?

 

[Thibsie]

We already had that very discussion earlier in this thread (mostly in agreement with your stance).

The interesting thing is with Xilinx and Pensando AMD now actually has multiple products that build on ARM, so the big question is: Where to go from there? Does AMD still have staff with significant knowledge in ARM from the K12 days, or will it all be dominated by new Xilinx and Pensando staff? Will the combined company keep using stock ARM cores, or will the K12 approach eventually be resurrected, consolidating the backend for both x86 and ARM?

From what I remember (dunno from where sorry), designing a x86 core or an ARM core isn't that different, most of the difference being in the front end.
Of course, I guess SIMD might more specific though.
AMD works with ARM too, there's one in all our AMD cpus.
I'm not convinced AMD would get significant difficulties designing an ARM core.

 

[Asterox]

From what I remember (dunno from where sorry), designing a x86 core or an ARM core isn't that different, most of the difference being in the front end.
Of course, I guess SIMD might more specific though.
AMD works with ARM too, there's one in all our AMD cpus.
I'm not convinced AMD would get significant difficulties designing an ARM core.

If old rumors are legit, we can expect various news in the near future.

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

 

[DisEnchantment]

The interesting thing is with Xilinx and Pensando AMD now actually has multiple products that build on ARM, so the big question is: Where to go from there? Does AMD still have staff with significant knowledge in ARM from the K12 days, or will it all be dominated by new Xilinx and Pensando staff? Will the combined company keep using stock ARM cores, or will the K12 approach eventually be resurrected, consolidating the backend for both x86 and ARM?

Not sure if there is a business case for AMD in the near future to make ARM CPUs for PCs. That's like supporting a crowded ecosystem where all their competitors can benefit from.
And cloud operators would rather fab their own CPUs from ARM reference designs.
At best semi custom or UCIe integration.
For automotive perhaps? But in this field reference designs are good enough.

And Xilinx/Pensando use older reference designs from ARM. I don't think they are using any recent licenses.

I don't know if there is some value to replacing a lot of ARM IPs in their GPUs, CPUs, FPGAs, SmartNICs etc.
ARM licenses are quite cheap, for now, might not be worth the hassle switching to x86 or RISC-V for example.
Although it seems they are starting to replace some microcontroller blocks within GPUs with RISC-V (AMD is now a strategic RISC-V member through Xilinx)

 

[moinmoin]

From what I remember (dunno from where sorry), designing a x86 core or an ARM core isn't that different, most of the difference being in the front end.

The K12 design was all about it, shared backend with Zen but ARM frontend for it. We discussed it before in this thread. Keller mentioned it again in the video I linked on the previous page.

Not sure if there is a business case for AMD in the near future to make ARM CPUs for PCs. That's like supporting a crowded ecosystem where all their competitors can benefit from.
And cloud operators would rather fab their own CPUs from ARM reference designs.
At best semi custom or UCIe integration.
For automotive perhaps? But in this field reference designs are good enough.

And Xilinx/Pensando use older reference designs from ARM. I don't think they are using any recent licenses.

I don't know if there is some value to replacing a lot of ARM IPs in their GPUs, CPUs, FPGAs, SmartNICs etc.
ARM licenses are quite cheap, for now, might not be worth the hassle switching to x86 or RISC-V for example.
Although it seems they are starting to replace some microcontroller blocks within GPUs with RISC-V (AMD is now a strategic RISC-V member through Xilinx)

I wasn't thinking of ARM for competing with its own Zen products in the PC market. I was more thinking about internally consolidating different cores across all three companies.

• Zen cores obviously stay the high performance cores, should need arise for high performance cores using the ARM or RISC-V instruction set, preparing a decoupling of the backend (again if the work indeed completely stopped after K12) may be a feasible path to there (if not happening already).
• AMD also uses an old small ARM core in every single Zen chip (for Infinity Fabric and PSP), and Xlinix and Pensando may have more expertise around that so that could probably unified and improved upon (likely especially software and firmware wise).
• And as you mentioned Xilinx contributes RISC-V which could be standardizes across the company as the next lower level, replacing microcontrollers.

All those can then be offered to semi custom business customers.

 

[maddie]

When we say ARM, do we mean ARM cores or designing cores for the ARM ISA? They can lead to different results.

 

[moinmoin]

When we say ARM, do we mean ARM cores or designing cores for the ARM ISA? They can lead to different results.

I think it's safe to assume that AMD in totality only uses stock ARM cores currently. This thread in particular is about K12 though which was an effort to create an ARM ISA compatible frontend for the Zen backend, an effort AMD may eventually resurrect if it didn't already.

 

[NostaSeronx]

imho, it will be funny if AMD decides to improve/refresh/next-gen AMD64 instead;
Big AMD64r2 (Rather than a complete break of compatibility, it is a partial break of compatibility?)
Tiny AMD64r2 (Processors using this ISA replace these: Cortex-A5/Cortex-A7+Dual-core Cortex-M4)

Supposedly, new/revolutionary/re-architected [AMD64] ISAs, developing in DCs located:
Boxborough, MA
Santa Clara, CA
Austin, TX
Bellevue, WA

// Tiny AMD64 Micro-controller: 2017+
// Revamped AMD64 ISA: 2020+
// Hiring extended AMD64 ISA+patent target(group that patents enhancements) staff: 2021+

 

[yuri69]

The article mentions "dense servers" as K12's target deployment. This was the niche all those early server ARM attempts targeted (including AMD's Seattle ARMv8). It kinda didn't work out.

 

[eek2121]

We already had that very discussion earlier in this thread (mostly in agreement with your stance).

The interesting thing is with Xilinx and Pensando AMD now actually has multiple products that build on ARM, so the big question is: Where to go from there? Does AMD still have staff with significant knowledge in ARM from the K12 days, or will it all be dominated by new Xilinx and Pensando staff? Will the combined company keep using stock ARM cores, or will the K12 approach eventually be resurrected, consolidating the backend for both x86 and ARM?

Every Zen product includes an ARM processor. The PSP is ARM-based.

 

[moinmoin]

Every Zen product includes an ARM processor. The PSP is ARM-based.

I know. I stated as much several times in this thread, last time in #181. I do wonder who all is not yet aware of that.

 

[eek2121]

I know. I stated as much several times in this thread, last time in #181. I do wonder who all is not yet aware of that.

Figured I'd bring it up, because some people aren't aware. ARM lives on within AMD, even if no public products are coming.

 

[NTMBK]

Yes, I'm sure AMD will be launching their FDSOI K12 derived secret weapon any day now

 

[Thibsie]

Yes, I'm sure AMD will be launching their FDSOI K12 derived secret weapon any day now

Ate Nosta food ? xD

 

[NostaSeronx]

Yes, I'm sure AMD will be launching their FDSOI K12 derived secret weapon any day now

The Mountain Peak "K12" ARM Core was only ever 16nm TSMC or 14nm GloFo. Any new ARM core, wouldn't be called "K12" but based off another Mountain Peak or completely different naming scheme.


FDSOI projects had no mountain peaks or K-series. The current plan for GloFo's FDSOI is reconfigurable extremely power-efficient "ULP" AMD64 processors. With the first one doing Dynamism with clusters and second one doing Dynamism with grids.

Dynamism = Dynamic Level of Parallelism :: Instruction Level Parallelism associated with OoO and Thread Level Parallelism associated with Multithreading.
Without loss in ILP efficiency or TLP efficiency.

1. ARM custom "Mountain Peak" cores was never FDSOI.
2. FDSOI project developed into aiming at below 4.5W/6W: Micro-family of GF28A Mullins and StoneyX GF28A. With dynamic ILP/TLP without efficiency loss.
TDP Saga:
2014+ - Mullins/Stoney start, the successor planned TDP was 3.6W
2017+ - Removal of Raven2 as G-series and insert of actual G-series successor, with the TDP being ~2W.
=> Client Products have reduced TDP but aren't as low as that.
Process saga:
2012+ - 28FD/20FD
2014+ - Advanced FDSOI
2018+ - Waiting for High Mobility SOI, which was actually suppose to come much, much sooner. This was presented and included into 12FDX in 2015-2016.
2020+ - 12FDX rather than sharing 20nm Lgate with 22FDX, went to sharing 14nm Lgate with 12nm FinFET.

~
ARM Core lineup would have halved AMD's development share between ~50% AMD64 and ~50% ARM64. In hindsight, the trend however would have had the cores split at Zen2.
ARMv8 (Post-K12/2nd Gen Mountain Peak) = VL128
AMD64 (Post-Zen/2nd Gen Zen: Valhalla) = VL256
AMD64 (Post-Zen/3rd Gen Zen: Cerberus) = VL256
AMD64 (Post-Zen/4th Gen Zen: Persephone) = VL512
The inclusion of SVE(purchase of SVE license) would have fixed that, but SVE doesn't perfectly align with AVX2. More issues would largely have popped up trying to keep the cores as similar as possible.

The only benefit of an ARM core is more software developers than AMD64. However, getting developers willing to spend money on getting AMD's custom ARM core is probably like Zen, and not in favor of AMD. Opteron A1100 eventually fell to $399 from $2999. It is unlikely that K12's dev kit would have been cheaper. AMD's development of Mtn. Peaks would have thus be as close as possible to Cortex-A. Where as AMD's development of Zen do to low amount of developers has to copy Intel's P-core. Both of which had a huge generational gap back then.