AMD's ARM implementation speculation

[Raqia]

AMD has survived the past decade by nimbly (or desperately) steering itself (and sometimes the industry) away from being trampled by Intel's in trying to occupy its space directly (see AMD64, Fusion). It announced today that it would be implementing a 64 bit ARM core:

http://www.amd.com/us/aboutamd/newsroom/Pages/presspage2012Oct29.aspx

ARM instructions probably mean that the decoding unit slims down, and for once, the execution units will actually be doing something more or less one to one to the instructions they fetch from memory; also this is probably the real reason they hired Jim Keller back, fresh from A6.

It's a server part, so it'll probably have something like ~16 cores per die w/ a much faster memory controller than most ARM parts on board. They could leverage this design for mobile space by cutting it down to 2 to 4 cores and adding a lightweight GCN core to it. I'll be excited to hear details of this implementation when Analyst's day comes.

 

[rpg.314]

They are using a stock armv8 core. They don't have an architecture license yet. At least not publicly.

 

[Pressure]

Hard to be anything else than next generation ARM architecture as armv7 is only 32-bit.

 

[3dilettante]

It's a server part, so it'll probably have something like ~16 cores per die w/ a much faster memory controller than most ARM parts on board.

That's more cores than AMD has ever been able to connect on x86, back when they had money.

The question now is whether AMD paid the cash for a soft macro it can lay down somewhat more freely, or a hard macro that even less engineering cash would go into.

They could leverage this design for mobile space by cutting it down to 2 to 4 cores and adding a lightweight GCN core to it. I'll be excited to hear details of this implementation when Analyst's day comes.

AMD's slides indicate a clear dilineation between the shared-nothing density server market this ARM chip is going into and anything media-related. That has a GPU and x86 only.
It doesn't mean it can't be done, but AMD could barely handle the workload of handling the software stack it had--again, back when it had money.

I'm more interested in seeing if they mention to the analysts where they're getting money to last until 2014. The layoffs are cutting into necessary functions AMD simply can't afford.

 

[Blazkowicz]

AMD is boasting about interconnexions with the "Freedom fabric" (which funnily is not free as it won't be licensed to others, or at a whim like Intel's QPI)

http://www.theinquirer.net/inquirer...cense-seamicros-freedom-fabric-to-cpu-vendors

16 core looks probable, and dual socket motherboards and systems I'd say. AMD probably intends you to pile up 1Us of this thing, if you stack ten dual socket units in a cabinet along with other stuff (a SAN bay, a switch, a PC server) that would be 320 cores here.

 

[Blazkowicz]

AMD's slides indicate a clear dilineation between the shared-nothing density server market this ARM chip is going into and anything media-related. That has a GPU and x86 only.
It doesn't mean it can't be done, but AMD could barely handle the workload of handling the software stack it had--again, back when it had money.

Yes, AMD will sell you an underclocked 2 core Jaguar if you need something low power, else a desktop Bulldozer version n APU or an Opteron APU.
The ARM Opteron is for web servers and generic VMs, where an Opteron APU has a perfectly useless GPU (which could be used to beam remote, 3D accelerated and/or GPGPU accelerated applications to users)

 

[cal_guy]

AMD is boasting about interconnexions with the "Freedom fabric" (which funnily is not free as it won't be licensed to others, or at a whim like Intel's QPI)

http://www.theinquirer.net/inquirer...cense-seamicros-freedom-fabric-to-cpu-vendors

16 core looks probable, and dual socket motherboards and systems I'd say. AMD probably intends you to pile up 1Us of this thing, if you stack ten dual socket units in a cabinet along with other stuff (a SAN bay, a switch, a PC server) that would be 320 cores here.

They won't go into a traditional rack server but rather they will go onto those SeaMicro PCI-E cards with storage and ethernet virtualized.

http://www.anandtech.com/show/3768/...ith-512-atom-cpus-and-low-power-consumption/2

AMD's slides indicate a clear dilineation between the shared-nothing density server market this ARM chip is going into and anything media-related. That has a GPU and x86 only.
It doesn't mean it can't be done, but AMD could barely handle the workload of handling the software stack it had--again, back when it had money.

I'm more interested in seeing if they mention to the analysts where they're getting money to last until 2014. The layoffs are cutting into necessary functions AMD simply can't afford.

With the exception of the final payment for the GF 28nm thing (which should be offset with an long-term investment coming in) AMD doesn't have any Notes due unto 2015 so AMD should be on fairly firm ground until then.

 

[Alexko]

With the exception of the final payment for the GF 28nm thing (which should be offset with an long-term investment coming in) AMD doesn't have any Notes due unto 2015 so AMD should be on fairly firm ground until then.

What are you referring to?

 

[cal_guy]

What are you referring to?

It's just corporate bonds that AMD owns that are due in 2013. It's about $140 million.

 

[Blazkowicz]

They won't go into a traditional rack server but rather they will go onto those SeaMicro PCI-E cards with storage and ethernet virtualized.

http://www.anandtech.com/show/3768/...ith-512-atom-cpus-and-low-power-consumption/2

Alright, thus it makes little sense to speak of multi-socket systems, instead you have mulitple independant systems accessing a common bus.

There still ought to be a self-contained 1U system for low end, we can't all buy expensive blade-like things.

 

[Alexko]

It's just corporate bonds that AMD owns that are due in 2013. It's about $140 million.

OK, thanks.

 

[Raqia]

More news on this front:

http://www.anandtech.com/show/7991/amd-is-also-working-on-a-new-64bit-x86-core

Jim Keller joined Mark Papermaster on stage at AMD's Core Innovation Update press conference and added a few more details to AMD's K12 announcement. Keller stressed AMD's expertise in building high frequency cores, as well as marrying the strengths of AMD's big cores with those of its low power cores. The resulting K12 core is a 64-bit ARM design, but Jim Keller also revealed that his team is working on a corresponding 64-bit x86 core.

The x86 counterpart doesn't have a publicly known name at this point, but it is a new design built from the ground up.

 

[Alexko]

It's been ages since I've heard anything about K12. Is there any indication that the project is still alive?

 

[Deleted member 13524]

Last I heard the only indication they gave was that is had been put on low priority.
With Zen APU SKUs starting at 4W TDP, I'm wondering if they haven't made K12 redundant.

 

[xEx]

I think needs to finish k12 and use it in smartphones and tablets. that market grow every year and AMD can compete against QC and bring some of the tech they already use in their CPUs.

 

[Blazkowicz]

Smartphone has zero-ish margins, many competitors, higher risk to fail to capture a part of the market, high SoC churn rate combined with a need for low end / mid range / high end SoCs. AMD knows nothing about integrated modems (Nvidia had a try, it ended up as a waste of money).

Tablets are more forgiving, but a cramped market too with not much growth. Many vendors and products there again. Intel/Nvidia still are there : Intel got out of smartphones and uses x86 that's the same as on netbooks or some embedded Atom basically. Nvidia sells a few Tegra X1, Chrome Pixel C and nothing else, again a reuse of a chip that's otherwise for high end embedded (cars, airplanes) and consoles.
Thus, I believe any AMD tablets we'll see will use Zen+ 2c/4t APU (or is there a 4W Zen before that one?) : reuse a chip that's for netbooks, tiny desktops, 12" MacBook-like things on tablets like the Atom tablets that use Windows 10, Android or dual boot.

Work on an ARM + GCN APU was terminated a while ago and I'd expect any further one exists in a very limited way on paper.
For K12 I would guess the plan is for AMD to wrestle out a share in big traditional x86 servers first (ARM and things like 512 Atom cores in a blade just failed to make a dent in the cheap and powerful dual-socket Xeon market) then sell some K12 server chips for networked storage appliances and the like.

 

[xEx]

Tell that to QC who is more value than intel right now...

The SP/Tablet markets are growing PC is shrinking and socs for smartphones can be use in different other things plus the image value of having ur brand in more products. Nvidia is a GPU maker, had no experience or tech for CPU and didn't want to compete with QC directly either. Intel tried to make x86 cpus but couldn't compete with QCs socs.

You don't need to be a theoretical physics to know that if ur future aims to a shrinking market then you have a problem. SP will eventually becomes the PCs of the future and gaming PC will also disappear in favor of online streaming games so AMD needs to start switching markets as soon as possible to have a presence in future markets rather than condem- to-dead present ones.

 

[Gubbi]

Tell that to QC who is more value than intel right now...

INTC market cap: $173B
QCOM market cap: $71.6B

Also, QCOM makes almost all of their profits from patent licensing, margins on their SOCs are razor thin.

Cheers

 

[xEx]

INTC market cap: $173B
QCOM market cap: $71.6B

Also, QCOM makes almost all of their profits from patent licensing, margins on their SOCs are razor thin.

Cheers

Oh Intel retake the lead over QC. either way the principle is the same, AMD can make tons of money in the AMR space and as I said SP and TB are the future and unless AMD wants to become a kind-of-IBM company who only serves enterprises sooner or later they will have to take that path.

 

[Gubbi]

Oh Intel retake the lead over QC.

????
Financially, Intel is more than twice the size of Qualcomm, and always has been.

And, as I said, Qualcomm gets most of their profits from their patent licensing, which will be drastically reduced in the upcoming 5G era.

Cheers