AMD Announcement: ARMv8 Opterons In 2014

[Blitzvogel]

This. x 10.
I think this will be a major winning strategy in the future.

Unfourtunately that pretty much also means withdrawal from the dedication to it's current major markets.
Which won't exactly be great for competition and technology.

If we look at AMDs current core types:

Low end/Power Efficient x86 block: Jaguar x86 core

Middle/High end x86 block: Piledriver/Steamroller x86 core

Graphics block: GCN

Future low end CPU block: ARM cores

While they are only semi-modular and are constantly being researched and improved, if we look at them as such, the picture becomes more clear in terms of what AMD could do or plan. Integration might be expensive or time consuming, but with research and development from a company looking for a customized product, it could be a big win for AMD.

AMD doesn't necessarily have to leave core/dedicated markets. It perhaps means that their dedicated products may suffer a bit in order to be modular, but AMD should do well if they dedicate enough engineers to "hand laying out" blocks in order to minimize wasted die area.

I wonder if AMD will design their own ARM cores. Even without dedicated ARM cores, AMD could immediately challenge Nvidia with ARM + GCN, and if x86 actually begins going the way of the Dodo, AMD will have a way out because they will have plenty of experience with ARM integration and engineering.

There is so much potential AMD could have with an ARM license, be it integration or custom ARM core design a la Qualcomm.

"Graphics, x86, ARM. We are AMD." - That's a big statement.

It does bring up thoughts on where Fusion is actually going. I'm not sure if AMD is focused on "true" integration of the graphics/GPGPU processing system into x86 cores or not. Obviously, a block system goes against this......

 

[mrmt]

AMD has an x86 license, and they have already created the IP blocks involving x86 to a degree that is unrivaled in that power footprint - neither Intel nor Via can touch them at this time in that segment (that may change with future Atom iterations)

All this talk about ambidextrous architecture to me means that they are going to throw the towel in the x86 market in the medium term. They simply cannot afford the fight with Intel and staying there would mean be sandwiched between the clash of the two giants, x86 and Intel.

The scenario they are making this calculations,l which fits very well the appalling financial assumptions they are working for Q313 onwards is something like this:

- Piledriver starts from a worse point than Llano in margin and MSRP (this already happened)

- Haswell pushes Piledriver further down into the value chain

- Haswell also checks Kabini on its upper market bracket

- Steamroller is late and when it comes it doesn't improve the situation very much. It will be toast against Broadwell in any case.

- Silvermont arrives and checks Kabini from the bottom.

So they need an exit from the PC market, ASAP. That's why I think they are working on embedded. Embedded is always high volume but lower margins. Also once you commit yourself to a supplier you don't change it for the product life cycle. They are also trying to make something out of Seamicro. I think they will fail. Intel will sooner or later push Silvermont on servers and there is Calxeda with more backing with AMD, so they may end up a master of none. Distant second x86 maker, distant 2358th ARM maker.

 

[ViRGE]

Yes all very puzzling. If AMD's R&D budget is too small to compete with Intel (although I'm sure Bulldozer used up a fair budget), why do they think they can compete in a market they have no experience in? Looks like someone on the board has got either a tablet or ARM obsession.

The board as a whole clearly has tablet fever. Given how hard AMD is gunning for tablets, I don't think there's any doubt about that at this point.

 

[mrmt]

IWhile they are only semi-modular and are constantly being researched and improved, if we look at them as such, the picture becomes more clear in terms of what AMD could do or plan

Isn't this reusable blocks, modular design AMD marketing bullshit, or at least a far fetched goal? If AMD designs are so modular, so easy to switch parts, why they take so long to refresh their line up with so modest improvements? Why Trinity sucked, and why take so long to update Brazos?

Modular might work for AMD GPUs, which they can be on schedule with refreshing and they do that switching core and uncore on different cycles, but their CPU? If there is something not modular at all on AMD is AMD CPUs. They are always late and take more time than they say they need even for small improvements.

 

[cbn]

But AMD might decide to go after the handset market, ala qualcomm and broadcom style, and for that they would do fine just dropping in a bog-standard ARM core surrounded by their own IP (specifically the GPU IP).

As far as handsets or fanless Tablets go, wouldn't it make sense for AMD to keep x86 and emulate ARM instruction set?

In this way they wouldn't AMD have an advantage over other fabless ARM players who don't have x86. (This could become a stronger advantage than we think if Intel ever takes off in handsets....which I think is definitely possible due to its projected widening process lead and its SOC integrated digital radio development.)

P.S. Speaking of Intel, really looking forward to seeing how the explanation for their xtor design at 22nm for Silvermont. According to the Anandtech 3D xtor article here adding or removing fins can alter the drive current.

It's also possible to vary the number of fins to impact drive strength and performance, allowing Intel to more finely tune/target its 22nm process to various products.

 

[MarkLuvsCS]

A big problem with them getting into a new market with new tech is their software integration. Enduro is still gravely lacking even though a year ago they released their new tech. If you think nvidias tegra platform was a bust, how can amd hope to cope when software plays a major part of mobility platform. I think a big part of that problem comes back to money. I'm sure spreading into so many markets spreads them pretty thin.

 

[pelov]

This is the part I don't get. I understand why everyone else on the planet has to pursue using bog-standard A15 cores embedded in their SoC, including Calxeda, but I don't understand why AMD would.

AMD has an x86 license, and they have already created the IP blocks involving x86 to a degree that is unrivaled in that power footprint - neither Intel nor Via can touch them at this time in that segment (that may change with future Atom iterations).

That's not true, as SeaMicro is still selling Atom-based products. So the atoms offer very good perf-per-watt. Where atom falls apart is the GPU... or lack of it ;P If AMD offers Jaguar with HSA improvements it would certainly be a very good sell and fill a significant niche in a growing market.

So they need an exit from the PC market, ASAP. That's why I think they are working on embedded.

They're too far behind to catch up, both in performance/R&D and in process disadvantage, unless they pursue it heavily in the next 2-3 product cycles. That just doesn't make sense from a economic standpoint, as they would be pursuing head on a market that's shrinking.

If they can start licensing their GPU cores, they could be on to something. If they decide to slap together an ARM SoC in-house, that could potentially be pretty good as well.

I think we'll see this flexibility that Read has been preaching so much about. Frankly, I think it's a good idea. Whether it's a case of too little too late, though, I don't know :/ It certainly looks like it.

 

[Barfo]

Announcement leaked:

 

[Blitzvogel]

Isn't this reusable blocks, modular design AMD marketing bullshit, or at least a far fetched goal? If AMD designs are so modular, so easy to switch parts, why they take so long to refresh their line up with so modest improvements? Why Trinity sucked, and why take so long to update Brazos?

Modular might work for AMD GPUs, which they can be on schedule with refreshing and they do that switching core and uncore on different cycles, but their CPU? If there is something not modular at all on AMD is AMD CPUs. They are always late and take more time than they say they need even for small improvements.

Core design revolves around multiple requirements including actual IPC, GFLOPS, power, TDW, etc. I would assume AMD didn't need to update Brazos because the design was already working so well in their favor. It's no wonder they cancelled Krishna, and decided to just focus on Jaguar while they still managed to sell Bobcat CPUs.

"Blocks" would more or less be established core designs that can be fitted in combination with more of the same blocks or different ones while maintaining a coherent communication and processing structure (which isn't easy, but not impossible) much like ARM SoCs. AMD would still have to heavily engineer the integration side of making for example 2 Jaguar Cores + 1 ARM Security/compatibility core + 192 GCN Stream processors to make it all work together + software.

 

[mrmt]

"Blocks" would more or less be established core designs that can be fitted in combination with more of the same blocks or different ones while maintaining a coherent communication and processing structure (which isn't easy, but not impossible) much like ARM SoCs. AMD would still have to heavily engineer the integration side of making for example 2 Jaguar Cores + 1 ARM Security/compatibility core + 192 GCN Stream processors to make it all work together + software.

Ok, but how would this work at the foundry level? They would still have to do this heavy engineering, send to the foundry, tape out, debug... I don't think they can take short cuts here, am I correct?

 

[CHADBOGA]

Announcement leaked:

:biggrin:

 

[Olikan]

Ok, but how would this work at the foundry level? They would still have to do this heavy engineering, send to the foundry, tape out, debug... I don't think they can take short cuts here, am I correct?

no and yes...

when AMD desing a GPU, it will use a similar aproach of PowerVR, small "gpu cores"... they tape out just one "gpu core", learn about the design and foundry...and then, glue several of the "cores" and make a flagship...

for new products, it saves money for R&D, but lose time to market... making it very bad if they wanna be competitive with nvidia... for old IP products, it very fast and cheap ( like the wiiU )

AMD probably will still go for a gpu flagship first, with several "gpu cores" in mind... when they wanna make a mid-range gpu, just use less "cores", easy and fast...right?

 

[anongineer]

Some work is saved if certain blocks are hard macros, so the transistor layout is fixed for a certain process. Graphics and CPU cores would probably fall into this category - take previous claims of these things being "synthesizable" with a grain of salt, because if they're smart about it they'll have locked down the layout to reduce time to market.

Being in design services could be brilliant. At least you can charge other people for work performed instead of selling to retail or OEM's. But getting business depends on reputation. APU's are basically SOC's, so AMD has been doing this integration stuff for a while. Has their track record in this area been confidence inspiring?

Now assume a 15% reduction in the people who are supposed to do this work.

 

[sontin]

no and yes...

when AMD desing a GPU, it will use a similar aproach of PowerVR, small "gpu cores"... they tape out just one "gpu core", learn about the design and foundry...and then, glue several of the "cores" and make a flagship...

for new products, it saves money for R&D, but lose time to market... making it very bad if they wanna be competitive with nvidia... for old IP products, it very fast and cheap ( like the wiiU )

AMD probably will still go for a gpu flagship first, with several "gpu cores" in mind... when they wanna make a mid-range gpu, just use less "cores", easy and fast...right?

So, what's different from today? GPUs are modular since 10 years. They have independent hardware blocks, rasterizer and memory controllers.

 

[Olikan]

So, what's different from today? GPUs are modular since 10 years. They have independent hardware blocks, rasterizer and memory controllers.

you should ask amd this

 

[pelov]

So, what's different from today? GPUs are modular since 10 years. They have independent hardware blocks, rasterizer and memory controllers.

I think the difference could potentially be the licensing of these IP blocks/GPU to someone else. Generally speaking, if you wanted to buy an AMD GPU, you had to buy a discrete GPU or an AMD APU, but they might be looking to change that.

 

[sontin]

I think the difference could potentially be the licensing of these IP blocks/GPU to someone else. Generally speaking, if you wanted to buy an AMD GPU, you had to buy a discrete GPU or an AMD APU, but they might be looking to change that.

And why would a)AMD do this and b) why would anybody do this?

I mean you get more money from hardware instead of IP. Look at ARM. Every quarter billions of ARM cores sold by their partner but yet the revenue is even under the one of nVidia.

I see no future in licensing of the GPU tech.

 

[Blitzvogel]

And why would a)AMD do this and b) why would anybody do this?

I mean you get more money from hardware instead of IP. Look at ARM. Every quarter billions of ARM cores sold by their partner but yet the revenue is even under the one of nVidia.

I see no future in licensing of the GPU tech.

If you can develop a substantial product that offers features and/or performance beyond the competition for the same price/power/etc then why not, *unless* AMD can manage to develop excellent SoCs themselves that everyone wants.

 

[sontin]

If you can develop a substantial product that offers features and/or performance beyond the competition for the same price/power/etc then why not, *unless* AMD can manage to develop excellent SoCs themselves that everyone wants.

And AMD can do this?
They lost the Perf/Watt battle against GK107 and the result is a massiv notebook share decline. They lost the Perf/Watt battle against GK104 and the result is that nVidia owned them in their MultiGPU game...

AMD has a better Chip with Pitcairn. But Pitcairn is to big and power hungry for 90% of the notebook market and to slow to win against GK104.

They can't beat nVidia and yet they think people will license their IP for ARM SoCs?!

 

[AtenRa]

Isn't this reusable blocks, modular design AMD marketing bullshit, or at least a far fetched goal? If AMD designs are so modular, so easy to switch parts, why they take so long to refresh their line up with so modest improvements? Why Trinity sucked, and why take so long to update Brazos?

1: Mobile Llano was launched 14th of June 2011, Trinity(mobile) was launched 15 May 2012, only 11 months away.

2: Trinity is better than Core i3 and it has the most advanced iGPU today. You are the only one to mention that Trinity Sucks

3: Brazos had no competition from Intel, AMD could afford to use 40nm for Brazos 2 and go 28nm for Kabini.

 

[LogOver]

3: Brazos had no competition from Intel, AMD could afford to use 40nm for Brazos 2 and go 28nm for Kabini.

Brazos sucks in a Big Way. AMD is pushing it in a "full size" notebook market where it sells for the same price, as SB-based Pentiums, which are not only by order of magnitude faster than Brazos on CPU side, but also faster on GPU side. Speaking about netbooks - you can already buy cheap Celeron based netbooks (Acer AO756) which will eat Brazos for breakfast.

 

[Idontcare]

1: Mobile Llano was launched 14th of June 2011, Trinity(mobile) was launched 15 May 2012, only 11 months away.

2: Trinity is better than Core i3 and it has the most advanced iGPU today. You are the only one to mention that Trinity Sucks

3: Brazos had no competition from Intel, AMD could afford to use 40nm for Brazos 2 and go 28nm for Kabini.

I agree with (1) and (2).

I cannot agree with (3). Node shrinks are done to enhance gross margins and improve profitability. AMD was not in any position to "afford to use 40nm" for Brazos 2.

Even if the ASP of Wichita and Krishna had not increased relative to Brazos 2, the cost/chip would have decreased and AMD would have made more money (or loss less per chip as it were).

Instead they didn't get to enjoy the gross margin benefits of shrinking Brazos, and they had to turn around and pay hundreds of millions of dollars to GloFo to negotiate out of their 28nm-exclusivity contract for Kabini production.

Possibly AMD will consider making some of its next generation APUs at Taiwan Semiconductor (TSM). However, the arrangement involves a cash payment of $425 million to Globalfoundries, to be paid over the next five quarters, and AMD will take a $703 million charge in the March 2012 quarter.

source

(3) is just more proof of how poorly AMD was managed as a fabless company heading into sub-40nm territory.

Not a single other fabless company had the kind of mis-execution and financial fallout that AMD did in going to sub-40nm. Others had bumps, but they still made money and managed to negotiate foundry contracts that were not so deleterious to their own business model as to cost them hundreds of millions to break out of in desperate moves of survival.

On top of that you have the financial fact that AMD was not in any position of being able to afford a misstep or two in transitioning to sub-40nm, and all the other companies who sell 32nm and 28nm products have the gross margins and profits to show from their management and execution.

 

[ViRGE]

(3) is just more proof of how poorly AMD was managed as a fabless company heading into sub-40nm territory.

Not a single other fabless company had the kind of mis-execution and financial fallout that AMD did in going to sub-40nm. Others had bumps, but they still made money and managed to negotiate foundry contracts that were not so deleterious to their own business model as to cost them hundreds of millions to break out of in desperate moves of survival.

On top of that you have the financial fact that AMD was not in any position of being able to afford a misstep or two in transitioning to sub-40nm, and all the other companies who sell 32nm and 28nm products have the gross margins and profits to show from their management and execution.

Is that an AMD problem or a GloFo problem though? I haven't been hearing anything good about GloFo's 28nm production line, which would absolve AMD to some extent and validate paying them to break the contract.

 

[Idontcare]

Is that an AMD problem or a GloFo problem though? I haven't been hearing anything good about GloFo's 28nm production line, which would absolve AMD to some extent and validate paying them to break the contract.

Even if it is a GloFo execution issue, it doesn't absolve AMD's management for having gotten AMD into the precarious and risky position of being dependent on their foundry, as a fabless company, as they did with GloFo and 32nm/28nm.

These are amateur mistakes for a fabless company to be making, if you look at the management decisions of the top-10 fabless companies you won't find a single one of them making these kinds of blunders (because usually it is fatal to the fabless company long before they become large enough to become a top-10 business).

This is what I mean when I say AMD did a very poor job of managing themselves as a fabless company in their sub-40nm transition and it shows.

The under-supply then excessively over-supplied inventory issue with Llano. The complete miscalculation and then cancellation of wichita and krishna is another. The hefty payment to Glofo just to get out of an essentially worthless (by then) 28nm exclusivity contract is yet another.

Fabless companies know they must not only work hand-in-hand with their business partners (the foundries) in dynamic fashion - there is a lot of iterative learning that flows back and forth that goes into defining design rules as well as engineering tradeoffs for the electrical properties of specific components including DFM models - but that they also must plan for fallback plans (contingency plans) if things do not turn out all rosy per Plan A.

The part of AMD that once upon a time knew how to successfully manage a fabless company (ATI) would have known all this, the part of AMD that appears to have made critical decisions at AMD leading up to 32nm and 28nm appear to have underestimated the criticality of this portion of their job as managers of a then fabless company and it shows in their misexecution and financial situation today. (not Rory's fault, he inherited this mess)

 

[podspi]

yes, it boggles my mind how if cash is changing hands, AMD is paying GF and not the other way around.

Still, if GF goes away where would AMD go? I'm sceptical TSMC has enough capacity to serve AMD completely.