AMD post-Bulldozer x86 CPU architecture

[NostaSeronx]

WSA last to 2024.

WSA lasts till 2019 with an option to go 10 more years.

E.g. what are the terms, does AMD have to buy a certain number of wafers per year? Are they free to also buy wafers from other suppliers? What options exist in that case?

AMD has to buy a certain amount of wafers from GlobalFoundries. The node exclusivity has been removed but there is a time limit on how long AMD can stick to a node.

AMD is free to port designs to TSMC, UMC, STMicroelectronics, and Samsung. While that will be an added cost towards AMD it isn't included in the WSA.

 

[SlowSpyder]

Does anyone know the terms of the WSA? Are console APU's made at TSMC or Global Foundries?

 

[NostaSeronx]

Does anyone know the terms of the WSA?

It is mostly a take and pay contract. GlobalFoundries says it will produce x amount of wafers at y cost. AMD then has to purchase x amount of wafers at y cost. If AMD does not purchase x amount of wafers instead purchases z amount of wafers. AMD would then need to pay the left over amount of y cost.

Are console APU's made at TSMC or Global Foundries?

AMD is not paying for the research and development costs. The research and development costs are paid upfront by Sony and Microsoft. It is up to Sony and Microsoft to choose the foundries they want their consoles to be ported to.

Sony and Microsoft might want to port to 28nm FDSOI. So the foundries that will be producing that are;
GlobalFoundries
Samsung
STMicroelectronics

 

[ShintaiDK]

WSA lasts till 2019 with an option to go 10 more years.

http://sec.edgar-online.com/advance...tatement-definitive/2014/03/25/section18.aspx

The WSA terminates no later than March 2, 2024

 

[mrmt]

Does anyone know the terms of the WSA? Are console APU's made at TSMC or Global Foundries?

The terms disclosed are in the link Shintai posted. The WSA is an exclusivity agreement coupled with a take or pay charge, meaning AMD can only manufacture CPUs with GLF *and* is bound by a minimum purchase commitment.

The console APUs are not currently manufacture at GLF, but will move to them this year, so the minimum commitment is more or less a non-issue for AMD this year.

AMD could manufacture the cat cores at TSMC because of a waiver in the agreement. This waiver was killed in 2012, so GLF is back porting or phasing out everything manufactured outside Globalfoundries. By end of 2014 GLF should be AMD's sole source for CPUs, and some GPUs.

 

[Fjodor2001]

Seems like a very risky deal to enter into for a fabless company like AMD. How can they sign up for such an agreement, where basically their future may depend completely on the success or failure of an external company with non-exceptional track record?

I could understand it if the terms were that AMD signed up to buy a certain number of wafers per year, but that they were still free to buy wafers from other foundries too if desired. But for AMD to only be allowed to buy wafers from GF from 2012/2014 until 2024 is a very high risk. (As the WSA says "Pursuant to the WSA, we [AMD] are required to purchase all of our microprocessor and APU product requirements from GF with limited exceptions.").

I do however notice that three amendments have been added to the WSA, and a fourth is in negotiation. So I guess anything is up for discussion. But if AMD would like to make changes to the WSA through amendments, they most certainly will have to pay for it.

 

[Ajay]

I would expect HBM, not EDRAM.

I think, they need one or the other. HBM would better if AMD can add it economically - even if it's their mainly for the iGPU (assuming, even with HSA, there are still direct memory mapping options).

 

[NostaSeronx]

@Fjodor2001, a high majority of employees at GlobalFoundries are ex-AMD. As of late, GlobalFoundries has been producing nodes specifically for AMD. These specific nodes are usually "Super" for 32nm and 28nm and "Plus" for 20nm and beyond.

 

[Ajay]

Always good to have a financial perspective on here. Not enough of this type of business discussion goes on in these forums, unfortunately, as these considerations above all else drive the products we see.

I appreciate mrmt's comments as well. Most of us here are from a technical as opposed to financial background. Just as IDC has educated us in process node development, mrmt has helped us understand some of the financial relationships and others have helped us with CPU designs considerations, software issues and broader market understanding (😉 ). It definitely takes a village - at least to have enough info to be dangerous 🙂

 

[Phynaz]

Seems like a very risky deal to enter into for a fabless company like AMD. How can they sign up for such an agreement, where basically their future may depend completely on the success or failure of an external company with non-exceptional track record?

I could understand it if the terms were that AMD signed up to buy a certain number of wafers per year, but that they were still free to buy wafers from other foundries too if desired. But for AMD to only be allowed to buy wafers from GF from 2012/2014 until 2024 is a very high risk. (As the WSA says "Pursuant to the WSA, we [AMD] are required to purchase all of our microprocessor and APU product requirements from GF with limited exceptions.").

I do however notice that three amendments have been added to the WSA, and a fourth is in negotiation. So I guess anything is up for discussion. But if AMD would like to make changes to the WSA through amendments, they most certainly will have to pay for it.

The CEO of AMD at the time - Hector Ruiz - negotiated the deal. As part of the deal he became the CEO of Global Foundries.

What better way to ensure his success at Global Foundries than to have his former company financially chained to him? He guaranteed himself a revenue stream.

 

[sirroman]

Minimum 1000. More likely closer to 2000.

Want to guess how many of those people have even met Keller (even once)?

Let alone being in the position of working hand-in-hand on a somewhat regular basis in the creation of specific circuits or logic blocks?

Keller himself said that now fewer people do the actual HW microarchitecture design. Most of the team codes the SW that simulates, lays down the transistors and tests it all.

 

[mrmt]

Keller himself said that now fewer people do the actual HW microarchitecture design. Most of the team codes the SW that simulates, lays down the transistors and tests it all.

What's the scope complexity of what AMD is developing now is comparatively smaller than the previous projects Keller worked in AMD? By the time Keller worked at AMD, the company was developing a bleeding edge, top of the line x86 core using technology from DEC, which was in another category compared to its peers.

Now they are developing something based on A57 and probably a "me too" x86 core. Add that, plus the synthesis tools, and what you could get the reason.

 

[Ajay]

What's the scope complexity of what AMD is developing now is comparatively smaller than the previous projects Keller worked in AMD? By the time Keller worked at AMD, the company was developing a bleeding edge, top of the line x86 core using technology from DEC, which was in another category compared to its peers.

Now they are developing something based on A57 and probably a "me too" x86 core. Add that, plus the synthesis tools, and what you could get the reason.

You've lost me a bit here.

The scope of complexity for a modern APU is huge compared to, say, the K8. A "me too" x86 uArch isn't that surprising for a mature market (shared IP, similar design tools, etc.). Intel has several substantial and carefully cultivated advantages over AMD (being an IDM is just one). The last time around, AMD gambled heavily on BD and lost - they literally cannot afford to gamble again.

Over a decade ago I worked as a firmware engineer for high-end switching products. I happened to have a cube in the ASIC design area for a while b/4 space was found for me with the other software engineers. I was amazed that they were writing simulation design code basically in 'C' (Verilog). The ASIC designer's spent most of their time coding, just like me.

Now CPUs are becoming so complex that not even Intel can afford the army (or management nightmare) that would be needed to hand design everything.

I wouldn't just pish/posh AMD out of hand if they are designing an APU uArch. I very much doubt they will achieve Intel's state-of-the-art CPU performance; but improved perf & perf/watt over Excavator plus best in class iGPU performance could be enough to lure some new customers to AMD - or not, it's just way to early to tell.

 

[mrmt]

I wouldn't just pish/posh AMD out of hand if they are designing an APU uArch. I very much doubt they will achieve Intel's state-of-the-art CPU performance; but improved perf & perf/watt over Excavator plus best in class iGPU performance could be enough to lure some new customers to AMD - or not, it's just way to early to tell.

Just look at the beating Nvidia is taking with Tegra. iGPU is not enough to drive customers towards Nvidia solution, why should it work for AMD? The main issue for AMD, far more critical than the performance of the core (assuming they don't pull another bulldozer from their hat) is not what the IP they have, but the IP they haven't. SATAe, modem, wireless radio, ring bus, security, just to name a few. Licensing tech from others alleviate this issue but doesn't solve it.

As much as AMD is trying to position itself as going beyond x86 traditional markets, lack of IP limits their ability to go to the more lucrative markets. What's left for them is 10-20% markets like console chips.

 

[Ajay]

Just look at the beating Nvidia is taking with Tegra. iGPU is not enough to drive customers towards Nvidia solution, why should it work for AMD? The main issue for AMD, far more critical than the performance of the core (assuming they don't pull another bulldozer from their hat) is not what the IP they have, but the IP they haven't. SATAe, modem, wireless radio, ring bus, security, just to name a few. Licensing tech from others alleviate this issue but doesn't solve it.

Leading edge graphics performance doesn't seem to be as important in smartphones as one would think. Good enough, or simply better, for more expensive phones, seems to be all that's needed. That, and the big players are already set with preferred vendors, so their only chance is small players and they are too expensive then, IIRC.

My understanding from the out of court settlement between AMD and Intel is, among other things, and extensive cross license. iGPU hasn't been a main differentiator for anyone so far. If AMD can leverage it's GPU technology to make a solid entry level gaming machine (something not accomplished yet) - I think they will have a market opportunity with some OEMs. I don't have a crystal ball, so I don't know if they can manage this, but a new uArch and what may be a solid 14nm node @ GF means that there is some hope.

Please note, I like to root for AMD just like I root for other underdogs (tech, sports, whatever). I still have some fondness for the company that created the K7 and introduced 64 bit technology in the K8. find it enjoyable to hold out some hope. That said, I wouldn't invest a dime in AMD. Some like to play ups and downs to turn some coin, but I'm a longer term investor and wouldn't touch it.

If I looked at AMD with my skeptical investor squint eyes - I'd consider the company a lost cause.

 

[mrmt]

My understanding from the out of court settlement between AMD and Intel is, among other things, and extensive cross license. iGPU hasn't been a main differentiator for anyone so far. If AMD can leverage it's GPU technology to make a solid entry level gaming machine (something not accomplished yet) - I think they will have a market opportunity with some OEMs. I don't have a crystal ball, so I don't know if they can manage this, but a new uArch and what may be a solid 14nm node @ GF means that there is some hope.

This is what I was talking about. We can see AMD entering the console gaming market. We also can see they entering in the micro server market, but we can't see them going much further than these adjacent markets.

Just look at what you are saying here: Gaming. AMD is dominating the *console* gaming market, but the *pc* market remains firmly in Intel's and Nvidia's hand, and this market is changing: The most important gaming devices out there are not the consoles IMO (they can't even decisively bea the old PC in terms of revenue), but phones and tablets. The sheer quantity of devices (which becomes more and and more powerful each generation) will create a market for gaming big enough to be as significant as the console markets, if not outright surpass it. And yet AMD have actually 0 share on this market, and don't have the IP to break into this market. Buying an AMD tablet is to buy top class GPU performance, but also buy bulkier, heavier power hogs that will be worse than vanilla tablets in almost every other aspect you measure. And AMD buying second rate IP like ASmedia SATAe alleviates the issue but doesn't solve it, because performance is still second rate, plus they have to *pay* for it.

All in all, I don't see them going much further away from the traditional PC market, because they are ill-equipped to deal with most of the embedded market. And the embedded market is no panacea, the margins there are really, really low. It's not for anything that Nvidia only ran to the embedded car market with Tegra only after they failed to break into the phone, server and tablet market. Let's see where's AMD's car market equivalent.

 

[chrisjames61]

AMD will be back, intels days as top dog are numbered. Apple has already destroyed them in mobile with little effort. Intels whole business model is based on exploiting manufacturing advantages, sadly for them those advantages are further between every time. Once Intel has no process advantage, thier lack of innovation will allow AMD to catch up


X86 is in it's final days. When apple releases it's own processors intel will become just another ARM manufacturer with x86 for servers and old gamers

Let's skip the doomsday projections and hyperbole for both sides, 'k?
-ViRGE

Intel has no competition in the x86 space. Their competition is ARM. They know it. Now they are playing catch up. They could have owned the mobile space too but they blew it. Now ARM is entrenched and it will be hard to displace. I think Apple will eventually release an ARM MacBook Air. Then other OEM's will follow suit. It won't bode well for Intel.

Which has absolutely nothing to do with AMD. Please stay on topic.🙂
-ViRGE

 

[AtenRa]

Intel has no competition in the x86 space. Their competition is ARM. They know it. Now they are playing catch up. They could have owned the mobile space too but they blew it. Now ARM is entrenched and it will be hard to displace. I think Apple will eventually release an ARM MacBook Air. Then other OEM's will follow suit. It won't bode well for Intel.

I beg to differ,

There maybe no competition in the High-End Desktop x86 segment but the fight has only been started for supremacy of the Low Power Tablet realm(too medieval ??).

http://www.anandtech.com/show/7974/...tecture-a10-micro-6700t-performance-preview/3

Looking at single threaded performance we immediately see the benefits of AMD's new boosting capabilities. The Puma+ cores are 35% faster than Intel's Silvermont cores, and can deliver nearly 80% of the performance of AMD's Piledriver cores found in Trinity.

Hell, and this is from 28nm bulk vs 22nm FinFet.
Ooh and not forget that both Mullins and Baytrail are almost the same in Die size (~110mm2). That means that AMD clearly has the better design because 28nm has lower density than Intel's 22nm FinFet.

And now the good part,

No contest here in the GPU department, Intel is way behind.

Arg! Can you guys please stick to AMD?
-ViRGE

 

[witeken]

Hell, and this is from 28nm bulk vs 22nm FinFet.
Ooh and not forget that both Mullins and Baytrail are almost the same in Die size (~110mm2). That means that AMD clearly has the better design because 28nm has lower density than Intel's 22nm FinFet.

Don't forget that the CPU cores themselves are only a small fraction (<5mm²) of the die area, so don't draw your conclusions too fast. Also, the density advantage of 22nm is only 10%.

 

[AtenRa]

Don't forget that the CPU cores themselves are only a small fraction (<5mm²) of the die area, so don't draw your conclusions too fast.

Well, smaller CPU cores with higher performance at a density disadvantage and still AMDs Puma is faster than Intel BayTrail. If that doesn't qualify for a better design then what is ??

Also, the density advantage of 22nm is only 10%.

How did you come to that conclusion ??
Even if that is true, the fact remains the same, AMDs design is better because they are using a lower density process.

 

[Enigmoid]

Puma looks good but we have no idea how much power it actually uses. I think a scaled up jaguar core (puma is essentially the same design with power gating and turbo) may do quite well.

For the gpu Puma does better but doesn't really allow for playable games. ULV HD 4000 is painful and puma is significantly less powerful than that.

 

[Sweepr]

I beg to differ,

There maybe no competition in the High-End Desktop x86 segment but the fight has only been started for supremacy of the Low Power Tablet realm(too medieval ??).

http://www.anandtech.com/show/7974/...tecture-a10-micro-6700t-performance-preview/3

Hell, and this is from 28nm bulk vs 22nm FinFet.
Ooh and not forget that both Mullins and Baytrail are almost the same in Die size (~110mm2). That means that AMD clearly has the better design because 28nm has lower density than Intel's 22nm FinFet.

Arg! Can you guys please stick to AMD?
-ViRGE

Good job at handpicking the CB 11.5 ST benchmark. Now let's take a look at MT performance:

Almost same score as old Atom Z3770.
Oh, and that's the fastest Mullins part (up to 2.2GHz), the second fastest is a much slower 1.6GHz max Turbo part with 350MHz graphics, probably slower than recent Bay Trail-T C0 parts in most benchmarks (especially MT) and with greatly reduced graphics perf. advantage.

And now the good part,

No contest here in the GPU department, Intel is way behind.

Yes, Haswell-Y is also way faster than BT-T, but you know what? Bay Trail-T does fit in very slim/thin 7-10'' tablets and draws quite a bit less power than that platform, what about the bulky/huge 11.6'' Discovery Tablet that AMD sent out for testing? Did they allow power cunsumption tests for Mullins? 🙂
Also, iGPU performance was never one of BT's strenghts, the EU count is way low and it's based on an outdated architecture compared to GCN/Kepler/Maxwell. Snapdragon 800 and Tegra 4 were already faster here last year so you'd expect a 2014 AMD reference platform tablet with agressive turbo to be a lot faster. Intel recognizes this, and that's why Cherry Trail will increase the EU count by 4x on a new architecture.
There were no Mullins design wins at all @ APU 2014, now let's wait for Computex 2014. Anyway, not the proper thread for this kind of discussion.

Since asking isn't getting anywhere, we're down to infractions
-ViRGE

 

[witeken]

Well, smaller CPU cores with higher performance at a density disadvantage and still AMDs Puma is faster than Intel BayTrail. If that doesn't qualify for a better design then what is ??

How did you come to that conclusion ??
Even if that is true, the fact remains the same, AMDs design is better because they are using a lower density process.

I don't think I can answer your questions, seeing ViRGE's comments above.

 

[parvadomus]

Good job at handpicking the CB 11.5 ST benchmark.

Both of you are actually handpicking. That puma+ @2.2Ghz is obviously throttling a lot in CB MT , maybe even at lower Ghz than the atom.
I will not comment until I see a proper mullins benchmark with real power consumption data. But I'm sure future 14nm cats will be a lot more competitive in performance/watt.

 

[AtenRa]

Good job at handpicking the CB 11.5 ST benchmark.

The reason I picked the Single Thread benchmark is because currently in Tablets and Extremely Low Power Devices, ST performance still matters more. And Puma is also faster in every ST benchmark in AT's review.

AMD could have created a smaller Die with lower iGPU performance and still be faster than BayTrail.
Truly if you think of it, AMD is in another league and way ahead of Intel in the Mobile (Tablet) segment.
Intel will most probable be in front in 2015 due to 14nm but the pain for Intel will come in 2016 when AMD will be on the low power high density 14nm FinFet process. If they could outclass Intel with 28nm, just imaging what they will bring with 14nm FinFets.

Now let's take a look at MT performance:
Almost same score as old Atom Z3770.

Old ?? That is current BayTrail