Best possible scenario for post-Bulldozer AMD x86 CPUs?

[Fjodor2001]

Hi,

It seems like AMD has lots of new tech coming up in the not so distant future.

To summarize:

* New post-Bulldozer x86 architecture, as discussed here.
* Stacked DRAM / High Bandwidth Memory (HBM), as discussed here.
* Indirectly Samsung/GF 14 nm FinFET, as discussed here.
* Anything else I missed?

It's unknown when all that will materialize, but from what most people seem to believe it'll be around 2016 or so.

So I'm just thinking, if all of these new technologies are released at the same time by AMD (possibly paired with a 6/8 core APU), couldn't it mean a new come back for them in the x86 desktop arena?

Please let me know what you think.

 

[MongGrel]

I'll refrain.

 

[ShintaiDK]

Its not possible for AMD to make an x86/ARM core on 20nm and below without having higher transistor cost than 28nm, or an IC design cost they can never recoup. So that alone limits the options drasticly.

2 posts in? Really?

Shintai, you certainly have earned enough vacations over the years that you know what thread crapping is.

-ViRGE

 

[SAAA]

Best possible scenario?

I expect any APU with decent amount of TDP, cores and HBM to kick Intel back in IGPs after they press with Broadwell/Skylake GT3/GT4 variants, this is ATI we are talking about, they are GPU makers!

For pure core performance the story is different: unless the very optimistic opinions of new ARM cores on par with Haswell IPC come true they will be still slower, especially because of even lower clocks than Steamroller/Piledriver.
Larger core count and no more module drawbacks can help in this regard (using denser nodes), but that's still the best chance they have and not an optimal solution.

A big if is compute and any IGP-assisted performance, if more programs can use this new feature even Kaveri could be seen as a killer CPU but to date is almost none. Plus old programs wouldn't benefit at all...

 

[Maximilian]

It not killing the company would be a pretty decent scenario.

Dont think they can afford another massive failure, they already had Phenom I and Bulldozer.

 

[frozentundra123456]

Pointless speculation at this point. I also think the entire thread is asking for trouble and will turn ugly really fast.

 

[SOFTengCOMPelec]

Its not possible for AMD to make an x86/ARM core on 20nm and below without having higher transistor cost than 28nm, or an IC design cost they can never recoup. So that alone limits the options drasticly.

Maybe not with X86, but I thought that the Arm cores can be developed relatively cheaply, because they are taking the "Arm" design and polishing/improving it, and building it with a chip plant(s), that they don't have to pay for (i.e. they pay for what they use, not the entire chip plant, like Intel).

But your point about being able to recoup the costs (profit from it), could well be true, so I'm contradicting myself here.

I guess it depends on how many Arm's it can sell.

 

[ShintaiDK]

Maybe not with X86, but I thought that the Arm cores can be developed relatively cheaply, because they are taking the "Arm" design and polishing/improving it, and building it with a chip plant(s), that they don't have to pay for (i.e. they pay for what they use, not the entire chip plant, like Intel).

But your point about being able to recoup the costs (profit from it), could well be true, so I'm contradicting myself here.

I guess it depends on how many Arm's it can sell.

Its the same for ARM. It doesnt matter what it is. ARM, x86, GPU, modems etc.

Example from ARM themselves:

Note the 14/16nm even have 1M transistors less than 28 and 20nm. Else it would go even further down.

Its unlikely we even see anything but FPGAs and high performance CPUs for 10nm for example as the next step.

28nm looks like a node that will have one of, if not the longest lifespan of any node.

 

[SOFTengCOMPelec]

Its the same for ARM. It doesnt matter what it is. ARM, x86, GPU, modems etc.

Example from ARM themselves:

Note the 14/16nm even have 1M transistors less than 28 and 20nm. Else it would go even further down.

Its unlikely we even see anything but FPGAs and high performance CPUs for 10nm for example as the next step.

28nm looks like a node that will have one of, if not the longest lifespan of any node.

There have been some rumours/signs/announcements that even Intel have been struggling to get 14 nm (Broadwell) production up and running (i.e. good yields), which further confirms what you have just said.

The whole semi-conductor industry, may have to make a huge technical leap, in order to move forward, as regards, super powerful chips, in the future.
I.e. Invent some fundamentally new way of making chips.

We maybe are ok for the next 5 or 10 years, but after that, it is not clear what the path forward is going to be.

This puts AMDs future cpus into even more difficulty.

 

[monstercameron]

They dont need new/better products, what they need is a killer app. When I say killer app I mean major design win or major software win.

[day-dreaming]
If some large company like apple suddenly endorsed AMD apus in their mac products[lower asp from AMD parts with their massive margins seem like a win] and lets be honest most mac users wont be able to tell.
[/day-dreaming]

 

[SOFTengCOMPelec]

They dont need new/better products, what they need is a killer app. When I say killer app I mean major design win or major software win.

[day-dreaming]
If some large company like apple suddenly endorsed AMD apus in their mac products[lower asp from AMD parts with their massive margins seem like a win] and lets be honest most mac users wont be able to tell.
[/day-dreaming]

Maybe what they could/should do, is create a new series of server chips, with optionally 8/16/24/32/48/64/96 or even 128 Arm Cortex A57 cores (with possible improvements to the basic Arm Cortex A57, and maybe a decent SoC as well). But they may need to use alternative Arm core types (i.e. smaller core types), to keep the power dissipation, and chip sizes under control.

They would have a choice of either lowish clocked versions (for the very high number of cores) and/or extreme cooling solutions, a bit like their 220 Watt FXs (top of range), or something like what they use to do for the old Mainframe computers.

Possibly even have 2, 4, 6 8 or even 16 way, massive cored Arm chips, possibly without shared memory, to avoid massive high speed communications issues.

The Server market is suppose to be VERY profitable, so it could be a viable option.

 

[jpiniero]

Honestly I don't know why AMD is even bothering.

[day-dreaming]
If some large company like apple suddenly endorsed AMD apus in their mac products[lower asp from AMD parts with their massive margins seem like a win] and lets be honest most mac users wont be able to tell.
[/day-dreaming]

AMD almost got the 2011 Macbook Air CPU contract, but Apple dropped them when they couldn't guarantee they could deliver the product. Super embarrassing I know.

 

[Fjodor2001]

Its not possible for AMD to make an x86/ARM core on 20nm and below without having higher transistor cost than 28nm

But that should apply to all companies in the semiconductor industry, right? I.e. it's not different for AMD compared to e.g. Qualcomm, NVidia or Intel. Because the the wafer price for different nodes is not set specifically for AMD.

 

[ShintaiDK]

But that should apply to all companies in the semiconductor industry, right? I.e. it's not different for AMD compared to e.g. Qualcomm, NVidia or Intel. Because the the wafer price for different nodes is not set specifically for AMD.

Thats correct. Hence there is only 3, possible 4 companies that can get lower transitors cost at 20nm and down. Same reason why its almost a death game for who gets the 14/16nm FF customer(s) and who doesnt.

 

[NTMBK]

Congratulations, you've successfully FUDed this thread into the ground. Meanwhile in the real world AMD has already committed to shipping 20nm parts next year.


If you have a problem with another poster, you need to report it. These kinds of callouts only make things worse
-ViRGE

 

[Fjodor2001]

Thats correct. Hence there is only 3, possible 4 companies that can get lower transitors cost at 20nm and down.

I'm not sure I follow. Why would only 3 or 4 companies get lower transistor cost at 20 nm and down? Are you suggesting TSMC/GF/Samsung will set one wafer price for 3-4 customers, and another higher price for the rest? Why would they do that?

Same reason why its almost a death game for who gets the 14/16nm FF customer(s) and who doesnt.

I agree with NTMBK - that's FUD and pure speculation. Are you suggesting either TSMC or Samsung/GF will stop developing and producing wafers at below 20 nm for certain? I.e. there will only be one foundry left in the world after that using the latest node process tech? In that case I'd be really worried about Intel, since that only foundry that is left will have such huge volumes and hence process tech R&D capacity that not even Intel can match it.

Regardless, it's another story since AMD can move the wafer production to those foundries that are left anyway if needed.

 

[NTMBK]

I'm not sure I follow. Why would only 3 or 4 companies get lower transistor cost at 20 nm and down? Are you suggesting TSMC/GF/Samsung will set one wafer price for 3-4 customers, and another higher price for the rest? Why would they do that?

He's talking about design costs. His current theory is that double patterning will make designs too expensive for AMD to survive.

 

[ShintaiDK]

Congratulations, you've successfully FUDed this thread into the ground. Meanwhile in the real world AMD has already committed to shipping 20nm parts next year.

They can still ship 20nm, they just wont get any lower transistor cost. Meaning the transistor amount isnt increasing without a price increase. And even with the same transistor count, price will increase too.

Or are you saying its FUD despite companies like ARM, Broadcom, nVidia and IBS supports me?

Please show me your edvidence of me being wrong.

 

[Fjodor2001]

He's talking about design costs. His current theory is that double patterning will make designs too expensive for AMD to survive.

How does that theory hold, if AMD has already committed to delivering 20 nm products next year? If it would have been too expensive for them, they should have known that by now.

Also, is it really reasonable that design costs should be that much higher at 20 nm and below anyway? I.e. that design costs will amount to such a huge part of the total R&D costs that only 3 or 4 semiconductor companies in the world can design chips for it. Are there any public cost figures or articles to back that theory up?

 

[ShintaiDK]

I'm not sure I follow. Why would only 3 or 4 companies get lower transistor cost at 20 nm and down? Are you suggesting TSMC/GF/Samsung will set one wafer price for 3-4 customers, and another higher price for the rest? Why would they do that?

Its about the IC design cost.

There is 2 solutions. Either you pay a huge premium in IC design cost (Around 1B$.) or you dont get cheaper transistor cost.

It shouldnt be news for you, since you also pacitipated in this thread:
http://forums.anandtech.com/showthread.php?t=2378755&page=1

Pretty well summoned up here:
http://forums.anandtech.com/showpost.php?p=36560660&postcount=264

 

[Fjodor2001]

They can still ship 20nm, they just wont get any lower transistor cost. meaning the transistor amount isnt increasing.

Or are you saying its FUD despite companies like ARM, Broadcom, nVidia and IBS supports me?

Please show me your edvidence of me being wrong.

It was your claim, so the burden of evidence is upon you.

So do you have any links to public cost figures and articles to back that theory up? I.e. that design costs at 20nm and below will amount to such a huge part of the total R&D costs that only 3 or 4 semiconductor companies in the world can design chips for it in such a way that they get lower transistor cost.

 

[ShintaiDK]

Even Samsung says the same.

 

[NTMBK]

Getting back on topic, one of the most interesting possibilities is pin compatible ARM and x86 SoCs. This will let OEMs offer Windows and Android devices with the same design, varying only a single component. This will hopefully reduce both design and inventory costs.

 

[witeken]

Android runs fine on x86.

 

[Idontcare]

Even Samsung says the same.

Yep, its true.

Per the thread's title question, the best possible post-bulldozer scenario is that Intel elects to become a division of UNICEF and thereby voluntarily chooses to manufacture AMD's future processors on a pro-bono basis. You know, for the betterment of mankind and all that.

Because right now AMD doesn't have too many other options when it comes to improving on price/performance. Only Intel's management and decision makers had the wisdom and fortitude year's ago to see to it that they were smartly investing in the R&D that they would need some 7-10 yrs later (now).

And for that they should be punished, nobody likes a smart kid who has planned smartly for his future. Bring in the FTC and bust them up!