AMD CPU speculation... and expert conjecture

[The Q6660 Inside]

Are steamrollers just APUs ?

Will they release a pure CPU ?

Yes and no, depends on how you define APU and CPU.

AMD is investing heavily into mixed processor's, otherwise known as HSA. They're current design's have a huge a$$ iGPU that doubles as a vector coprocessor bolted onto the side of every chip. That's the APU btw. Since all they need is the giant vector coprocessor it's theoretically possible to build a 3~4 module chip with a scaled down iGPU that only functions for HSA work (GPU components being stripped off). In that case it wouldn't be an "APU" anymore as it couldn't do graphics but it would still have the vector processing capabilities.

Also I caution everyone to take all "HSA" benchmark projections with a very large grain of salt. The vast majority of code is scalar, meaning HSA does jack sh!t for it. The workloads that can be done in vector or massively parallel (GCN can do regular integer operations) will get a very large boost from HSA. Any benchmark that only tests the parallel will dramatically inflate the worth of HSA.

Hmm, that is an quite interesting assumption..

 

[Cazalan]

Since all they need is the giant vector coprocessor it's theoretically possible to build a 3~4 module chip with a scaled down iGPU that only functions for HSA work (GPU components being stripped off). In that case it wouldn't be an "APU" anymore as it couldn't do graphics but it would still have the vector processing capabilities.

How much die savings would they get though? If AMD wanted a bigger piece of the HPC market they should do that.

 

[palladin9479]

Since all they need is the giant vector coprocessor it's theoretically possible to build a 3~4 module chip with a scaled down iGPU that only functions for HSA work (GPU components being stripped off). In that case it wouldn't be an "APU" anymore as it couldn't do graphics but it would still have the vector processing capabilities.

How much die savings would they get though? If AMD wanted a bigger piece of the HPC market they should do that.

Some savings by not including the texture units and some of the prefetch stuff. GCN is an interesting design, each CU is essentially a weak 64-core vector CPU that can also do scalar instructions. You could theoretically run an entire OS off a GCN processor, if you could find some way to socket it on a MB with some form of BIOS boot function. In a HSA environment the scalar components aren't needed as your gonna use the CPU for that anyway.

The big decision comes in how many CU's to employ. Each CU is a vector processor with 64 vector units (believe each is 32-bit's wide but not sure), that is one helluva SIMD processor. A CU has pathetic scalar performance and outstanding, phenomenal vector performance. It's also the component that takes up the most room on the iGPU side. You can reduce the number of CU's to fit in another scalar module. That reduces your vector performance while enhancing your scalar performance. Desktop world needs more scalar then vector while HPC world often purchase's dedicated vector processors (single dGPU has 32CU's for comparison).

It's a big balancing act, I can see a 3M 8CU design being made.

 

[esrever]

Nah, GF are totally brilliant 🙂

http://www.youtube.com/watch?v=eX9HP_rzQc8&feature=youtu.be

video doesn't tell me anything about why globalfounderies is stuck on 28nm for so long.

 

[juanrga]

Good to see AMD correcting the misunderstandings in this thread

“Sometimes there’s a misconception that ARM means very low power and very low performance,” says Kamal Khouri, director of product management for AMD’s embedded division, ” but this is truly a high performance SOC taking advantage of ARM’s first 64-bit architecture."

I estimate HieroFalcon would perform as the top Kaveri Steamroller.

AMD continues without confirming/denying the 3 ALUs per core diagram.

 

[juanrga]

Are steamrollers just APUs ?

Will they release a pure CPU ?

Berlin has been presented as both APU and CPU, therefore I would expect the same for Steamroller in the desktop: Kaveri (APU) and Athlon (CPU)

 

[The Q6660 Inside]

Are steamrollers just APUs ?

Will they release a pure CPU ?

Berlin has been presented as both APU and CPU, therefore I would expect the same for Steamroller in the desktop: Kaveri (APU) and Athlon (CPU)

He is likely reffering to FX...

 

[Cazalan]

Here's one tiny FM2+ board for you.

http://www.fudzilla.com/home/item/32878-gigabyte-has-a-killer-fm2%20-itx-board

 

[juanrga]

Are steamrollers just APUs ?

Will they release a pure CPU ?

Berlin has been presented as both APU and CPU, therefore I would expect the same for Steamroller in the desktop: Kaveri (APU) and Athlon (CPU)

He is likely reffering to FX...

I think there is no FX Steamroller.

 

[Cazalan]

Good to see AMD correcting the misunderstandings in this thread

“Sometimes there’s a misconception that ARM means very low power and very low performance,” says Kamal Khouri, director of product management for AMD’s embedded division, ” but this is truly a high performance SOC taking advantage of ARM’s first 64-bit architecture."

I estimate HieroFalcon would perform as the top Kaveri Steamroller.

AMD continues without confirming/denying the 3 ALUs per core diagram.

SR code has been posted for some time now. (BDVER3.MD) There are 2 ALUs.

http://gcc.gnu.org/ml/gcc-patches/2012-10/msg01079/bdver3.md

;; The bdver3 contains three pipelined FP units and two integer units.
;; Fetching and decoding logic is different from previous fam15 processors.
;; Fetching is done every two cycles rather than every cycle and
;; two decode units are available. The decode units therefore decode
;; four instructions in two cycles.


HieroFalcon is such a different product than Kaveri. It is CPU only and no graphics. They're not very comparable.

 

[8350rocks]

Are steamrollers just APUs ?

Will they release a pure CPU ?

Berlin has been presented as both APU and CPU, therefore I would expect the same for Steamroller in the desktop: Kaveri (APU) and Athlon (CPU)

He is likely reffering to FX...

I think there is no FX Steamroller.

It may not be called FX specifically, though I have assurances that there will be a suitable replacement. It may be on FM2+, there may be an entirely new socket for it...I haven't been able to pry out any further information yet. (The rep has been extremely resourceful about dodging questions in such a manner as to address them, but ever so vaguely as to give nothing concrete...)

 

[Ags1]

Given that AMD sales of HEDT went up this quarter, and sales of APU-based mobile devices went down, I would be pretty surprised if they didn't produce something for the HEDT crowd with Steamroller.

Make that rep squirm! He or she gonna break eventually 🙂

 

[blackkstar]

http://www.globalfoundries.com/technology/14XM-FAQ.aspx

GloFo is already claiming that tape-outs should be happening in 2013.

Also, yes, the xbit article is high on speculation. But they are also speculating that 14nm will only be used for mobile parts while GloFo is claiming it can scale to multi-core and advanced servers.

There is a lot of contradicting information and whatever side you pick probably depends on if you're more optimistic or pessimistic about AMD. I'm personally more optimistic but that's just me.

I have no doubt AMD is playing the best possible hand they can with the cards they were dealt.

That GF page says they were expecting deliveries of 20nm in 2H2012. We're a year beyond that and I don't think any 20nm part from GF has hit the market yet. Kaveri will likely be the first 28nm part.

AMD has sizable wafer commitments to GF and they're currently all 32nm. Something just doesn't add up there. Now GF is talking about 14nm. That's 3 deep in the pipeline of vaporware (28nm/20nm/14nm). Maybe I missed some press releases but has anyone found any chips GF made with these nodes?

My point is that GloFO can not keep having vaporware nodes. They have to be cancelled for a reason and I'm suggesting that the reason is that they are just moving faster and cancelling what they were working on.

GloFo is backed by ATIC and a whole lot of oil money. I really wouldn't be surprised if the folks in charge of GloFo had no qualms about just tossing money away.

Also, the Extreme in Extreme Mobility can either mean it's

1. Made for extremely mobile products
2. is designed for mobile products but can scale extremely well to higher end products.

I have not seen proof towards either one, however I am at least pointing out that there are alternatives to the suggestions provided about 14nm XM. For all we know GloFo got lucky and has a 14nm/20nm hybrid process that sips power and scales better than everyone else. Or worst case it is another vaporware product.

Also, new roadmaps from AMD:

http://www.slideshare.net/wilfredlin/amdembeddedroadmapunveilvfinal-130906181555

Of note (copy pasted from my OCN account of a different name)


1. AMD is not telling us what 60% to 50% of their future business plans are (slide 3).
2. ARM is not replacing Steamroller APUs (slide 9)
3. ARM is not replacing Jaguar (slide 9)
4. AMD is not even getting close to replacing Jaguar with ARM as the only ARM part on slide 9 is an (up to) 8 core A57 APU

Thus, I think it's safe to infer that:

1. AMD's products are either mobile/embedded or classic HEDT/server. If AMD is moving 40% to 50% of their product lineup to mobile/embedded high growth markets, it means that the remainder can ONLY go to HEDT/server as that's the only other things AMD makes

2. AMD has a Steamroller v2 out there and it has better HSA features according to these slides.

3. AMD has little faith that ARM will cause a massive disruption in x86 but instead will co-exist.

AMD's own market projections have ARM taking about .5% from x86 and .5% from MIPS by 2016. That is hardly disruptive at all.

I find it curious that there are no low power ARM chips with GCN cores. Perhaps AMD is looking at Tegra and realizing doing something similar is a giant waste of time and resources. Regardless it means Temash is here to stay.

Surface 2 Pro reviews popped up today with the new Haswells.

http://www.anandtech.com/show/7440/microsoft-surface-pro-2-review/4

~50% increase in battery life and it still is blown away by ARM. And this is what Intel is optimizing their big cores for.

Now do you see why I think it's absolutely foolish for AMD to walk away from the HEDT/Gaming CPU market? They can beat Intel with Temash and then beat Intel in HEDT because Intel is off trying to compete with efficient, simple RISC ARM designs with a bloated x86 CPU.

 

[Cazalan]

This is a bit of a surprise if true. Xbox One being able to run Windows 8 applications.

http://www.brightsideofnews.com/news/2013/10/21/dell-confirms-xbox-one-will-run-windows-8-apps.aspx

 

[jdwii]

http://www.3dnews.ru/assets/external/illustrations/2013/09/23/762494/sm.sandra.600.png

 

[The Q6660 Inside]

http://www.3dnews.ru/assets/external/illustrations/2013/09/23/762494/sm.sandra.600.png

Well, looks like we found the A10-X800K for Kaveri. Because of the DDR3 bottleneck, I would assume its iGPU has about 7770GHz performance.

 

[Kulasko]

Look at the iGPU frequency and take the DDR3 RAM into account... I think the performance will be around a 7750, maybe less... still very impressive.

832sp would make a hell of an iGPU.

 

[Cazalan]

I find it curious that there are no low power ARM chips with GCN cores. Perhaps AMD is looking at Tegra and realizing doing something similar is a giant waste of time and resources. Regardless it means Temash is here to stay.

Good find of that slide, hadn't seen the updated pics for the GPU+RAM before. It seems to correspond with what others are saying, that the ARM parts are tailored more as network processors than a replacement for Kabini/Temash.

 

[de5_Roy]

AMD makes GPU comute a reality with hQ
Heterogeneous Queues make GPUs a full peer in compute
http://semiaccurate.com/2013/10/21/amd-makes-gpu-comute-reality-hq/
http://techreport.com/news/25545/amd-heterogeneous-queuing-aims-to-make-cpu-gpu-more-equal-partners
looks good for kaveri
http://www.tomshardware.com/news/amd-apu-performance-increase-drivers-windows-8.1,24773.html

I find it curious that there are no low power ARM chips with GCN cores. Perhaps AMD is looking at Tegra and realizing doing something similar is a giant waste of time and resources. Regardless it means Temash is here to stay.

Good find of that slide, hadn't seen the updated pics for the GPU+RAM before. It seems to correspond with what others are saying, that the ARM parts are tailored more as network processors than a replacement for Kabini/Temash.

yeah. so far none of amd's arm socs have gcn attached. amd associates the 8 core arm soc with a57 cores as high performance... yet the soc has vanilla arm cores and looks.... generic.

the slideshow was about amd's embedded cpus/apus/socs but it points to amd's general direction and it's not HEDT. amd seems to assume that arm and x86 will have near-equal overall marketshare in 2016 and it wants a piece of both. good call, highly dependent on timely execution and consumers - as current arm devices are highly dependent on consumers. to have that much marketshare, arm cannot slow down. we'll get a clearer idea when amd seperates it's semi-custom dept. numbers from gfx numbers.

as for amd's HEDT sales going up: amd said fx cpus sold more, never said which ones. so far, 4 core fx cpus (e.g. fx4300) have far outsold 6 and 8 core ones and those are much cheaper too. quad core fx is part of amd's hedt lineup by brand. and amd's hedt is very different from intel's hedt, don't confuse the two. richland seems to have underwhelmed in laptops - that was expected. amd can't expect to screw with customers for long.

 

[Ags1]

so far, 4 core fx cpus (e.g. fx4300) have far outsold 6 and 8 core ones and those are much cheaper too

Do you have a reference for that? I can't imagine why anyone would buy FX4300 when FX6300 is almost the same price. Not sure what you mean by 'those are much cheaper too'- where I live only 40 euros separate the FX4300 and the FX8320.

 

[de5_Roy]

Do you have a reference for that? I can't imagine why anyone would buy FX4300 when FX6300 is almost the same price. Not sure what you mean by 'those are much cheaper too'- where I live only 40 euros separate the FX4300 and the FX8320.

prices are different in different places.

the latest reference i found was this:
http://www.techpowerup.com/175751/apus-make-up-nearly-75-of-amds-processor-sales.html
the link has a pie chart showing core numbers. i think 2013 numbers will come out later this year or in 2014.

edit: this year could be different if new games and softwares, consoles and price drops pursuade users to buy 6-8 core fx cpus. all the info i've read so far don't indicate if the situation in the tpu link has changed.

 

[juanrga]

Good to see AMD correcting the misunderstandings in this thread

“Sometimes there’s a misconception that ARM means very low power and very low performance,” says Kamal Khouri, director of product management for AMD’s embedded division, ” but this is truly a high performance SOC taking advantage of ARM’s first 64-bit architecture."

I estimate HieroFalcon would perform as the top Kaveri Steamroller.

AMD continues without confirming/denying the 3 ALUs per core diagram.

SR code has been posted for some time now. (BDVER3.MD) There are 2 ALUs.

http://gcc.gnu.org/ml/gcc-patches/2012-10/msg01079/bdver3.md

;; The bdver3 contains three pipelined FP units and two integer units.
;; Fetching and decoding logic is different from previous fam15 processors.
;; Fetching is done every two cycles rather than every cycle and
;; two decode units are available. The decode units therefore decode
;; four instructions in two cycles.


HieroFalcon is such a different product than Kaveri. It is CPU only and no graphics. They're not very comparable.

Each Steamroller _module_ contains three FPU units and two integer units. But I am discussing about something different. I am asking AMD if each _integer unit_ has 2 or 3 ALUs. Hot Chips 2012 showed two integer units with 2 ALUS each (4 ALUs per module). Sep Kaveri talk showed two integer units with 3 ALUS each (6 ALUs per module).

When I wrote "Kaveri Steamroller" I did really mean only the CPU part of Kaveri. I was comparing CPU to CPU.

I think there is no FX Steamroller.

It may not be called FX specifically, though I have assurances that there will be a suitable replacement. It may be on FM2+, there may be an entirely new socket for it...I haven't been able to pry out any further information yet. (The rep has been extremely resourceful about dodging questions in such a manner as to address them, but ever so vaguely as to give nothing concrete...)

My point was not about branding name.

Also, new roadmaps from AMD:

http://www.slideshare.net/wilfredlin/amdembeddedroadmapunveilvfinal-130906181555

Of note (copy pasted from my OCN account of a different name)


1. AMD is not telling us what 60% to 50% of their future business plans are (slide 3).
2. ARM is not replacing Steamroller APUs (slide 9)
3. ARM is not replacing Jaguar (slide 9)
4. AMD is not even getting close to replacing Jaguar with ARM as the only ARM part on slide 9 is an (up to) 8 core A57 APU

You mention slide 3 and 9 but I think you mean 4 and 10. AMD plans are about moving outside the PC up to a point where most of the revenue comes from non-PC market.

Steamroller APUs will be replaced by Carrizo APUs.

The replacement of x86 by ARM will be made in phases. AMD is _first_ replacing jaguar in server and tablets _then_ other products will follow.

AMD's own market projections have ARM taking about .5% from x86 and .5% from MIPS by 2016. That is hardly disruptive at all.

With new head, AMD projections are overly-conservative. Their Q3 results are much better than predicted, for instance. Also relevant this AMD slide:

The slide is pretty clear as are also AMD statements about Warsaw and their future server plans.

I find it curious that there are no low power ARM chips with GCN cores. Perhaps AMD is looking at Tegra and realizing doing something similar is a giant waste of time and resources. Regardless it means Temash is here to stay.

As mentioned before, AMD will be replacing Temash tablets by ARM+GCN tablets this or next year

http://techreport.com/news/25461/report-amd-to-introduce-arm-based-tablet-chip-this-year

Finally I want to emphasize again that AMD is already devoting time and resources to a custom high-performance ARM chip as Nvidia does. The problem with AMD was the myopia of the _former_ head. AMD is too late to the party, that is why Seattle and HieroFalcon use standard A57 cores. AMD will compete with standard cores before has ready its own cores.

 

[juanrga]

This is a bit of a surprise if true. Xbox One being able to run Windows 8 applications.

http://www.brightsideofnews.com/news/2013/10/21/dell-confirms-xbox-one-will-run-windows-8-apps.aspx

No surprise here. It has been know since day one Xbox1 has three OSs, one for games, another for w8 apps and a third one that manages the other two

http://www.engadget.com/2013/05/21/xbox-one-runs-three-operating-systems/

You can have a game on the one part of the screen and then a W8 app in other side. Something looking as this

Not all W8 apps will work on Xbox1.

Do you have a reference for that? I can't imagine why anyone would buy FX4300 when FX6300 is almost the same price. Not sure what you mean by 'those are much cheaper too'- where I live only 40 euros separate the FX4300 and the FX8320.

prices are different in different places.

the latest reference i found was this:
http://www.techpowerup.com/175751/apus-make-up-nearly-75-of-amds-processor-sales.html
the link has a pie chart showing core numbers. i think 2013 numbers will come out later this year or in 2014.

edit: this year could be different if new games and softwares, consoles and price drops pursuade users to buy 6-8 core fx cpus. all the info i've read so far don't indicate if the situation in the tpu link has changed.

This is the chart

That 2012 data shows that dual and quad-cores make for >90% of the market, whereas octo-cores make for 1%. That data coincides very well with last Steam stats, except that last Steam statistics give 0.3% for octo-cores.

Those numbers are the reason why Kaveri comes as dual and quad-core and there is no FX octo-core Steamroller.

 

[de5_Roy]

http://www.techpowerup.com/img/12-11-19/146a.jpg

2012 data and coincides very well with last Steam stats. Dual and quad-cores make for >90% of the market, whereas octo-cores make for 1%. last Steam statistics give 0.3% for octo-cores.

Those numbers are the reason why Kaveri comes as dual and quad-core and there is no FX octo-core Steamroller.

those numbers are not the reason kaveri is dual module. apus are not aimed at high-perf. high-end. they're chiefly for laptops and entry level desktops. that's why they incorporate cost-cutting measures like lack of L3 cache, less number of modules, dual channel memory controller, etc. price point, die-area, fab. process and kaveri's architecture are the reasons kaveri is dual module.

 

[The Q6660 Inside]

http://www.techpowerup.com/img/12-11-19/146a.jpg

2012 data and coincides very well with last Steam stats. Dual and quad-cores make for >90% of the market, whereas octo-cores make for 1%. last Steam statistics give 0.3% for octo-cores.

Those numbers are the reason why Kaveri comes as dual and quad-core and there is no FX octo-core Steamroller.

those numbers are not the reason kaveri is dual module. apus are not aimed at high-perf. high-end. they're chiefly for laptops and entry level desktops. that's why they incorporate cost-cutting measures like lack of L3 cache, less number of modules, dual channel memory controller, etc. price point, die-area, fab. process and kaveri's architecture are the reasons kaveri is dual module.

+1