AMD CPU speculation... and expert conjecture

[Cazalan]

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).

It's all in the code there. A module has 2 bdver3 cores. The FP units are shared of course but each core can access them, just they might have to wait.

These are your 2 ALUs.

(define_cpu_unit "bdver3-ieu0" "bdver3_ieu")
(define_cpu_unit "bdver3-ieu1" "bdver3_ieu")
(define_reservation "bdver3-ieu" "(bdver3-ieu0|bdver3-ieu1)")


These are your 2 AGUs.

(define_cpu_unit "bdver3-agu0" "bdver3_agu")
(define_cpu_unit "bdver3-agu1" "bdver3_agu")
(define_reservation "bdver3-agu" "(bdver3-agu0|bdver3-agu1)")

Which makes up the typical 4 integer pipelines you see in the block diagrams. Same as Bulldozer.

Further down you can see how many clock cycles each operation takes. That's where you can do a more fine comparison between architecture changes.

 

[juanrga]

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.

My argument had two parts: one about Kaveri, another about FX. It was not only about APUs...

If market share for dual-cores was of a 0.1% I can sure you that Kaveri 2C wouldn't exit. The dual-core Kaveri exists because dual-cores are popular for both APUs and CPUs.

According to the above data:
- dual cores are about one half the market
- quad cores are about one third the market

AMD could release an octo-core Kaveri APU with quad memory channel and 1024SP tomorrow, just how they could release a hypothetical FX-8570 CPU tomorrow.

They aren't doing it because the market is close to non-existent. Nobody buys a 28nm production line for making APUs/CPUs for ~0.3% of the whole market, except if one is going to sell them at $1200 each, which is not the case.

Nvidia is preparing ultra-high-performance APUs with 8-cores + 1024 SP. I expect an enthusiast line of APUs from AMD, something similar to this

but no today neither tomorrow...

 

[de5_Roy]

My argument had two parts: one about Kaveri, another about FX. It was not only about APUs...

If market share for dual-cores was of a 0.1% I can sure you that Kaveri 2C wouldn't exit. The dual-core Kaveri exists because dual-cores are popular for both APUs and CPUs.

i didn't address the fx bit because i have too little information on that, and i didn't find it worth addressing.

it's not about marketshare. your argument(!) has a gigantic flaw: 2c kaveri (i.e. single module cpu) exists/will exist because of die harvesting and amd making money off partially defective kaveri silicon instead of throwing those away after looking at some marketshare pie chart.

According to the above data:
- dual cores are about one half the market
- quad cores are about one third the market

AMD could release an octo-core Kaveri APU with quad memory channel and 1024SP tomorrow, just how they could release a hypothetical FX-8570 CPU tomorrow.

They aren't doing it because the market is close to non-existent. Nobody buys a 28nm production line for making APUs/CPUs for ~0.3% of the whole market, except if one is going to sell them at $1200 each, which is not the case.

they're not doing that because apus are not aimed high end and becaue it'll take up a lot of space, affecting yields. it'll also use a lot of power and be hard to cool. glofo does not have good track record for new nodes as early llano and zambezi supplies showed. kaveri is dual module by design.
i could see amd launching a 6 core 'big' kaveri but i dunno if amd would start off with such an asymmetric cpu core design ( odd number of modules).

 

[blackkstar]

Did some of you even look at the slides I posted?
http://www.slideshare.net/wilfredlin/amdembeddedroadmapunveilvfinal-130906181555

1. AMD only expects embedded/mobile market to account for 40% to 50% of their sales in the next couple of years. THATS LESS THAN HALF OF THEIR PRODUCTS

2. The roadmap CLEARLY SHOWS Temash is not going to be replaced by anything ARM. There are no ARM + GCN low power products on the roadmap at all for 2014.

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

Sweclockers doesn't offer more specifics on the chip (the small SoC for mobile devices), which will reportedly be announced in Q4 and launch some time next year.

This completely contradicts the official slide AMD released in the slides I linked where there is no small ARM with GCN, only one for servers and Temash going on.

I do not understand why you are taking one source (SweClockers) over another source (AMD) about what AMD is planning on doing. I am taking the newer, official AMD embedded/mobile roadmaps over a rumor site's opinion.

 

[8350rocks]

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.

My argument had two parts: one about Kaveri, another about FX. It was not only about APUs...

If market share for dual-cores was of a 0.1% I can sure you that Kaveri 2C wouldn't exit. The dual-core Kaveri exists because dual-cores are popular for both APUs and CPUs.

According to the above data:
- dual cores are about one half the market
- quad cores are about one third the market

AMD could release an octo-core Kaveri APU with quad memory channel and 1024SP tomorrow, just how they could release a hypothetical FX-8570 CPU tomorrow.

They aren't doing it because the market is close to non-existent. Nobody buys a 28nm production line for making APUs/CPUs for ~0.3% of the whole market, except if one is going to sell them at $1200 each, which is not the case.

Nvidia is preparing ultra-high-performance APUs with 8-cores + 1024 SP. I expect an enthusiast line of APUs from AMD, something similar to this

but no today neither tomorrow...

When 2 core CPUs came out, there were FAR more single core CPUs than dual core CPUs. But they continued making dual core CPUs in spite of low market share initially. Now, they no longer make single core CPUs for mainstream consumer PCs, even though they were once 100% of the market.

What's your point?

 

[tracker45]

Okay. How will steamroller compare to the 8320, and 8350.

 

[blackkstar]

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.

My argument had two parts: one about Kaveri, another about FX. It was not only about APUs...

If market share for dual-cores was of a 0.1% I can sure you that Kaveri 2C wouldn't exit. The dual-core Kaveri exists because dual-cores are popular for both APUs and CPUs.

According to the above data:
- dual cores are about one half the market
- quad cores are about one third the market

AMD could release an octo-core Kaveri APU with quad memory channel and 1024SP tomorrow, just how they could release a hypothetical FX-8570 CPU tomorrow.

They aren't doing it because the market is close to non-existent. Nobody buys a 28nm production line for making APUs/CPUs for ~0.3% of the whole market, except if one is going to sell them at $1200 each, which is not the case.

Nvidia is preparing ultra-high-performance APUs with 8-cores + 1024 SP. I expect an enthusiast line of APUs from AMD, something similar to this

but no today neither tomorrow...

What is the point of Nvidia doing this? So they can make ARM APUs that play Android games super fast?

 

[-Fran-]

What is the point of Nvidia doing this? So they can make ARM APUs that play Android games super fast?

I would say the mobile gaming market is pretty huge.

Cheers!

 

[8350rocks]

IIRC Nvidia DENVER arm is to be for cluster servers

I would imagine the format of that would be closer to a vector coprocessor for the iGPU then.

Either way, DENVER from NVidia is going to be about as successful as Tegra, Tegra 2, Tegra 3, and Tegra 4 before it...

Which is to say...not very successful...

http://hexus.net/business/news/components/54085-nvidia-tegra-sales-hit-qualcomms-lte-equipped-chips/

Qualcomm is kicking them around...

 

[8350rocks]

I would not be so sure about denver falling flat. It was originally to be used in both desktop and server applications using transmeta code morphing software to run 32 , 64bit windows and android dual boot. The legal team advised nvida against trying to enforce it's x86 license acquired from transmeta. So denver became a server product. There is a demand for these low power server clusters, The market is just not there yet. Based on nvidias size and channel partners Denver will win some designs,

Somehow, I doubt NVidia wins too many...they have some partners and channels for that, but then why hasn't Tegra 2 or Tegra 3 or Tegra 4 or Tegra 4i taken off like a rocket?

Additionally, AMD has SeaMicro to pedal it's commercial micro server clusters...

EDIT: As for transmeta's software, well...their designs were terrible at serial computing...parallel...yeah...they could do parallel. However, when it came to running serial compute functions where x86 really shines...the difference was massive between "real" x86 and "abstracted" x86.

 

[jdwii]

Off topic but man i can't wait for GTA5 for the PC
http://www.incgamers.com/2013/10/more-gta-5-for-pc-evidence-surfaces-in-amd-catalyst-drivers

More evidence that its coming sooner than later

 

[8350rocks]

Off topic but man i can't wait for GTA5 for the PC
http://www.incgamers.com/2013/10/more-gta-5-for-pc-evidence-surfaces-in-amd-catalyst-drivers

More evidence that its coming sooner than later

If they put MANTLE in that game...look out...it could be something crazy good.

 

[8350rocks]

Good question, maybe because to get into the thermal envelop of the target product it had to be clocked way down hurting performance.

It is not hard to imagine nvidia using cash and strong arm tactics (carrot and Stick) to force there product on the market, this has been a dream for them.

Seamicro was a great buy for amd, but pushing atom micro servers is not going to fair well in a sea of arm nodes.

Edit
Regarding transmeta. The crusoe 400 was what caused all the bad press. I had the sonic blue pro book and it was not as bad as the reviews made it out to be. the 800 was ok but the 1ghz model in the t1000 was indistinguishable from the t1100 pentium model in normal use. It was around that time intel sent the lawyers lose on transmeta basically suing them out of business. Code morphing was good and getting better,

Well, you do realize that all of that code morphing, and those patents, are what is the driving foundation for GPU compute these days...right? GCN can actually do serial computing if you wanted, though serial programs would feel lethargic compared to the CPUs we have now if you wanted to try to do it.

 

[tracker45]

Is steamroller is only 15% faster than current AMD CPUs, is there even any point of releasing a new range of CPUs ?

They could fail like bulldozer did.

 

[8350rocks]

Is steamroller is only 15% faster than current AMD CPUs, is there even any point of releasing a new range of CPUs ?

They could fail like bulldozer did.

Well let's look at Intel improvements recently, shall we?

Nehalem to Sandy Bridge: Sandy Bridge was 15% better
Sandy Bridge to Ivy Bridge: Ivy was about 8% better
Ivy Bridge to Haswell: Haswell is about 5% better

So, to recap...Haswell is only a 13% improvement over Sandy Bridge...so, why did Intel bother releasing those?

Steamroller is projected to be a 20-30% increase over Piledriver...

15% doesn't sound like a lot, but in the CPU world, 15% is a big improvement. Intel hasn't improved 15% generation over generation in 4 years, and since then they haven't improved 15% over that generation(Sandy Bridge).

 

[blackkstar]

Is steamroller is only 15% faster than current AMD CPUs, is there even any point of releasing a new range of CPUs ?

They could fail like bulldozer did.

Well let's look at Intel improvements recently, shall we?

Nehalem to Sandy Bridge: Sandy Bridge was 15% better
Sandy Bridge to Ivy Bridge: Ivy was about 8% better
Ivy Bridge to Haswell: Haswell is about 5% better

So, to recap...Haswell is only a 13% improvement over Sandy Bridge...so, why did Intel bother releasing those?

Steamroller is projected to be a 20-30% increase over Piledriver...

15% doesn't sound like a lot, but in the CPU world, 15% is a big improvement. Intel hasn't improved 15% generation over generation in 4 years, and since then they haven't improved 15% over that generation(Sandy Bridge).

Hate to be that guy, but that's not how percentages work.

Say SB is performance of:
performance = 1.0

IB is an 8% increase according to you, so:
performance = 1.0 * 1.08 = 1.08

Haslel is a 5% increase according to you, so:
performance = IB performance * haswell increase
performance = 1.08 * 1.05 = 1.134

You claimed it was 13% faster, so you'd have
performance = 1.13

It isn't a huge difference in percentage this time but it can be if you continue to do math improperly in threads.

Conversely, if we give bdver1 an arbitrary performance index of 1.0

bdver1 -> bdver2 = 15% increase overall
1.15
bdver2 -> bdver3 = 20% increase overall
1..38

bdver3 being 20% faster than bdver2 would mean bdver3 is 38% faster than bdver1

However if bdver3 is only 30% faster than bdver1 it would have to mathematically be a smaller than 15% increase in performance over bdver2.

Sorry, maybe I am still resentful of all the derps in my CS program who blew at math yet did alright in the class.

 

[8350rocks]

Is steamroller is only 15% faster than current AMD CPUs, is there even any point of releasing a new range of CPUs ?

They could fail like bulldozer did.

Well let's look at Intel improvements recently, shall we?

Nehalem to Sandy Bridge: Sandy Bridge was 15% better
Sandy Bridge to Ivy Bridge: Ivy was about 8% better
Ivy Bridge to Haswell: Haswell is about 5% better

So, to recap...Haswell is only a 13% improvement over Sandy Bridge...so, why did Intel bother releasing those?

Steamroller is projected to be a 20-30% increase over Piledriver...

15% doesn't sound like a lot, but in the CPU world, 15% is a big improvement. Intel hasn't improved 15% generation over generation in 4 years, and since then they haven't improved 15% over that generation(Sandy Bridge).

Hate to be that guy, but that's not how percentages work.

Say SB is performance of:
performance = 1.0

IB is an 8% increase according to you, so:
performance = 1.0 * 1.08 = 1.08

Haslel is a 5% increase according to you, so:
performance = IB performance * haswell increase
performance = 1.08 * 1.05 = 1.134

You claimed it was 13% faster, so you'd have
performance = 1.13

It isn't a huge difference in percentage this time but it can be if you continue to do math improperly in threads.

Conversely, if we give bdver1 an arbitrary performance index of 1.0

bdver1 -> bdver2 = 15% increase overall
1.15
bdver2 -> bdver3 = 20% increase overall
1..38

bdver3 being 20% faster than bdver2 would mean bdver3 is 38% faster than bdver1

However if bdver3 is only 30% faster than bdver1 it would have to mathematically be a smaller than 15% increase in performance over bdver2.

Sorry, maybe I am still resentful of all the derps in my CS program who blew at math yet did alright in the class.

I understand your point, and I rounded down, because the math came out to a number just slightly over 13%.

However, my point about the fact that they haven't improved 15% since SB, whether it's my rough 13% or 13.4% is unchanged.

 

[Cazalan]

And they're saying a 9% improvement with APUs just from going from Windows 8 to 8.1. 😉

With MS working so closely with AMD for XBone there could be more bonuses in store.

 

[lilxskull]

I hope AMD can make Intel up their game like they are currently doing to NVIDIA.

 

[juanrga]

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).

It's all in the code there. A module has 2 bdver3 cores. The FP units are shared of course but each core can access them, just they might have to wait.

These are your 2 ALUs.

(define_cpu_unit "bdver3-ieu0" "bdver3_ieu")
(define_cpu_unit "bdver3-ieu1" "bdver3_ieu")
(define_reservation "bdver3-ieu" "(bdver3-ieu0|bdver3-ieu1)")


These are your 2 AGUs.

(define_cpu_unit "bdver3-agu0" "bdver3_agu")
(define_cpu_unit "bdver3-agu1" "bdver3_agu")
(define_reservation "bdver3-agu" "(bdver3-agu0|bdver3-agu1)")

Which makes up the typical 4 integer pipelines you see in the block diagrams. Same as Bulldozer.

Further down you can see how many clock cycles each operation takes. That's where you can do a more fine comparison between architecture changes.

Yes, but that is 2012 code, aka Steamroller A. Can you sure me that AMD is not releasing a Steamroller B with 3 ALUs per core?

After days talking with AMD representative he cannot confirm/deny.

 

[juanrga]

My argument had two parts: one about Kaveri, another about FX. It was not only about APUs...

If market share for dual-cores was of a 0.1% I can sure you that Kaveri 2C wouldn't exit. The dual-core Kaveri exists because dual-cores are popular for both APUs and CPUs.

i didn't address the fx bit because i have too little information on that, and i didn't find it worth addressing.

it's not about marketshare. your argument(!) has a gigantic flaw: 2c kaveri (i.e. single module cpu) exists/will exist because of die harvesting and amd making money off partially defective kaveri silicon instead of throwing those away after looking at some marketshare pie chart.

You seem to believe that using defective silicon always generate revenue, which is not true. Plot the money from reusing that silicon against the money spent on fabricating, testing, packaging, storing, shipping... products. There is a minimum market/price-point beyond which you lost more money by making those products than by just rejecting the defective silicon. That is why AMD is not selling single cores, for instance.

For illustration: consider the extreme case you reuse that defective silicon ($100M) and fabricate product Y ($100M) which nobody purchases. You are loosing $200M, but if you ignore the defective silicon you only miss $100M.

Did some of you even look at the slides I posted?
http://www.slideshare.net/wilfredlin/amdembeddedroadmapunveilvfinal-130906181555

1. AMD only expects embedded/mobile market to account for 40% to 50% of their sales in the next couple of years. THATS LESS THAN HALF OF THEIR PRODUCTS

2. The roadmap CLEARLY SHOWS Temash is not going to be replaced by anything ARM. There are no ARM + GCN low power products on the roadmap at all for 2014.

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

Sweclockers doesn't offer more specifics on the chip (the small SoC for mobile devices), which will reportedly be announced in Q4 and launch some time next year.

This completely contradicts the official slide AMD released in the slides I linked where there is no small ARM with GCN, only one for servers and Temash going on.

I do not understand why you are taking one source (SweClockers) over another source (AMD) about what AMD is planning on doing. I am taking the newer, official AMD embedded/mobile roadmaps over a rumor site's opinion.

Embedded != mobile.

AMD expects embedded market to rise about a 36%, whereas the desktop market is falling. That is why AMD is moving away from tradittional desktop and focusing in new markets.

The embedded roadmap shows how _G-series_ is replaced by Eagle. The roadmap says nothing about Temash; there is no contradiction with the docs reported by techreport and Sweclockers about mobile roadmaps and ARM+GCN tablets replacing Temash tablets. You are confounding chips and roadmaps.

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.

My argument had two parts: one about Kaveri, another about FX. It was not only about APUs...

If market share for dual-cores was of a 0.1% I can sure you that Kaveri 2C wouldn't exit. The dual-core Kaveri exists because dual-cores are popular for both APUs and CPUs.

According to the above data:
- dual cores are about one half the market
- quad cores are about one third the market

AMD could release an octo-core Kaveri APU with quad memory channel and 1024SP tomorrow, just how they could release a hypothetical FX-8570 CPU tomorrow.

They aren't doing it because the market is close to non-existent. Nobody buys a 28nm production line for making APUs/CPUs for ~0.3% of the whole market, except if one is going to sell them at $1200 each, which is not the case.

Nvidia is preparing ultra-high-performance APUs with 8-cores + 1024 SP. I expect an enthusiast line of APUs from AMD, something similar to this

but no today neither tomorrow...

When 2 core CPUs came out, there were FAR more single core CPUs than dual core CPUs. But they continued making dual core CPUs in spite of low market share initially. Now, they no longer make single core CPUs for mainstream consumer PCs, even though they were once 100% of the market.

What's your point?

When you release a new product, you predict which will be the market for it. The market share for a dual core is zero before releasing the first dual core (that is evident, is not?), but it is not zero in the future when you are already selling the product.

The situation with a falling market is different. If you don't sell the products today you will not tomorrow. Then better you stop making those products. That is why AMD is no more making single core CPUs.

My point is: AMD is not releasing products for a ~0.3% of market. That is why there is no FX 8-core Steamroller, neither Steamroller Warsaw.

 

[juanrga]

Nvidia is preparing ultra-high-performance APUs with 8-cores + 1024 SP. I expect an enthusiast line of APUs from AMD, something similar to this

but no today neither tomorrow...

What is the point of Nvidia doing this? So they can make ARM APUs that play Android games super fast?

It was explained before, during several days... Nvidia is going to attack the HPC and the high performance server/desktops market with those products.

I can sure you that people developing the ARM supercomputer will not use it for playing Android games :lol: It will be used for less mundane workloads, such as

http://news.illinois.edu/news/13/0529HIVcapsid_KlausSchulten.html

 

[juanrga]

IIRC Nvidia DENVER arm is to be for cluster servers

I would imagine the format of that would be closer to a vector coprocessor for the iGPU then.

Either way, DENVER from NVidia is going to be about as successful as Tegra, Tegra 2, Tegra 3, and Tegra 4 before it...

Which is to say...not very successful...

http://hexus.net/business/news/components/54085-nvidia-tegra-sales-hit-qualcomms-lte-equipped-chips/

Qualcomm is kicking them around...

I would not be so sure about denver falling flat. It was originally to be used in both desktop and server applications using transmeta code morphing software to run 32 , 64bit windows and android dual boot. The legal team advised nvida against trying to enforce it's x86 license acquired from transmeta. So denver became a server product. There is a demand for these low power server clusters, The market is just not there yet. Based on nvidias size and channel partners Denver will win some designs,

Somehow, I doubt NVidia wins too many...they have some partners and channels for that, but then why hasn't Tegra 2 or Tegra 3 or Tegra 4 or Tegra 4i taken off like a rocket?

Additionally, AMD has SeaMicro to pedal it's commercial micro server clusters...

LOL

Are you aware that Nvidia is competing against giants such as Samsung, Apple, Qualcomm, Intel...? It is not that Tegra 2/3/4 are bad products, it is that better products are available.

Are you aware that Nvidia was a GPU company? You don't transform into a CPU/SoC company in two days. The CPU/SoC dept. is getting experience and polishing its designs closing the gap with competitors.

Tegra 4/4i is already changing things

http://www.fool.com/investing/general/2013/10/17/nvidia-may-finally-have-a-tegra-4-hit-on-its-hands.aspx

But the more important, are you aware that Tegra 2/3/4 is a secondary project for Nvidia? Their main emphasis is Denver and related projects. Nvidia is using standard cores before they get ready its custom core. Developing a custom high-performance core needs lots of time and money. Why do you believe AMD is releasing Seattle/HieroFalcon with standard cores? And that is a company with lots of experience in CPUs

 

[The Q6660 Inside]

My argument had two parts: one about Kaveri, another about FX. It was not only about APUs...

If market share for dual-cores was of a 0.1% I can sure you that Kaveri 2C wouldn't exit. The dual-core Kaveri exists because dual-cores are popular for both APUs and CPUs.

i didn't address the fx bit because i have too little information on that, and i didn't find it worth addressing.

it's not about marketshare. your argument(!) has a gigantic flaw: 2c kaveri (i.e. single module cpu) exists/will exist because of die harvesting and amd making money off partially defective kaveri silicon instead of throwing those away after looking at some marketshare pie chart.

You seem to believe that using defective silicon always generate revenue, which is not true. Plot the money from reusing that silicon against the money spent on fabricating, testing, packaging, storing, shipping... products. There is a minimum market/price-point beyond which you lost more money by making those products than by just rejecting the defective silicon. That is why AMD is not selling single cores, for instance.

For illustration: consider the extreme case you reuse that defective silicon ($100M) and fabricate product Y ($100M) which nobody purchases. You are loosing $200M, but if you ignore the defective silicon you only miss $100M.

Did some of you even look at the slides I posted?
http://www.slideshare.net/wilfredlin/amdembeddedroadmapunveilvfinal-130906181555

1. AMD only expects embedded/mobile market to account for 40% to 50% of their sales in the next couple of years. THATS LESS THAN HALF OF THEIR PRODUCTS

2. The roadmap CLEARLY SHOWS Temash is not going to be replaced by anything ARM. There are no ARM + GCN low power products on the roadmap at all for 2014.

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

Sweclockers doesn't offer more specifics on the chip (the small SoC for mobile devices), which will reportedly be announced in Q4 and launch some time next year.

This completely contradicts the official slide AMD released in the slides I linked where there is no small ARM with GCN, only one for servers and Temash going on.

I do not understand why you are taking one source (SweClockers) over another source (AMD) about what AMD is planning on doing. I am taking the newer, official AMD embedded/mobile roadmaps over a rumor site's opinion.

Embedded != mobile.

AMD expects embedded market to rise about a 36%, whereas the desktop market is falling. That is why AMD is moving away from tradittional desktop and focusing in new markets.

The embedded roadmap shows how _G-series_ is replaced by Eagle. The roadmap says nothing about Temash; there is no contradiction with the docs reported by techreport and Sweclockers about mobile roadmaps and ARM+GCN tablets replacing Temash tablets. You are confounding chips and roadmaps.

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.

My argument had two parts: one about Kaveri, another about FX. It was not only about APUs...

If market share for dual-cores was of a 0.1% I can sure you that Kaveri 2C wouldn't exit. The dual-core Kaveri exists because dual-cores are popular for both APUs and CPUs.

According to the above data:
- dual cores are about one half the market
- quad cores are about one third the market

AMD could release an octo-core Kaveri APU with quad memory channel and 1024SP tomorrow, just how they could release a hypothetical FX-8570 CPU tomorrow.

They aren't doing it because the market is close to non-existent. Nobody buys a 28nm production line for making APUs/CPUs for ~0.3% of the whole market, except if one is going to sell them at $1200 each, which is not the case.

Nvidia is preparing ultra-high-performance APUs with 8-cores + 1024 SP. I expect an enthusiast line of APUs from AMD, something similar to this

but no today neither tomorrow...

When 2 core CPUs came out, there were FAR more single core CPUs than dual core CPUs. But they continued making dual core CPUs in spite of low market share initially. Now, they no longer make single core CPUs for mainstream consumer PCs, even though they were once 100% of the market.

What's your point?

When you release a new product, you predict which will be the market for it. The market share for a dual core is zero before releasing the first dual core (that is evident, is not?), but it is not zero in the future when you are already selling the product.

The situation with a falling market is different. If you don't sell the products today you will not tomorrow. Then better you stop making those products. That is why AMD is no more making single core CPUs.

My point is: AMD is not releasing products for a ~0.3% of market. That is why there is no FX 8-core Steamroller, neither Steamroller Warsaw.

Have faith.. According to your logic, AMD would not have produced the 7970, it is only 0.7% of the share, like plenty of other cards... Aaand because steam is the only service people use for gaming..... The 8 core share is slowly growing..

 

[de5_Roy]

You seem to believe that using defective silicon always generate revenue, which is not true.

not seem, i don't believe it at all. dunno what gave you that idea. none of my recent posts contain anything that points to such conclusion.

Plot the money from reusing that silicon against the money spent on fabricating, testing, packaging, storing, shipping... products. There is a minimum market/price-point beyond which you lost more money by making those products than by just rejecting the defective silicon. That is why AMD is not selling single cores, for instance.

waste not, want not.
how can amd make a single 'core' cpu after going through all the trouble of introducing modules and bulldozer uarch? the minimum cpu 'block' is a single module, dual 'core' cpu.
to calculate that minimum point, you need to calculate the cost of everything associated with fabrication to the point of sale. you'll see that the point you're talking about is much lower than you're making out to be. i won't go further into detail because most people here will explain it much better than i can. iirc.... you'll find a detailed explanation or binning and cost in the piledriver thread. i only replied because you claimed that kaveri is dual module because of marketshare and steam's hardware survey. it is not.
.... i found this forum thread that sheds some light (i am not claiming that these are facts):
http://beyond3d.com/showthread.php?t=47829

For illustration: consider the extreme case you reuse that defective silicon ($100M) and fabricate product Y ($100M) which nobody purchases. You are loosing $200M, but if you ignore the defective silicon you only miss $100M.

riight... it's not as simple as that. you're better off asking amd to disclose their per-cpu costs. good luck.