AMD CPUs, SoC Rumors and Speculations Temp. thread 2

[jimmysmitty]

Rough estimation of Zen costs

Die size

1) About 50% of Kaveri die is 4 SR-core CPU.

2) Samsung/Glofo 14nm is about 2x more dense than 28nm.

3) Zen core ~ 2 SR cores

4) Thus 8 Zen-core CPU (without L3) would occupy about 245mm² on 14nm.

5) Add L3 and the die size will be close to 300mm².

Transistor cost

6) Not all transistors are the same. 14FF is more expensive than 28PL; first because of the jump to 3D and, second, because the smaller size requires extra techniques as multi-patterning.

7) Thus ~300mm² die on 14FF will be about so expensive as ~600mm² on 28PL [I am using cost data from another guy].

Design/Validation

8) All of above simply considers number of transistors and transistor costs. Add the R&D cost from designing Zen micro-architecture from scratch. Add the cost from validating such design (validation costs increase nonlinearly with design complexity).

Extras

9) AMD has a huge debt and other economic problems and needs huge margins. This is why AMD will not focus on cheap products anymore. In Lisa Su terms: AMD is no longer the cheap company.

Conclusion: As I stated many time ago, Zen will be an expensive product. Anyone expecting octo-core Haswell-like performance at 8350 prices or i5 prices will be disappointed.

I would imagine you will see it for roughly the same as a higher end i7 quad though...which is very fair, especially considering.

Only if the performance is the same. If it is lower it wont be. If it is higher it will be priced higher.

AMD needs to make money and playing the cheap company gets them no where.

You make the assumption price is based off performance. The top chips will be selling north of $500 at launch, regardless of performance. They'll correct over time as supply/demand, which is performance driven, takes hold. But initial prices will be jacked up to catch the early adopter.

That is sort of what I mean. Pricing starts somewhere and ends based on performance, supply and demand.

 

[ComputerSecurityGuy]

I know- start high and lower down is the rule. People complain and jump ship if you raise prices. But if you lower them, people come on just because of the "value". I would expect, at launch, the top end Zen chip priced around $300-500, and the top end Zen APU at $200-300. In reality they will get cheaper after a month or two, slowly decreasing in price until Zen+.

 

[jimmysmitty]

I know- start high and lower down is the rule. People complain and jump ship if you raise prices. But if you lower them, people come on just because of the "value". I would expect, at launch, the top end Zen chip priced around $300-500, and the top end Zen APU at $200-300. In reality they will get cheaper after a month or two, slowly decreasing in price until Zen+.

Unless it actually does beat Intels top offering then AMD will price it accordingly to start.

 

[ComputerSecurityGuy]

If it beats Intel then I would expect it to be the same pricing as Intel chips. Lower if the Intel ones are better. Nobody would buy an i5 if a Zen FX priced identically or lower annihlated it. They might if they priced it higher.

 

[de5_Roy]

AMD's Next Gen Products Taped Out, Possibly Delayed
http://www.tomshardware.com/news/amd-zen-pascal-taped-out,30751.html

Unless it actually does beat Intels top offering then AMD will price it accordingly to start.

and retailers will add on top of that depending on chip supply.

 

[Vogner16]

I know- start high and lower down is the rule. People complain and jump ship if you raise prices. But if you lower them, people come on just because of the "value". I would expect, at launch, the top end Zen chip priced around $300-500, and the top end Zen APU at $200-300. In reality they will get cheaper after a month or two, slowly decreasing in price until Zen+.

I would be shocked to see anything less than $600 for the top end 8 core zen chip.

You heard it here first! I will link this post in the future for the haters!

top end apu I don't think we have enough information on yet. how many cores? how many CU's? what will intels new L4 cashe cost? we don't know...

 

[8350rocks]

I know- start high and lower down is the rule. People complain and jump ship if you raise prices. But if you lower them, people come on just because of the "value". I would expect, at launch, the top end Zen chip priced around $300-500, and the top end Zen APU at $200-300. In reality they will get cheaper after a month or two, slowly decreasing in price until Zen+.

I would be shocked to see anything less than $600 for the top end 8 core zen chip.

You heard it here first! I will link this post in the future for the haters!

top end apu I don't think we have enough information on yet. how many cores? how many CU's? what will intels new L4 cashe cost? we don't know...

likely 4-6 cores and 8-12 CUs on the top end APU...assuming at most a combination of something like 4c/12CU-6c/10CU

 

[Vogner16]

likely 4-6 cores and 8-12 CUs on the top end APU...assuming at most a combination of something like 4c/12CU-6c/10CU

does anyone have a link to something to shed some light to this? 4 core 12 CU would be amazing, but maybe too good to be true.

I know juan can do some math to determine based on die size 😛

 

[Cazalan]

You make the assumption price is based off performance. The top chips will be selling north of $500 at launch, regardless of performance. They'll correct over time as supply/demand, which is performance driven, takes hold. But initial prices will be jacked up to catch the early adopter.

I don't think there are that many die hard AMD HEDT fans left these days. You'll be looking at 4+ years since the 8350 launch (Oct 2012) and the Zen launch (Early 2017). If Intel is selling Hex cores for $350-400, and Zen cores are behind on IPC then not enough people will buy. You may however be able to pitch more cores for the same price. Hex-Zen vs Quad-KabyLake, or Octo-Zen vs Hex-KabyLake.

We still have AMD launching more 32nm parts. In this case selling a Hex core vs a Dual core. They are very cognizant that performance is what sells.

http://wccftech.com/amd-fx-6330-black-edition-processor-silent-launch/

 

[Cazalan]

likely 4-6 cores and 8-12 CUs on the top end APU...assuming at most a combination of something like 4c/12CU-6c/10CU

does anyone have a link to something to shed some light to this? 4 core 12 CU would be amazing, but maybe too good to be true.

I know juan can do some math to determine based on die size 😛

They're going to need way more CU than that. By the time Zen APUs launch the 10nm CannonLake will be out. You can bet that Intel will double up on their EUs again @10nm. They have to keep ahead of Apple SoCs which are gaining massively on the iGPU front.

 

[hawkeyethe2nd]

These next several months are gonna be interesting. I can't wait

 

[gamerk316]

These next several months are gonna be interesting. I can't wait

You mean "boring". AMD isn't releasing anything for at least a year, possibly longer. Likewise, Intel's cadence is clearly starting to slip. 2016 is going to be a cosmic wasteland on the CPU front.

 

[jimmysmitty]

likely 4-6 cores and 8-12 CUs on the top end APU...assuming at most a combination of something like 4c/12CU-6c/10CU

does anyone have a link to something to shed some light to this? 4 core 12 CU would be amazing, but maybe too good to be true.

I know juan can do some math to determine based on die size 😛

They're going to need way more CU than that. By the time Zen APUs launch the 10nm CannonLake will be out. You can bet that Intel will double up on their EUs again @10nm. They have to keep ahead of Apple SoCs which are gaining massively on the iGPU front.

Apples iGPU is pretty overrated. In most tests it does great in the off screen and at whatever resolution it is running at but on screen and higher resolutions tend to show how weak it can be.

I don't think Intel will worry about that. When the entire AMD SoC can give the performance Apple has in their Macbooks with Intel then Intel would worry. Only thing I see is if Apple is willing to have two separate ecosystem for their PC lineup and if they are willing to throw their customers under the bus again with a hardware change. That or they just throw iOS on touchscreen Macbooks and call it "Amazing".

 

[Vogner16]

Apples iGPU is pretty overrated.

and so is the "amazing" A9x. find me a benchmark other than geekbench 3...

apple is still far behind intel and amd for that matter on the soc front. they have made major leaps however.

2016 does look to be quite barron until the end.

amd will have their desktop Carrizo come out soon, but nobody here is looking forward to that.

Intel kaby lake is looking to be on schedule to release with zen.
zen apu's will follow soon after but before cannon lake.

 

[hawkeyethe2nd]

These next several months are gonna be interesting. I can't wait

You mean "boring". AMD isn't releasing anything for at least a year, possibly longer. Likewise, Intel's cadence is clearly starting to slip. 2016 is going to be a cosmic wasteland on the CPU front.

Well I mean some of the rumors and leaks have been really fun to watch and listen to. There is only going to be more of them as we get closer, Some guy got the R9 390x early at Best Buy or something and that was exciting. Whether or not AMD releases actual information I love the hype. ]These next several months will be interesting here and hopefully AMD can turn it around and retake some market share, I guess we will see.

 

[con635]

Apples iGPU is pretty overrated.

2016 does look to be quite barron until the end.

amd will have their desktop Carrizo come out soon, but nobody here is looking forward to that.

I'm looking forward to it, it looks quite interesting in aida64 fp vs steamroller, nearly catches @1.6ghz an fx4300 at full speed and we also have ddr4 which I dont believe is used in these benches, (also not counting ddr4 is 40% better igp vs kaveri as well iirc)
http://www.planet3dnow.de/cms/18564-amd-piledriver-vs-steamroller-vs-excavator-leistungsvergleich-der-architekturen/subpage-aida-fpu/

(credit to 'themissable' on ocn)

 

[Vogner16]

I'm looking forward to it, it looks quite interesting in aida64 fp vs steamroller, nearly catches @1.6ghz an fx4300 at full speed and we also have ddr4 which I dont believe is used in these benches, (also not counting ddr4 is 40% better igp vs kaveri as well iirc)
http://www.planet3dnow.de/cms/18564-amd-piledriver-vs-steamroller-vs-excavator-leistungsvergleich-der-architekturen/subpage-aida-fpu/

(credit to 'themissable' on ocn)

Even with floating point calculations about the different competent units makes Excavator an excellent figure. While Steamroller must struggle with its stripped-FPU in order not to be caught unawares by Piledriver, lays Excavator by an impressive 50 to 99% to - at the same clock speeds mind.

the translation to that link. specific benchmark for FPU performance gains dosent say much for excavator in general, but than to show IPC in excavator is better than piledriver and amd's zen claim over excavator is 40% so once again don't compare to piledriver directly.

 

[Vogner16]

http://www.planet3dnow.de/cms/18564-amd-piledriver-vs-steamroller-vs-excavator-leistungsvergleich-der-architekturen/subpage-rendering-cinebench/

a much more relevant link to what I use daily. shows a 13% IPC gain over Piledriver in multicore and 10.8% IPC gain over Piledriver in single core.

once again proof of my point to AMD's claim of 50% IPC over Piledriver. (Take a note here. AMD claimed 10% IPC over piledriver with excavator and they delivered.)

EDIT: they originally claimed 5% over PD for steamroller and then 5% over steamroller for excavator. total of 10% IPC gain

 

[Cazalan]

http://www.planet3dnow.de/cms/18564-amd-piledriver-vs-steamroller-vs-excavator-leistungsvergleich-der-architekturen/subpage-rendering-cinebench/

a much more relevant link to what I use daily. shows a 13% IPC gain over Piledriver in multicore and 10.8% IPC gain over Piledriver in single core.

once again proof of my point to AMD's claim of 50% IPC over Piledriver. (Take a note here. AMD claimed 10% IPC over piledriver with excavator and they delivered.)

EDIT: they originally claimed 5% over PD for steamroller and then 5% over steamroller for excavator. total of 10% IPC gain

I don't think there was ever any question that the newer cores made their stated IPC gains. The issue was that the process changes impaired the final clock speeds (regression). This made for more efficient processors but not more powerful processors.

At higher clocks Steamroller throws out some big numbers.
http://browser.primatelabs.com/geekbench3/3353217

 

[Vogner16]

http://www.planet3dnow.de/cms/18564-amd-piledriver-vs-steamroller-vs-excavator-leistungsvergleich-der-architekturen/subpage-rendering-cinebench/

a much more relevant link to what I use daily. shows a 13% IPC gain over Piledriver in multicore and 10.8% IPC gain over Piledriver in single core.

once again proof of my point to AMD's claim of 50% IPC over Piledriver. (Take a note here. AMD claimed 10% IPC over piledriver with excavator and they delivered.)

EDIT: they originally claimed 5% over PD for steamroller and then 5% over steamroller for excavator. total of 10% IPC gain

I don't think there was ever any question that the newer cores made their stated IPC gains. The issue was that the process changes impaired the final clock speeds (regression). This made for more efficient processors but not more powerful processors.

At higher clocks Steamroller throws out some big numbers.
http://browser.primatelabs.com/geekbench3/3353217

if you read this thread through there is lots of question to amd's ability to hit advertised IPC. gamerk even goes as far as to think amd will only hit 20%. (not calling him out but he is one of the doubtful group.)

and Yes juanrga has stated and linked a source to estimated clocks for8 core zen to be around 2.8 and 3.5 turbo.

this article while it is in german, does a great job in showing how overall excavator shows a 4-13% ipc improvement over steamroller and steamroller showed a 3-10% ipc gain over piledriver. this includes the lack of an L3 cashe on excavator and steamroller not being factored in. if you include that extra gain in performance from an L3 that will be on zen the 40% gain over excavator has no reason to not be found truthful.

I know you all look to BD and say that failed, but what was said about BD wasn't physical data such as a value of improvement, but more so of "faster than intel" a hype statement with no background as faster than intel could mean faster in one test or faster in all or higher clock speeds etc... there was also a new process to consider and a new core design CMT.

zen is only one of these things. new process. we know smt is great. we know a % performance that amd claims to be able to gain from zen. we know the number of cores. we have a reasonable range of clock speeds zen will reach. what is left to speculate? actual benchmarks? the chip was just taped out!

all im saying is amd HAS delivered in the past 5 years to what they said IPC gains would be and what power improvement will be. Carrizo uses 40% of the power kaveri used at idle. Carrizo is 10% faster than piledriver at the same clocks even without an L3 cashe.

I understand why you think they wont deliver, but look at the last 5 years and tell me they wont deliver and I don't buy it.

 

[Cazalan]

all im saying is amd HAS delivered in the past 5 years to what they said IPC gains would be and what power improvement will be. Carrizo uses 40% of the power kaveri used at idle. Carrizo is 10% faster than piledriver at the same clocks even without an L3 cashe.

I understand why you think they wont deliver, but look at the last 5 years and tell me they wont deliver and I don't buy it.

I think they can hit the 40% target they claim over Excavator. My only issue is that it won't be available when the 16nm GPUs launch. Next year we will be seeing a lot of DX12 games and I've been holding off for a CPU upgrade.

 

[jimmysmitty]

all im saying is amd HAS delivered in the past 5 years to what they said IPC gains would be and what power improvement will be. Carrizo uses 40% of the power kaveri used at idle. Carrizo is 10% faster than piledriver at the same clocks even without an L3 cashe.

I understand why you think they wont deliver, but look at the last 5 years and tell me they wont deliver and I don't buy it.

I think they can hit the 40% target they claim over Excavator. My only issue is that it won't be available when the 16nm GPUs launch. Next year we will be seeing a lot of DX12 games and I've been holding off for a CPU upgrade.

IDK. 40% is a lot. I mean it is plausible, look at Netburst -> Conroe, but it still seems like a very big jump.

 

[hawkeyethe2nd]

all im saying is amd HAS delivered in the past 5 years to what they said IPC gains would be and what power improvement will be. Carrizo uses 40% of the power kaveri used at idle. Carrizo is 10% faster than piledriver at the same clocks even without an L3 cashe.

I understand why you think they wont deliver, but look at the last 5 years and tell me they wont deliver and I don't buy it.

I think they can hit the 40% target they claim over Excavator. My only issue is that it won't be available when the 16nm GPUs launch. Next year we will be seeing a lot of DX12 games and I've been holding off for a CPU upgrade.

IDK. 40% is a lot. I mean it is plausible, look at Netburst -> Conroe, but it still seems like a very big jump.

Big jump indeed. But With all of the improvements they are bringing (Die size , SMT, DDR4) they should fairly easily hit it.. will they? I believe so but others won't and we will have to wait and see.

 

[gamerk316]

http://www.planet3dnow.de/cms/18564-amd-piledriver-vs-steamroller-vs-excavator-leistungsvergleich-der-architekturen/subpage-rendering-cinebench/

a much more relevant link to what I use daily. shows a 13% IPC gain over Piledriver in multicore and 10.8% IPC gain over Piledriver in single core.

once again proof of my point to AMD's claim of 50% IPC over Piledriver. (Take a note here. AMD claimed 10% IPC over piledriver with excavator and they delivered.)

EDIT: they originally claimed 5% over PD for steamroller and then 5% over steamroller for excavator. total of 10% IPC gain

I don't think there was ever any question that the newer cores made their stated IPC gains. The issue was that the process changes impaired the final clock speeds (regression). This made for more efficient processors but not more powerful processors.

At higher clocks Steamroller throws out some big numbers.
http://browser.primatelabs.com/geekbench3/3353217

if you read this thread through there is lots of question to amd's ability to hit advertised IPC. gamerk even goes as far as to think amd will only hit 20%. (not calling him out but he is one of the doubtful group.)

20% PERFORMANCE, due to an expected decline in base clocks. I wouldn't be shocked to see IPC gains in the 25-35% range, but if you have to decline clocks 20% in the process, what have you really gained performance wise?

Note there are still unknowns, specifically with the SMT configuration. Obviously, a more friendly SMT algorithm will show larger gains, which is why I focus on the "typical case", not "best case" performance numbers.

 

[Vogner16]

@gamerk which is why I did some tests to show what piledriver was capable of at expected clock speeds of zen and that's how we got the final conclusion of ivy to haswell performance.

your correct with the SMT configuration however. we don't know how well it will scale.

the important thing to show with zen is not to overhype it as BD was or it will fail for sure. to promote false promises such as faster than intel or great for gamers etc, without concrete evidence to back it up will surely prove to zen's failure just as BD was.

@calazan I think a January 2017 timeline for a zen build is very likely. you can probably get a 16nm gpu early if you want. but I bet release dates will be within 6 months of each other.