AMD CPU speculation... and expert conjecture

[juanrga]

It is cubic only as approximation. Your mistake was that you treated voltage and frequency as independent quantities, but V = V(F)

It's an approximation only valid when V = F.

No! It is an approximation valid when V~F i.e. when V=aF with "a" being the proportionality constant (Idontcare got a=0.4588 for a concrete i7). Why don't you read what I wrote and the references cited, instead posting nonsense?

 

[juanrga]

Samsung pushing forward with FD-SOI @ 28nm and 14nm.

https://www.semiwiki.com/forum/content/4194-fdsoi-jump-start-2015-tokyo-week.html

Hmm...Samsung is going to offer 14nm FD-SOI in their foundry? Really?

Someone here said FD-SOI was dead. Who was it??? I am trying to remember, but memory is failing at the moment.

I think Samsung just partnered with a big chipmaker too...gosh, all the recent news, I cannot recall who it was.

Oh well...I must be senile.

ST-Micro is the only foundry that chose FD-SOI for the 14nm node. Any other foundry (Intel, Samsung, Glofo, TSMC. UMC, IBM...) rejected FD-SOI and chose FinFETs for their main node because FinFETs are better for high-performance chips and provide a clever evolution path towards the 10nm node and beyond.

ST-Micro will use FD-SOI for low performance chips used in phones and tablets. ST-Micro expect a huge increase in demand of cheap phones coming from Asia markets and has partnered with Samsung to use them as second source foundry to satisfy the demand.

Less than the 6% of total volume of foundries production for 14nm node will be FD-SOI; despite the multi-year hype, the technology is only used in niche cases.

Of course, this partnership between Samsung and ST-Micro means nothing for AMD. Or does someone here pretend that Zen and K12 will be made on 14nm FD-SOI? Because has the same validity (i.e. none) how when some people said us that Kaveri was made on SOI...

 

[juanrga]

For those of you confused about the new memory tech coming out in the next few years:

http://www.extremetech.com/computing/197720-beyond-ddr4-understand-the-differences-between-wide-io-hbm-and-hybrid-memory-cube

The article contradicts when in the table, puts to HBM as giving more bandwidth than HMC, but in the figure just below puts HMC as giving more bandwidth than HBM. I have to check my docs, but I think HMC achieves more than the 240 GB/s mentioned in the table. I think 320 was the top bandwidth.

One of the goals of HMC is lower latency because has been designed for future very high-performance CPUs. However HBM seems to be only addressing bandwidth:

HBM is explicitly designed for graphics

AMD and Nvidia chose HBM. Intel, HP, and ARM chose HMC. I prefer HMC because will also improve CPUs.

 

[juanrga]

AMD doesn't have the money to do a second die on a different process. In fact it seems that no Kaveri refresh (a la Richland) is coming. It seems that AMD will release some new Kaveri models, but without anything really new.

Firmware changes to Kaveri, a more mature node, and little more. There is nothing like "stacked cache/hbm".

I was going to argue contradiction but then realized that by a la Richland you meant small tweaks on the architectural level. So a refresh would entail higher frequencies on nearly identical parts? Or even just higher binned parts?

I was going to point out the texture fill numbers in the SiSoft benchmarks, however, looks like wccftech at least got the benchmark commentary correct.

Carrizo will have massive iGPU included, aimed specifically at mobile/ULP designs, and will probably have the top part coming in around ~65W at most.

There is another part coming for DT, different codename, will not be the same as the ULP parts scaled upwardly.

Zen and K12 still in the works...

I have a strange feeling that 'different codename' is Kaveri or some clever play on the word.

There is no refresh with higher clocks. Don't expect a 4.1GHz 'Kaveri'. Don't expect a new 'Kaveri' with full HSA. Don't expect a new 'Kaveri' with HBM. The only 'refresh' will be firmware changes, different steeping, and minor changes like that. A more mature process could reduce power consumption a bit.

There will be new models above Carrizo but there is absolutely nothing really new under the hood.

 

[noob2222]

2) To compete with 14nm Broadwell on efficiency they have to reduce clocks. Whereas performance increases with frequency is about linear, power increases as cube frequency: P ~ f^3

Where do you come up with this pseudo math?

P = C*F*V^2

http://en.wikipedia.org/wiki/Frequency_scaling

Power remains linear with frequency. The reason the power starts going crazy is they have to bump the core voltage to remain stable.

Voltage is a function of frequency. As explained in my article about power consumption of CPUs, substituting the (very good) approximation V ~ f on the expression for the dynamic power gives the well-known cubic dependence that is quoted in many sites

https://software.intel.com/en-us/articles/the-new-era-of-tera-scale-computing

A concrete case. First the almost linear dependence of voltage with frequency

http://i272.photobucket.com/albums/jj163/idontcare_photo_bucket/Intel%20Core%20i7-2600K/VoltageCurve.png

and then the almost cubic dependence of (dynamic) power with frequency

http://i272.photobucket.com/albums/jj163/idontcare_photo_bucket/Intel%20Core%20i7-2600K/PtotalVccTGHzDataGraph.png

Cazalan 1 > juanrga 0

Real math > made up math

ya, but V = V(F) = ∞ so juan can't be wrong infinitely, the sheer size of his brain makes the entire population of earth look infinitesimally small. Obviously that's why hes working for the worlds greatest think tank (consisting of himself) and arguing on forums.

If everyone would stop arguing with Juan and just listen to him, the world would be at peace, starvation wouldn't exist, we would be traveling to other galaxies in a fraction of a second, and most importantly, AMD would belong to Intel.

 

[juanrga]

http://www.tomshardware.com/news/samsung-amd-win-win-acquisition,28403.html#xtor=EPR-8886

Agree with the part reporting AMD troubles and why AMD doesn't have future: lack of competitive products, lack of money, not enough R&D, behind competence on all markets,...

Disagree with the part about Samsung purchasing AMD. I agree with gamerk316 on that there's zero benefit for Samsung to buy AMD.

 

[esrever]

amd Q4 numbers

http://ir.amd.com/phoenix.zhtml?c=74093&p=irol-newsArticle&ID=2008997

"Cash, cash equivalents and marketable securities were $1.04 billion at the end of the quarter, up $102 million from the end of the prior quarter.
Total debt at the end of the quarter was $2.21 billion, flat from the prior quarter. "
Looks like they will be fine from cash perspective which means they will probably stay afloat pretty easily through 2015.

 

[juanrga]

Some new info about Carrizo (I hope is accurate).

The IPC gain of Excavator over Steamroller will be smaller than I expected: Excavator is about 15% faster than Steamroller clock for clock.

However, part of that IPC gain will be compensated by lower clocks. Excavator doesn't scale up well beyond 3GHz. Any Carrizo above 35W was canceled because would be slower than Kaveri!

Base frequencies for Carrizo SoCs will be from 1.6GHz to 2.8GHz or so. Some models will report turbos of between 3.0GHz and 3.4GHz. However those turbos will be almost unreachable (only short time and when GPU is almost iddle). Sustainable turbos will be sub 3.0GHz very probably.

The top 35W Carrizo SoC will be probably clocked at 2.8GHz (base) and 3.4GHz (turbo). However sustainable turbo must be about 2.9GHz, which is less than the 3.6GHz turbo of the FX-7600P.

Combining the expected IPC gain (15%) with the expected reduction on max clocks (24%) means that the top Carrizo will barely match the Kaveri FX-7600P on pure CPU workloads. Where Carrizo will shine will be on the iGPU part.

I welcome a better iGPU but the iCPU part is completely disappointing. Excavator is nowhere what AMD promised us years ago.

 

[8350rocks]

Juan: find a way to doom and gloom this news:

http://www.amd.com/en-us/press-releases/Pages/amd-firepro-awarded-2014nov20.aspx

http://www.green500.org/

"bb.b.b.b.b.b.bbb...but, NVidia/Intel...."

Always something with you...surprised you overlooked AMD making #1 on the green500. Then again, if a new NVidia HPC had moved up, you would have been shouting from the hills about how great it was.

 

[juanrga]

amd Q4 numbers

http://ir.amd.com/phoenix.zhtml?c=74093&p=irol-newsArticle&ID=2008997

Wow. It is even poor than expected! All divisions lost revenue with 22% lost since last year. Net loss of $364 million and total debt of $2.21 billion.

But the expectations for 2105 are not better. For Q1 2015, AMD expects revenue to decrease another 15 percent. This will translate into more personnel cuts, more cuts on R&D, more canceled/delayed projects,...

 

[juanrga]

Juan: find a way to doom and gloom this news:

http://www.amd.com/en-us/press-releases/Pages/amd-firepro-awarded-2014nov20.aspx

http://www.green500.org/

"bb.b.b.b.b.b.bbb...but, NVidia/Intel...."

Always something with you...surprised you overlooked AMD making #1 on the green500. Then again, if a new NVidia HPC had moved up, you would have been shouting from the hills about how great it was.

That old news was given and discussed before. I already explained why that has little relevance and even did a prediction that will be confirmed soon.

 

[juanrga]

New benchmarks: http://www.sisoftware.eu/rank2011d/show_system.php?q=cea598aa98af9ba391b7d0edc0f1d7a598a88ee7daebcda598ad8bf3ceffd9bcd9e4d4f281bc84&l=en

This time 2.5GHz/3.5GHz

Turbo is 3.4GHz.

The base frequency is also wrong. It is 2.1GHz.

This Carrizo sample gives about 15% more GOPS than Kaveri.

Just default clock. You know, CPU clocks can be changed :>

Either was overclocked or Sandra is reading wrong the clocks of the sample. I hope is the latter, otherwise would be slower than Kaveri.

Ok, I got confirmation the sample was overclocked (Pb2 profile) for the run.

 

[juanrga]

AMD Carrizo Benchmarks Leaked – Double The Performance of Kaveri
http://wccftech.com/amd-carrizo-graphics-benchmarks-leaked-double-performance-kaveri/

I was going to point out the texture fill numbers in the SiSoft benchmarks, however, looks like wccftech at least got the benchmark commentary correct.

No.

Wccftech are comparing Carrizo texture fill numbers (607 Mpix/s) against Kaveri APUs scores using an old driver (the 270 Mpix/s mentioned by Wccftech). Kaveri with the last driver update increases the GP Processing score to 543 Mpix/s, which is a 90% of the score achieved by Carrizo.

Carrizo brings a better iGPU than Kaveri, but Wccftech's "Double The Performance of Kaveri" was nonsense or hype or just both.

 

[palladin9479]

Cease the back and forth, personal insults and multiple serial posts. Final warning, this thread won't go to sh!t again.

 

[esrever]

http://tabtec.com/windows/heres-microsoft-actually-prices-windows-8-1-bing

Why we don't see any amd tablets. I don't know how this is legal. Maybe it will change with Windows 10. I'm sure intel is paying microsoft for this.

 

[juanrga]

http://tabtec.com/windows/heres-microsoft-actually-prices-windows-8-1-bing

Why we don't see any amd tablets. I don't know how this is legal. Maybe it will change with Windows 10. I'm sure intel is paying microsoft for this.

Why wouldn't be legal? Intel has the public contra-revenue program to increase the market share of x86-based tablets against ARM-based tablets. The goal was to ship about 40 million x86-based tablets. I wouldn't expect Intel to incentive OEMs to use chips from AMD.

 

[de5_Roy]

AMD Carrizo launch in Q2 2015
http://www.fudzilla.com/news/notebooks/36810-amd-carrizo-launch-in-q2-2015
i bet wccfaultytech will do a die shot analysis(!).

 

[8350rocks]

http://tabtec.com/windows/heres-microsoft-actually-prices-windows-8-1-bing

Why we don't see any amd tablets. I don't know how this is legal. Maybe it will change with Windows 10. I'm sure intel is paying microsoft for this.

Why wouldn't be legal? Intel has the public contra-revenue program to increase the market share of x86-based tablets against ARM-based tablets. The goal was to ship about 40 million x86-based tablets. I wouldn't expect Intel to incentive OEMs to use chips from AMD.

Because AMD already won a lawsuit against Intel for anti-competitive practices before and Intel never stopped?

 

[truegenius]

Voltage is a function of frequency. As explained in my article about power consumption of CPUs, substituting the (very good) approximation V ~ f on the expression for the dynamic power gives the well-known cubic dependence that is quoted in many sites

https://software.intel.com/en-us/articles/the-new-era-of-tera-scale-computing

A concrete case. First the almost linear dependence of voltage with frequency

http://i272.photobucket.com/albums/jj163/idontcare_photo_bucket/Intel%20Core%20i7-2600K/VoltageCurve.png

and then the almost cubic dependence of (dynamic) power with frequency

http://i272.photobucket.com/albums/jj163/idontcare_photo_bucket/Intel%20Core%20i7-2600K/PtotalVccTGHzDataGraph.png

I see your rooky mistake now. You took the dumbed down example from Intel where they normalized voltage and frequency to 1, and assumed that is valid for real processors as well. It was just an example so the layman can follow along.

No. I gave you the link as simple example of the cubic law being quoted. The cubic law appears in books and papers dealing with CPU design, which evidently you did never read. I also give you the figures with data measured for a real processor: a SB i7. Of course the cubic law also applies to AMD designs. It is the reason why a FX-CPU @ 4GHz is rated at 125W, whereas the same CPU @ 4.7GHZ is rated at 220W.

Intel spends more money but they're also in more segments and a significant portion of that is on the fabs. If AMD can focus their R&D wisely they can still compete in some areas. ARM spends 1/3 of what AMD does on R&D and that isn't holding them back at all.

Apples to oranges. We are comparing budgets for x86 (Zen) vs x86 (Skylake). Thanks to the ISA, it is cheaper to develop an ARM CPU than a x86 CPU.

Andrew Feldman, former server head at AMD, mentioned that one could build an entirely custom chip using the ARM architecture in about 18 months for about $30 million in development costs. But it requires three or four-years and $300 million to $400 million to build an x86-based one.

That is why several companies are developing ARM server chips with same performance or even more than the best x86 designs.

it is up to us to increase frequency while increasing voltage and vice versa
so we can't just replace V with F or vice versa
take it like this
i have 1090t, running @3.6ghz @1.3v
according to your V ~ F
it will mean that if i increase voltage to 2v and keep same frequency then power will remain same because we kept frequency at same, but we would be consuming ~136% more power ( neglecting electron migration / leakage )
or if we keep increase F to 4Ghz and decrease V to 1V then according to your " P is directly proportional to F^3 " thing , we would be consuming ~38% more power but in reality we will be consuming less ( due to lower voltage )

to get a better idea of this, i will ( strongly ) suggest ( else you know peeps of toms, they will fight (any overclocker will) with you on your P ∝ F^3 comment ) you to experiment with your pc, IBT stress testig software and a kill-a-watt like instrument to see power consumption variance with same speed different volts, and different speed same volts

 

[juanrga]

http://tabtec.com/windows/heres-microsoft-actually-prices-windows-8-1-bing

Why we don't see any amd tablets. I don't know how this is legal. Maybe it will change with Windows 10. I'm sure intel is paying microsoft for this.

Why wouldn't be legal? Intel has the public contra-revenue program to increase the market share of x86-based tablets against ARM-based tablets. The goal was to ship about 40 million x86-based tablets. I wouldn't expect Intel to incentive OEMs to use chips from AMD.

Because AMD already won a lawsuit against Intel for anti-competitive practices before and Intel never stopped?

My point is that the contra-revenue program is legal.

 

[anxiousinfusion]

There is no refresh with higher clocks. Don't expect a 4.1GHz 'Kaveri'. Don't expect a new 'Kaveri' with full HSA. Don't expect a new 'Kaveri' with HBM. The only 'refresh' will be firmware changes, different steeping, and minor changes like that. A more mature process could reduce power consumption a bit.

There will be new models above Carrizo but there is absolutely nothing really new under the hood.

First you say there will be no Kaveri refresh. Then when a refresh is confirmed you change to "there will be no higher clocks". Now higher frequency updates appear to be planned.

Sometime in the second quarter of the year AMD is also expected to release new APUs for desktops, which are currently known as the “Kaveri Refresh”. The new chips will operate at higher clock-rates than the currently available “Kaveri” and will increase performance of AMD’s APUs a bit.
-http://www.kitguru.net/components/graphic-cards/anton-shilov/amd-we-will-release-new-products-starting-in-the-second-quarter-of-2015/

 

[8350rocks]

There is no refresh with higher clocks. Don't expect a 4.1GHz 'Kaveri'. Don't expect a new 'Kaveri' with full HSA. Don't expect a new 'Kaveri' with HBM. The only 'refresh' will be firmware changes, different steeping, and minor changes like that. A more mature process could reduce power consumption a bit.

There will be new models above Carrizo but there is absolutely nothing really new under the hood.

First you say there will be no Kaveri refresh. Then when a refresh is confirmed you change to "there will be no higher clocks". Now higher frequency updates appear to be planned.

Sometime in the second quarter of the year AMD is also expected to release new APUs for desktops, which are currently known as the “Kaveri Refresh”. The new chips will operate at higher clock-rates than the currently available “Kaveri” and will increase performance of AMD’s APUs a bit.
-http://www.kitguru.net/components/graphic-cards/anton-shilov/amd-we-will-release-new-products-starting-in-the-second-quarter-of-2015/

You cannot fight illogical with logic. Because illogical will just ignore your completely relevant and totally valid argument as if it does not exist.

 

[juanrga]

This is my last word about this issue.

I was the only one that provided real data for a concrete processor and the empirical fits. My point was to use the well-known cubic law to explain why AMD is reducing base clocks on Carrizo to compete with Broadwell.

You can avoid the cubic law approximation if you don't like it and you can wait to Carrizo to hit the market, then measure final silicon to get the empirical relations between V and F and between P and F.

Instead the well-known generic cubic law proportionality, you will get a more complex power series expression with lots of empirical parameters valid only for a concrete Carrizo processor (I have said "processor", not model), a concrete ambient temperature..., but the conclusion will be the same. Both the more complex expression and the simple cubic law approximation predict that efficiency improves by reducing clocks. This is why we don't see phones clocked at 4GHz, for instance, or why the more efficient Intel Xeons are clocked so low as 1.8GHz.

I have tried to explain why AMD does what it does with Carrizo. I have tried to explain why FX-8350 @4GHz is rated at 125W, but FX-9590 @4.7GHz is rated at 220W.

Any time I have tried to explain why AMD does what it does (since the epoch when I tried to explain why AMD chose bulk instead SOI) my points have been either ignored or blatantly misinterpreted.

Intel has a nice blog entry explaining the relation between power and frequency to broad audiences

https://software.intel.com/en-us/blogs/2014/02/19/why-has-cpu-frequency-ceased-to-grow

They also use the well-known V~F approx to derive the cubic law except that they write it as P ~ V^3 but evidently it is equivalent to P ~ F^3, and then write

Linear frequency growth causes power dissipation to be increasingly cubed! If the frequency is raised only twice, there will be eight times greater heat that must be accommodated or the processor will melt or shutdown.

It’s obvious that this method of increasing the frequency is not suitable for processor manufacturers because of low efficiency. However, it is used by extreme overclockers.

As said, this is my last word about this issue. People can feel free to misread me misinterpret me or just ignore my point or pretend again that the cubic law is my invention or pretend that I wrote "V=F", when I never did. I don't care. I suspect which will be the next misunderstanding of my point (hint: "without voltage changing"). I will not even care to reply that. What is really interesting is that I wrote virtually the same thoughts about why AMD is reducing frequencies (cubic law), in another forum, and my thoughts were well-received by the people therein, including the ones with internal data about Carrizo

http://semiaccurate.com/forums/showpost.php?p=227588&postcount=182
http://semiaccurate.com/forums/showpost.php?p=227607&postcount=184

As both mention my only mistake about Carrizo was in the "pipeline depth". I believed AMD would reduce the pipeline in Excavator core but it is the same inefficient pipeline that Steamroller. I hope that Zen includes a shorther pipeline.

 

[noob2222]

The problem with Intels monopolistic practices is they are targeting AMD, not ARM. The incentive is for blocking AMD tablets. Intel owns over 90% of AMD's market share so that is what mskes it illegal. This is the same bs they pulled in the 90s, paying vendors not to sell AMD products. BEEMA and MUNLLINS are better than intels crap Atom, just like in the 90s, intel is blocking amd sales.

 

[jdwii]

I find it to be a problem as well something Intel should get in trouble about they did this back in the day. Is there anyway we can do something? Thinking this should be a bit more important then it is.

I really wish we had a country that brought ethics into business. However here in the United States companies like Comcast and so on can just screw us over.