AMD Carrizo APU Details Leaked

[ikachu]

I must have confused stepping with revision, what's the difference anyway? My point was that there were some changes in the silicon itself.

Trinity and Richland are actually the same chip, with some process tweaks at GF and updated firmware. The model was updated (in the fuse recipe) from 0 to 3 for firmware identification purposes.

 

[Enigmoid]

I must have confused stepping with revision, what's the difference anyway? My point was that there were some changes in the silicon itself.

Trinity and Richland are actually the same chip, with some process tweaks at GF and updated firmware. The model was updated (in the fuse recipe) from 0 to 3 for firmware identification purposes.

I always thought stepping was referring to the masks. Same stepping means same masks. I'm not sure though.

 

[ikachu]

I always thought stepping was referring to the masks. Same stepping means same masks. I'm not sure though.

Yeah the 'stepping' referred to in those screenshots corresponds to a metal layer spin of the processor with an unchanged base layer.

The 'model' should correspond to an all layer spin, but in the case of Richland, it was just used to differentiate the two CPU types.

 

[NostaSeronx]

I'm gonna post this here, ciao!

Market Demand

• Now mainstream processors within smart phone and tablet focus more on energy efficiency than performance.
• 28nm FD-SOI promises good energy efficiency as well as low die cost.

Higher Performance

• 28nm FD-SOI can achieve higher performance to meet the demand of current mainstream GHz processors.
• 28nm FD-SOI allows better dynamic performance/leakage tradeoff by body-bias.

Smaller Area

• 28nm FD‐SOI can save 6-20% area compared with other 28nm technologies.
• Smaller area enables higher chip density and lower die cost.

Lower Power

• 28nm FD‐SOI consumes much lower leakage power while achieving the same performance as 28nm HPM technology.
• 28nm FD‐SOI consumes lower dynamic power comparing with 28nm SLP technology at similar frequency.

- VeriSilicon Microelectronics
- Synapse Design
- SoCtronics

^-- All have vouched for FDSOI.

----
Up till November 2012, Carrizo/Excavator was to be on 20-nm LPM. Something happened in the year 2012; http://i.imgur.com/TBYwc5j.jpg // http://www.st.com/st-web-ui/active/en/press/jp/c2680

 

[Ajay]

I'm gonna post this here, ciao!- VeriSilicon Microelectronics
- Synapse Design
- SoCtronics

^-- All have vouched for FDSOI.

----
Up till November 2012, Carrizo/Excavator was to be on 20-nm LPM. Something happened in the year 2012; http://i.imgur.com/TBYwc5j.jpg // http://www.st.com/st-web-ui/active/en/press/jp/c2680

Is there *any* evidence that GFL has developed a 28nm SOI process? If it is going to be used for Carrizo, wouldn't it need to be ready for risk production already?

 

[Homeles]

As much as I'd love to believe that AMD's back on SOI, I have no choice but to wait until the proof pudding is at my doorstep.

 

[NostaSeronx]

Is there *any* evidence that GFL has developed a 28nm SOI process? If it is going to be used for Carrizo, wouldn't it need to be ready for risk production already?

The lack of mention of SOI does not mean that SOI is not on the plate.

AMD profiles with exact wording to lax wording of 28SOI;
http://www.linkedin.com/pub/john-chan/a3/178/348
http://www.linkedin.com/in/parasshah55
http://www.linkedin.com/pub/victor-f-andrade/b/822/814
http://www.linkedin.com/in/jeffcorrell

There is a lack of profiles at GlobalFoundries/Samsung regarding FDSOI.

STMicroelectronics/GlobalFoundries ramped up in 1H 2014, source: STMicroelectronics. Samsung will be ramping up in 1H 2015, source: Samsung.

I expect the actual ramp and products release won't happen till Samsung ramps up.

Off-topic;
RISC-V is a 28-nm FDSOI design win.

From left to right: a dual-core Rocket+Hwacha system in IBM’s 45nm SOI process, running up to 1.35GHz, a single core Rocket+Hwacha system in ST 28nm FDSOI process running down to 0.45V, and an FPGA prototype running on a Xilinx Zybo board.

https://blog.riscv.org/wp-content/uploads/2014/09/photo2.jpg (Middle one)

FDSOI is in some level of production.

 

[Ajay]

STMicroelectronics/GlobalFoundries ramped up in 1H 2014, source: STMicroelectronics. Samsung will be ramping up in 1H 2015, source: Samsung.

I expect the actual ramp and products release won't happen till Samsung ramps up.

Off-topic;
RISC-V is a 28-nm FDSOI design win.
https://blog.riscv.org/wp-content/uploads/2014/09/photo2.jpg (Middle one)

FDSOI is in some level of production.

Ramped up would mean in production. When I see production volumes of some 28nm SOI part - I will believe. A quick search finds recent results for deal between STM & Samsung, but all SOI news about GFL was old.

 

[Abwx]

wt
http://www.hardware.fr/articles/880-17/moyennes.html

 

[Fjodor2001]

http://www.digitimes.com/news/a20141007PD202.html

AMD is planning to announce next-generation Carrizo APUs in March 2015 to replace its existing Kaveri APUs for the mainstream performance notebook segment and will release Carrizo-L APUs for the entry-level notebook segment in December 2014 at the earliest to challenge Intel's Pentium and Celeron processors, according to sources from notebook players.


AMD's Carrizo processors will feature quad-core architecture and adopt a 28nm process and Excavator Core. The APUs will also support DDR3-2133 memory, Windows 10/Windows 8.1, Ubuntu and SLED operating systems, the sources noted.
Carrizo-L APUs will replace AMD's existing Beema and Mullins APUs for entry-level notebooks and tablets, the sources added.

So Carrizo will go up against Broadwell already this year. Interesting... I wonder how much improvement we can expect though, since AMD is abandonning the Bulldozer based uArch in favour of their next-gen x86 uArch Zen that should be coming soon after that...

 

[NostaSeronx]

Ramped up would mean in production. When I see production volumes of some 28nm SOI part - I will believe. A quick search finds recent results for deal between STM & Samsung, but all SOI news about GFL was old.

Update from April 23, 2014.

STMicroelectronics has become the FD-SOI promoter. Last year, STM opened its
FD-SOI library to GF and it has signed a foundry deal. STM had already signed three
contracts for its 28nm FD-SOI when it held its analyst day last May. We understand
that Sony [6758 JP] was one of the clients for its consumer products. We also
understand that GF had signed a major contract with one of its fabless clients for its
14nm FD-SOI.

STM has now signed a total of 15 design wins in FD-SOI including multiple design
wins in ASICs for networking and consumer markets. The group was very vocal
during its FY results presentation about the FD-SOI revenue growth prospects.
According to STM’s management, the 15 wins are complex, high-volume, high-ASP
products and there are lots more in the pipeline. The group said that they are major
wins with major customers, and that they fall into two main categories: network
infrastructure and consumer.

Then the next one is;

Volume production of 28nm FD-SOI will be ramping up in 4Q this year. STM will also
soon be ready with 14nm prototypes, and Mr Cherry cited the groundwork laid by the
alliance in Albany (with IBM, Leti, GF and Samsung).

28-nm FDSOI is ramping at Samsung, GlobalFoundries, and STMicroelectronics in Q4 2014. Much like how 14-nm LPE/LPP is ramping at both Samsung/GlobalFoundries at the same time

Full volume isn't expected till the end of Q1 2015.

 

[Abwx]

http://www.digitimes.com/news/a20141007PD202.html

So Carrizo will go up against Broadwell already this year. Interesting... I wonder how much improvement we can expect though, since AMD is abandonning the Bulldozer based uArch in favour of their next-gen x86 uArch Zen that should be coming soon after that...

30% better perf at 15W TDP is all that is known more or less officialy.

 

[mrmt]

Ramped up would mean in production. When I see production volumes of some 28nm SOI part - I will believe. A quick search finds recent results for deal between STM & Samsung, but all SOI news about GFL was old.

I think SOI for the bleeding edge applications died with the demise of IBM fab operation, the leaders of the pack are all betting on finfet/bulk for their bleeding edge nodes. SOITEC itself, even with Samsung signing an agreement with STM for 28nm FDSOI, is expecting its volumes to stay constant YoY, and those volumes are already very, very low. It will go above the paltry 5.6 million they are selling today, but not by much.

Plus there is the question of STM itself. The company is struggling for some years and is scalling back R&D (fell ~35% YoY). I don't think they will be able to afford to develop nodes at the speed and quality of Samsung, Intel and TSMC.

 

[NostaSeronx]

http://forums.anandtech.com/showpost.php?p=36799107&postcount=361

Just an update to the above post.

28-nm and 14-nm FDSOI will be ramping up to volume in Q4 2014, not just 28-nm FDSOI. Samsung will be ramping up to volume with the same processes in Q1 2015.

JULY 10, 2014;

GlobalFoundries committed to using ST’s FD-SOI tech for 28nm and 20nm production in 2012, and expects to put it into volume production by the end of this year, for 28nm and 14nm processes. In May, STMicroelectronics announced that Samsung would use ST’s 28nm FD-SOI tech for foundry customers. Samsung plans to offer the process in early 2015.

We know Samsung is committed to 14nm FDSOI as well do to it being on the roadmap; http://i.imgur.com/TBYwc5j.jpg

Comparison between the 20-nm nodes;
90/86nm CPP / 64nm M1 = 20-nm Bulk Planar
84nm CPP / 64nm M1= 14-nm FDSOI Planar
84/78nm CPP / 64nm M1 = 14-nm Bulk FinFET

What is introduced over 28-nm FDSOI;

New process elements compared to the 28nm FD-SOI technology developed in this project include a strained-SiGe channel and dual source-drain SiC: P and SiGe:B in-situ doped epi-taxies for performance boost.

In addition, this technology features 84nm contacted poly-pitch, new contact interconnect constructs, and 64nm metal 1 pitch, providing 0.55x area scaling over the 28nm technology.

Strained SiGe-on-insulator <3

If and when Excavator is FDSOI; Excavator (28-nm FDSOI) -> Excavator+ (14-nm FDSOI) -> Zen (14-nm FDSOI) -> Zen+ (10-nm FDSOI)?

 

[Vesku]

I could see them partnering with STM to convert GF 32 and current 28nm lines to 28nm SOI, STM having blazed the development trail already. Would be a long life SoC node. After all, it will probably be 2-3 years before FinFet transistor costs drop down to what the budget SoCs can handle. We'll find out when Carrizo is revealed, AMD would surely want to talk it up if they are using an improved version of 28nm.

 

[Justinbaileyman]

30% better perf at 15W TDP is all that is known more or less officialy.

So will this be a drop in for FM2+ sockets or will there be a new socket and chipset? I really hope that 30% increase is true clock for clock and not just AMD pulling one of there lowering TDP cop-out crap like they did with the FX-8320e and FX-8370e. I would really love to have a new socket so we could use DDR4 which should bump up the IGPU performance abit for sure.

 

[NostaSeronx]

So will this be a drop in for FM2+ sockets or will there be a new socket and chipset?

Kaveri 4C/8CU Mobile -> FP3
Carrizo 4C/8CU Mobile -> FP4

Kaveri 4C/8CU Desktop -> FM2+
Carrizo 4C/8CU Desktop -> FM2+

Kaveri 4C/8CU Server/Embedded -> FP3
Carrizo 4C/8CU Server/Embedded -> SP2

I really hope that 30% increase is true clock for clock and not just AMD pulling one of there lowering TDP cop-out crap like they did with the FX-8320e and FX-8370e.

"Greater Performance" for AMD always outcomes "Higher IPC"

Greyhound+(Stars) -> Bulldozer(Piledriver) = "Greater Performance", as one great ex-AMD employee once said.

Steamroller -> Excavator = "Greater Performance", as one great AMD roadmap stated.

 

[cbn]

I wonder if anyone will make a mainboard just using the integrated ones. Would be a killer HTPC chip. Especially if it has a H.265 decoder built-in.

That would be a good idea, especially with FM2+ traditionally lacking on budget side of things for Mini-ITX.

 

[cbn]

So will this be a drop in for FM2+ sockets or will there be a new socket and chipset? I really hope that 30% increase is true clock for clock and not just AMD pulling one of there lowering TDP cop-out crap like they did with the FX-8320e and FX-8370e. I would really love to have a new socket so we could use DDR4 which should bump up the IGPU performance abit for sure.

That diagram posted above your response points the platform being BGA.

 

[NostaSeronx]

I did some research and if Excavator is using High Density Libraries. It is impossible to get higher performance with a node lesser than 28-nm SHP.

The only three nodes, I consider to be better than 28-nm SHP;
28-nm FDSOI
20-nm Low Power Manufacturing
14-nm FDSOI

AMD has stated they achieved 30% area density and 30% lower total power consumption. The performance loss is not linear, my estimations are ~40%.

Excavator with a "bulk" node less than TSMC 28-nm HP and GlobalFoundries 28-nm SHP. Would need to see a performance increase architecturally if not frequency of....

~40%(high density cell libraries) + ~15%(lesser node) + ~30%(slide says Excavator performs this much better).

That is a total of an ~85% barrier wall of performance to improve. That points to Excavator incapable of being faster with bulk.

With 28-nm FDSOI;
~40%(high density cell libraries) - 40%(high performance(FBB+Poly Biasing) 28nm FDSOI node(1.2V-1.3V Vdd)) + ~30%(slide says Excavator performs this much better).

Bringing down the architectural improvement to just 30%.

With FDSOI there is an area shrink and a leakage drop at high performance-ended designs. The lower leakage would allow the designs to have higher frequencies at lower TDPs. FDSOI doesn't need a large architectural improvement.

Also, for the IVR, 28-nm FDSOI supports the same Switched Capacitors(iVRMs) from POWER8. You want to know why that is important? Well if you check IBM's research on switch capacitors you will find it is closed-loop.

Dynamic voltage and frequency scaling (DVFS) techniques&#8212;along with associated techniques
such as dynamic voltage scaling (DVS) and adaptive voltage and frequency scaling (AVFS)&#8212;
are very effective in reducing power, since lowering the voltage has a squared effect on active
power consumption. DVFS techniques provide ways to reduce power consumption of chips on
the fly by scaling down the voltage (and frequency) based on the targeted performance
requirements of the application. Since DVFS optimizes both the frequency and the voltage, it
is one of the only techniques that is highly effective on both dynamic and static power.
Dynamic voltage scaling is a subset of DVFS that dynamically scales down the voltage (only)
based on the performance requirements. Adaptive voltage and frequency scaling is an
extension of DVFS. In DVFS, the voltage levels of the targeted power domains are scaled in
fixed discrete voltage steps. Frequency-based voltage tables typically determine the voltage
levels. It is an open-loop (no feedback) system with large margins built in, and therefore the
power reduction is not optimal. On the other hand, AVFS deploys closed-loop voltage scaling
and is compensated for variations in temperature, process, and V=IR drop using dedicated
circuitry (typically analog in nature) that constantly monitors performance and provides active
feedback. Although the control is more complex, the payoff in terms of power reduction is
higher.

Analog in FDSOI(& PDSOI?) is superior than analog in bulk and FinFETs.

That IVR is positioned pretty close to;

GlobalFoundries committed to using ST&#8217;s FD-SOI tech for 28nm and 20nm production in 2012, and expects to put it into volume production by the end of this year, for 28nm and 14nm processes. In May, STMicroelectronics announced that Samsung would use ST&#8217;s 28nm FD-SOI tech for foundry customers. Samsung plans to offer the process in early 2015.

 

[Justinbaileyman]

Sorry guys my example was a poor one I should have used 6800K vs. the 7850K where there really are no improvements between the two.If there are its very minute. Not to mention the 6800K could be overclocked well past the 4.4-4.5Ghz mark more closely to 4.8-5.2Ghz where as the 7850K fails to get past 4.5Ghz most the time.Clock for clock I dont see any differences though. I have tested in Cinebench ,Winrar ,Handbreak , and a slew of other games and I see no improvements in bench scores between the two APU's/CPU's. I am just hoping this Carrizo isn't another 6800K like the 7850K is. So when are we expected to see them for sale?? Hopefully this year yet or is that asking to much??

 

[Paul98]

I am interested to see what performance boost it gets from color compression.

 

[MisterLilBig]

Wouldn't the color compression performance boost be similar or above what the R9 285 delivers, percentage wise of course?

 

[NostaSeronx]

Wouldn't the color compression performance boost be similar or above what the R9 285 delivers, percentage wise of course?

It could be bigger..

Since, Kaveri to Carrizo is 128-bit to 128-bit. While, Tahiti to Tonga is 384-bit to 256-bit.

The scaling for similar sized memory interfaces has not been seen yet.

 

[Brekyrself]

All this talk about Carrizo, which is great, but where are all the kaveri laptops?!