AMD Carrizo APU Details Leaked

[mikk]

http://www.bitsandchips.it/hardware...2-fino-al-2016-niente-ddr4-per-le-apu-desktop

Carrizo possibly is for mobile only.

 

[NostaSeronx]

Apparently, there is two Carrizo platforms. The "Carrizo" platform and the "Carrizo-L" platform.

I've been searching but I only know that these two exist. I have nothing to tell them apart so far.

The mentions of both platforms side by side with Intel platforms places it with;
Carrizo => Haswell/Broadwell?
Carrizo-L => Skylake

http://tw.linkedin.com/pub/%E7%9B%88%E5%A6%82-%E9%99%B3/98/673/b8

Lenovo Intel Shark Bay/Skylake,AMD Carrizo/Carrizo-L platform projects of Compal.

 

[monstercameron]

I reckon someone was bored http://www.overclock.net/t/1504122/...-features-fm3-and-20nm-afterall#post_22620345

 

[AtenRa]

I reckon someone was bored http://www.overclock.net/t/1504122/...-features-fm3-and-20nm-afterall#post_22620345

Its made for April the 1st next year.

 

[itsmydamnation]

apart from clock, number of cores, voltage and process it should be pretty close

 

[Rickyyy369]

I reckon someone was bored http://www.overclock.net/t/1504122/...-features-fm3-and-20nm-afterall#post_22620345

Reminds me of this work of art:

Its obviously not difficult to fake.

 

[NTMBK]

Apparently, there is two Carrizo platforms. The "Carrizo" platform and the "Carrizo-L" platform.

I've been searching but I only know that these two exist. I have nothing to tell them apart so far.

The mentions of both platforms side by side with Intel platforms places it with;
Carrizo => Haswell/Broadwell?
Carrizo-L => Skylake

http://tw.linkedin.com/pub/盈如-陳/98/673/b8

Perhaps Carrizo = DDR3 on FM2+, Carrizo-L = DDR4 and integrated southbridge?

 

[NostaSeronx]

Perhaps Carrizo = DDR3 on FM2+, Carrizo-L = DDR4 and integrated southbridge?

Compal does notebook platforms not desktop platforms.

Carrizo = FP4, DDR3 + Integrated Southbridge
Carrizo-L = SP2 or FP4r2, DDR4 + Integrated Southbridge

Everything I have seen is pointing at a double Carrizo launch though. One this year, August to October, for Desktop. One early next year, January to March, for Mobile. Then, a refresh of the desktop part around May to July. With a refresh of the mobile part July to September.

Carrizo launching at Broadwell, and Carrizo-L launching at Skylake.

 

[Erenhardt]

No HBM will be quite sad. It probably requires new platform.

 

[inf64]

Everything I have seen is pointing at a double Carrizo launch though. One this year, August to October, for Desktop.

So you expect AMD to let Kaveri only live for ~6 months on the market and replace it with a minor bump like Carrizo? What's the point and where is the economic logic in that?

 

[Ajay]

So you expect AMD to let Kaveri only live for ~6 months on the market and replace it with a minor bump like Carrizo? What's the point and where is the economic logic in that?

If Carrizo has a better perf/watt, it should be replaced if the design is done. It's still being manufactured on the same process node (HPM?), so cost the AMD should be about the same (+ new masks).

Has Carrizo taped out?

 

[Enigmoid]

If Carrizo has a better perf/watt, it should be replaced if the design is done. It's still being manufactured on the same process node (HPM?), so cost the AMD should be about the same (+ new masks).

Has Carrizo taped out?

On desktop there is little point in refreshing for efficiency. I doubt AMD has sold enough Kaveri chips to even think about selling Carrizo and mobile has just been released.

Personally I believe that AMD will refresh kaveri before releasing Carrizo much like they did with richland and with beema/mullins.

 

[NostaSeronx]

No HBM will be quite sad. It probably requires new platform.

High Bandiwdth Memory does not need a new platform. It just needs enough area within the package to fit a HBM Stack or a couple HBM stacks.

So you expect AMD to let Kaveri only live for ~6 months on the market and replace it with a minor bump like Carrizo? What's the point and where is the economic logic in that?

AMD releases new parts for the desktop platform around every 9 months. If Excavator launches beyond that point. Then, Excavator is not by absolute definition a 15h architecture.

Trinity -> Richland -> Kaveri
Each transition on the desktop side took only up to 9 months.

While for Llano(1st gen SKUs) -> Trinity(1st gen SKUs)
Took 15 months; 2nd gen Llano SKUs to 1st gen Trinity SKUs only took 9 months.

 

[itsmydamnation]

i guess its possible that carrizo isn't late,
kaveri is and thats why we have richland that would create an issue where carrizo is ready mid kaveri life cycle.

if we assume something like 1 year release schedule
late 2011 bulldozer ,
late 2012 pile-driver
late 2013 (non exsistant steamroller A)
late 2014 excavator

 

[Enigmoid]

High Bandiwdth Memory does not need a new platform. It just needs enough area within the package to fit a HBM Stack or a couple HBM stacks.AMD releases new parts for the desktop platform around every 9 months. If Excavator launches beyond that point. Then, Excavator is not by absolute definition a 15h architecture.

Trinity -> Richland -> Kaveri
Each transition on the desktop side took only up to 9 months.

While for Llano(1st gen SKUs) -> Trinity(1st gen SKUs)
Took 15 months; 2nd gen Llano SKUs to 1st gen Trinity SKUs only took 9 months.

Trinity to richland were the same die (basically) with better manufacturing and binning. Much like the 4790k is to the 4770k. It is absolutely not a "new" platform.

I'm not sure what happened with trinity (llano inventory) but AT had their review for mobile trinity on May 15, 2012, just under a year from their llano review.

 

[NostaSeronx]

Trinity to richland were the same die (basically) with better manufacturing and binning.

Trinity and Richland were not the same die. Richland had a couple layers changed, it looks the same but it isn't the same.

Richland;
- Hybrid Boost (17 Temperature Sensors, to evaluate boosts in both CPU and GPU)
- cTDP on certain platforms
- Reduced BAPM latency (Faster P-State switching)
- Etc

TN-A1 = Base Model - 0 / Stepping - 1
RL-A1 = Base Model - 3 / Stepping - 1

The differences between Trinity and Richland are within 10h-1Fh specification. While Kaveri's and Carrizo's are within separate specifications.

Kaveri - Family 15h Models 30h-3Fh Processors.
Carrizo - Family 15h Models 60h-6Fh Processors.

Where Richland is from the same model tree as Trinity. Carrizo is not from the same model tree as Kaveri.

You can stretch Richland to being the same or near the same as Trinity. The same can not be said between Kaveri and Carrizo.

 

[pTmdfx]

High Bandiwdth Memory does not need a new platform. It just needs enough area within the package to fit a HBM Stack or a couple HBM stacks.AMD releases new parts for the desktop platform around every 9 months.

Of course it needs a new socket. Otherwise what pins would supply the power for it?

 

[NostaSeronx]

Of course it needs a new socket. Otherwise what pins would supply the power for it?

The same pins that send power through the Integrated Voltage Regulation.

 

[sefsefsefsef]

Of course it needs a new socket. Otherwise what pins would supply the power for it?

Considering an entire DIMM of DDRx memory only adds ~5W to a system (including 16 DDRx chips, and high powered drivers for communicating over long motherboard traces), and that HBM is much more power efficient than DDRx, HBM will not contribute significantly to the power requirement of a socket. Similarly, it will not contribute significantly to the cooling demands of the socket.

 

[Enigmoid]

Trinity and Richland were not the same die. Richland had a couple layers changed, it looks the same but it isn't the same.

Richland;
- Hybrid Boost (17 Temperature Sensors, to evaluate boosts in both CPU and GPU)
- cTDP on certain platforms
- Reduced BAPM latency (Faster P-State switching)
- Etc

TN-A1 = Base Model - 0 / Stepping - 1
RL-A1 = Base Model - 3 / Stepping - 1

The differences between Trinity and Richland are within 10h-1Fh specification. While Kaveri's and Carrizo's are within separate specifications.

Kaveri - Family 15h Models 30h-3Fh Processors.
Carrizo - Family 15h Models 60h-6Fh Processors.

Where Richland is from the same model tree as Trinity. Carrizo is not from the same model tree as Kaveri.

You can stretch Richland to being the same or near the same as Trinity. The same can not be said between Kaveri and Carrizo.

Sorry I meant completely similar, the actual logic layout is virtually identically except richland has better power delivery, etc. Similar to beema/mullins and kabini/temash.
Either way extremely small differences as a whole and no performance/core logic changes.

 

[Lepton87]

Trinity to richland were the same die (basically) with better manufacturing and binning. Much like the 4790k is to the 4770k. It is absolutely not a "new" platform.

Not really, the only similarity is that the difference is minor. 4790K has exactly the same silicon as 4770k but has an improved package while Richland is a different silicon stepping then Trinity while retaining the same package.

 

[Enigmoid]

Not really, the only similarity is that the difference is minor. 4790K has exactly the same silicon as 4770k but has an improved package while Richland is a different silicon stepping then Trinity while retaining the same package.

I'm not sure how AMD and intel define it but trinity and richland have the same stepping, different revisions.

Both are stepping 1.

 

[Lepton87]

I'm not sure how AMD and intel define it but trinity and richland have the same stepping, different revisions.

Both are stepping 1.

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

 

[VirtualLarry]

Richland;
- Hybrid Boost (17 Temperature Sensors, to evaluate boosts in both CPU and GPU)

And no program can currently properly report temps for that CPU! Seriously, WTF AMD...

 

[Ajay]

And no program can currently properly report temps for that CPU! Seriously, WTF AMD...

These sensors may not be addressable outside of the APU (though that seems unlikely) or AMD cares not to publish the means to access them.