Kaveri performance

[NTMBK]

Berlin has been explicitly stated to be fabbed at TSMC. Not on a different process, not on a different node offering, but on the same process with the same node offering as Jaguar.

Where? I have not seen this, anywhere. Please post a link- I will happily admit I am wrong, if Berlin is on TSMC (given that it's just a rebadged Kaveri as far as I can tell).

 

[frozentundra123456]

Intel rebates/bribery/etc. Do we really need to go through this one AGAIN frozentundra or do you own any other arguments?

It was a rhetorical question. Obviously CPU performance and power usage DOES matter. And no we don't need to hear your conspiracy theories again.

 

[NostaSeronx]

Richland would have had to have been in production at the point that you posted in order for it to come out in June 2013 (and some OEM parts launched even earlier, if I remember correctly). Arguing "gap of information" is nonsense. If you had inside information about what was being made where, then you would have known that Richland was coming out on 32nm, and that it was just a respin of Trinity as opposed to a completely new part (which a move to 28nm would imply).

I didn't have any secret information about Richland pre-2013. I don't have any secret information for anything AMD either. It is all public domain if you know how to use aggressive search engine techniques. If you can't get it that way use a brute force search engine for more power.

Where? I have not seen this, anywhere. Please post a link- I will happily admit I am wrong, if Berlin is on TSMC (given that it's just a rebadged Kaveri as far as I can tell).

Can we leave it to this guy(TPM):

AMD used Taiwan Semiconductor Manufacturing Corp to fab its Kyoto Opteron X-Series processors using its 28-nanometer processes, and the kicker Berlin Opteron X chips will also come out of the same wafer baker and use the same 28nm processes.

http://www.theregister.co.uk/2013/06/18/amd_opteron_arm_server_chips/

 

[mrmt]

*sarcasm* So, you are saying my confidential info from TSMC & AMD backstabbing GlobalFoundries is wrong? */sarcasm*

Nope, I'm saying you have no confidential info of anything, that you are pulling this stuff out of thin air. The second WSA amendment is dead and there would be time to move stuff up to Globalfoundries from november last year to november this year. And how could they delay something in a mature node like TSMC's 28nm?

 

[NTMBK]

I didn't have any secret information about Richland pre-2013. I don't have any secret information for anything AMD either. It is all public domain if you know how to use aggressive search engine techniques. If you can't get it that way use a brute force search engine for more power.

Well it looks like your aggressive search engine techniques gave you dodgy information based on forum speculation.

Can we leave it to this guy(TPM):
http://www.theregister.co.uk/2013/06/18/amd_opteron_arm_server_chips/

That looks like the Register misreporting to me, as opposed to an actual release by AMD. Why have no other news sites said that?

 

[NostaSeronx]

How could they delay something in a mature node like TSMC's 28nm?

Who said the delay was for the node?

Well it looks like your aggressive search engine techniques gave you dodgy information based on forum speculation.

Filetypes are pdf, doc, xls, md. I don't use forums for aggregate information.

That looks like the Register misreporting to me, as opposed to an actual release by AMD. Why have no other news sites said that?

TPM is no TV, :thumbsup:

 

[NTMBK]

Filetypes are pdf, doc, xls, md. I don't use forums for aggregate information.

They still gave you bad information. But if these sources are publicly available, why not post links to them? I'm always interested in the stuff that a deep-dive can dig up.

TPM is no TV, :thumbsup:

I'm still unconvinced. A direct quote from AMD stating that Trinity's 28nm successor is at Globalfoundries, as opposed to a single line in a Reg post stating something which no other news outlet (reporting on the same press release) has said. That is a very long way from a confirmation.

 

[NostaSeronx]

I'm still unconvinced. A direct quote from AMD stating that Trinity's 28nm successor is at Globalfoundries, as opposed to a single line in a Reg post stating something which no other news outlet (reporting on the same press release) has said. That is a very long way from a confirmation.

Would Morgan Stanley Research be more convincing?

 

[mrmt]

Who said the delay was for the node?Filetypes are pdf, doc, xls, md. I don't use forums for aggregate information.TPM is no TV, :thumbsup:

As I said, back it up. Show us what your aggressive search techniques can do. Right now it's just hot air.

Would Morgan Stanley Research be more convincing?

Oh yes, it would. Link?

Ed: Seifert said that Kaveri would be built on GLF 28nm bulk at Morgan Stanley Technology, Media and Telecom Conference in 2012. Oh, the irony.

 

[NTMBK]

Would Morgan Stanley Research be more convincing?

Yup, sounds good, point us towards it. :thumbsup:

 

[NostaSeronx]

Source: http://www.morganstanleychina.com/conferences/apsummit2011/research/63ICManufacturingFoundries.pdf

It doesn't out right say it but it gives you reasons why and I'll help with that. It is before AMD cancelled 28-nm APU at GlobalFoundries as well.

We see several significant shifts happening: 1.) TSMC
continues to take share at 28nm. We would not be
surprised if AMD shifts some of its “Fusion” chips to TSMC
and the win with Xilinx is well documented.

We would expect it to lose some market share to TSMC, and
recent yield issues with AMD would suggest to us that AMD
could move its Fusion products to TSMC.

TSMC could gain additional share at 28nm. Given recent
announcements by AMD that its foundry partner GF can’t
deliver enough chips and has yield issues, our expectation is
that AMD likely transfers some of its production to TSMC at
28nm. Our view is the CPU likely stays at GlobalFoundries,
but the Fusion product line, which combines the CPU and
GPU, likely shifts to TSMC. We note that TSMC has
experience with GPU and on the Fusion the GPU takes up
more real estate than the CPU and thus the GPU expertise is
more critical.

We believe the yield issues give TSMC an opportunity to
produce some of AMD’s processors at 28nm in 2012.
Specifically, AMD’s Fusion product line, which combines the
graphics chip with the processor, seems ideally suited for
TSMC since TSMC has strong knowledge of the graphics
process. We would not be surprised if TSMC receives all of a
significant portion of AMD’s Fusion products going into 2012,
although we do expect that GF would take back the share
when/if its yield improves, given the historical relationship
between AMD and GF and given GF’s lower pricing versus
TSMC.

With the bulk of quotes gone: Nvidia Kepler and AMD Jaguar were both done on TSMC's HP offering. AMD Tahiti, Pitcarn, Cape Verde, Bonaire is up in the air for HPL or HPM or HP.

Image #1: (July 2013?)
http://i.imgur.com/kF1Ocon.jpg

GloFo's HPP which is the competitor of TSMC's HP wasn't finished in 2012.

http://ir.amd.com/phoenix.zhtml?c=74093&p=quarterlyearnings
2012 Q4 QE, Transcript.

Rory Read:

We introduced last year the first 28 nanometer graphics products and leadership. And now, as we move forward, we are positioning ourselves with
that refocus, better-executing supply chain to move forward in 28.

We had a really awesome 28-nm GPU supply chain and we are refocusing all our efforts towards that supply chain. While, we wait for GlobalFoundries to catch up.

Here comes the damage control

Lisa Su:

-------, just to your question about 28 nanometer and GLOBALFOUNDRIES, we are pleased with our overall 28 nanometer bring-up in all of our
founders. So from a 28 nanometer standpoint, we feel very good about where the technology is and what it is delivering for us from a product
standpoint.

Thank Goodness, for TSMC or we would have had to wait till 2015!

 

[mrmt]

It doesn't out right say it but it gives you reasons why and I'll help with that. It is before AMD cancelled 28-nm APU at GlobalFoundries as well.

The first document is from *before* Seifert statement. Document is from 2011, Seifert mentioned 28nm at GLF in 2012.

In the second document, more hot air. What you have is Lisa Su and Rory Read tap dancing around the issue. He didn't disclose whether they are going for Globalfoundries for 28nm or whether they are ditching the process. In fact, that sounds more like they cannot disclose whether they will give a go/no-go for this product with the then-current state of the process.

So what we have here is:

2011 ====> Morgan Stanley stating a possibility that Kaveri might end up in Globalfoundries.
2012 ====> Seifert saying that Kaveri will be at Globalfoundries
Q4 2012 => Rory and Lisa tap dancing around the issue.

I don't think your spider sense has credibility enough to tip the balance here.

Btw, what happened with your IBM 20nm process?

 

[NostaSeronx]

Btw, what happened with your IBM 20nm process?

Supposedly there is two nodes:
22-nm PDSOI and 20-nm ETSOI

Common Platform SOI Division
High Performance Extremely Thin SOI (ETSOI) Hybrid CMOS with Si Channel NFET and Strained SiGe Channel PFET, IBM, STMicroelectronics, GLOBALFOUNDRIES, Renesas, Soitec, CEA-LETI

IBM
22nm High-Performance SOI Technology Featuring Dual-Embedded Stressors, Epi-Plate High-K Deep-Trench Embedded DRAM and Self-Aligned Via 15LM BEOL, IBM Semiconductor Research and Development Center

I some how mixed them together without realizing these documents existed.

IBM Next-Power will be on it.
AMD BigAPU and BigCPU will be on it as well.

The BigAPU isn't the PS4/XB1 chip if you were wondering.

http://i.imgur.com/IZSQf5S.jpg <-- BigAPU

This 22-nm FD-SOI uses HDL as well.
CPP - ~9x nm
M1 - 80-nm
assuming here on after:
to M11 80-nm
M12 - ~500-nm
M13 - ~800-nm
M14 - ~800-nm
M15 - ~800-nm
^-- merely making the assumption based off TSMC's HDL.

This node is already in some form of state allowing for IBM and AMD to test their chips at Fab 1.

^--All supposedly.

---
http://semiaccurate.com/forums/showpost.php?p=190352&postcount=107

Metal 1-5 Pitch: 80nm
6-7: 144 nm
8-10: 288nm
11-13:640nm
14-15: 2400nm

Past M6 my estimation deviates. M14-15 are way off.

 

[NTMBK]

Source: http://www.morganstanleychina.com/conferences/apsummit2011/research/63ICManufacturingFoundries.pdf

It doesn't out right say it but it gives you reasons why and I'll help with that. It is before AMD cancelled 28-nm APU at GlobalFoundries as well.

With the bulk of quotes gone: Nvidia Kepler and AMD Jaguar were both done on TSMC's HP offering. AMD Tahiti, Pitcarn, Cape Verde, Bonaire is up in the air for HPL or HPM or HP.

Image #1: (July 2013?)
http://i.imgur.com/kF1Ocon.jpg

GloFo's HPP which is the competitor of TSMC's HP wasn't finished in 2012.

http://ir.amd.com/phoenix.zhtml?c=74093&p=quarterlyearnings
2012 Q4 QE, Transcript.

Rory Read:We had a really awesome 28-nm GPU supply chain and we are refocusing all our efforts towards that supply chain. While, we wait for GlobalFoundries to catch up.

Here comes the damage control

Lisa Su:Thank Goodness, for TSMC or we would have had to wait till 2015!

All that Morgan Stanley article is, is speculation. They had some good reasons... in 2011. GloFo was certainly not ready for 28nm in 2011, as the canning of Wichita (28nm Bobcat shrink) proved, and AMD moved Jaguar to TSMC as a result. But the important part is at the end of what you yourself quoted:

We would not be surprised if TSMC receives all of a
significant portion of AMD&#8217;s Fusion products going into 2012,
although we do expect that GF would take back the share
when/if its yield improves, given the historical relationship
between AMD and GF and given GF&#8217;s lower pricing versus
TSMC.

Kaveri is going to be a 2014 part- I think we can (hopefully) say that GloFo finally has their 28nm process working properly by that point. Even Morgan Stanley didn't think that AMD would move everything to TSMC indefinitely.

And yes, those quotes from the call sound like damage control- but given how long it has taken GloFo to get 28nm working (and how much it has delayed Kaveri) I am not surprised! But it certainly doesn't say that they're moving everything to TSMC.

20nm ETSOI sounds promising- lets hope that GloFo can get it working a bit faster than they did with 28nm. (There was a quote somewhere saying AMD were skipping 22nm and going straight to 20nm... I'll see if I can dig it up.)

EDIT: Oh, and that "big APU" image is a fake. It's been known to be a fake for a long time. It was a mockup created by a S|A forum poster, and that (stupid) news sites mistook for a leak.

 

[NostaSeronx]

EDIT: Oh, and that "big APU" image is a fake. It's been known to be a fake for a long time. It was a mockup created by a S|A forum poster, and that (stupid) news sites mistook for a leak.

It was only for reference there is a BigAPU product. Like Richland, there is no info on it except for the socket which has 1.5x more pins than G34. For all I know it could be ARMv8 with GCN. It could also be cancelled with the "Boulder" and "Austin" platforms.

 

[raghu78]

if Kaveri was being made on TSMC 28nm why the the massive delays. TSMC 28nm HP started volume production in Oct 2011. TSMC has primarily made 28nm ARM SOCs and GPUs with their HP process. these chips do not require high performance transistors at 4+ Ghz speeds. TSMC has not manufactured high performance microprocessors at 4+ Ghz speeds ever. Intel and IBM are only the companies who have the processes suitable for high performance and only Intel with bulk technology. IBM used SOI to provide similar transistor performance as Intel's bulk at same node.

Globalfoundries 32nm SOI has shown thats its capable of 4.5 - 5 Ghz speeds. But with 28nm bulk Globalfoundries might be struggling to get the same transistor performance as 32nm SOI. Intel is the only company which has 4.5 Ghz speeds on bulk processes, and even they are on FINFET at 22nm.

Kaveri/Berlin at TSMC seems far fetched because it essentially means AMD's entire 28nm desktop and notebook APU production in 2014 will be at TSMC, with only server CPUs on GF 32nm SOI.

AMD has wafer volume commitments with GF which cannot be met with server CPU volumes which are very very small. even PS4 and Xbox One APUs cannot make up for the loss of AMD's entire APU production. Also its unlikely that TSMC has a process capable of delivering 4+ Ghz speeds.

 

[mrmt]

TSMC has not manufactured high performance microprocessors at 4+ Ghz speeds ever. Intel and IBM are only the companies who have the processes suitable for high performance and only Intel with bulk technology. IBM used SOI to provide similar transistor performance as Intel's bulk at same node.

It's funny to read about TSMC lack of ability to raise clocks or not being able to manufacture AMD's punny APUs when in reality TSMC manufactures a 478mm^2 200W+ TDP 3.6GHz monster using their 28nm process. Yes, that's SUN SPARC T5, a processor that dwarfs *anything* that AMD has manufactured to date in both performance and complexity. Clearly some people don't know AMD's place in the food chain.

 

[Exophase]

I don't know about 4GHz but TSMC does manufacture SPARC processors, currently T5 which clocks up to 3.6GHz. We don't really know if competing nodes have a clock advantage or if it just so happens that no one wants to do 4+GHz processors at TSMC because that's such a niche.

Same story with GF bulk performance, just because someone hasn't been done yet doesn't mean it can't be done - until recently GF wasn't making anything in bulk at all.

 

[raghu78]

It's funny to read about TSMC lack of ability to raise clocks or not being able to manufacture AMD's punny APUs when in reality TSMC manufactures a 478mm^2 200W+ TDP 3.6GHz monster using their 28nm process. Yes, that's SUN SPARC T5, a processor that dwarfs *anything* that AMD has manufactured to date in both performance and complexity. Clearly some people don't know AMD's place in the food chain.

its not just the size of the chip. its whether the desired clocks can be hit. Richland hits 4.4 Ghz with max Turbo. with voltage overclocking it can go upto 4.6 - 4.8 ghz. Kaveri needs to atleast be able to do similar clocks.
if Kaveri has a 20% higher IPC and 20% lower max clock compared to Richland , net CPU perf gain would be very close to nothing.

http://hexus.net/tech/reviews/cpu/56657-amd-a10-6800k-32nm-richland/?page=10

http://techreport.com/review/24954/amd-a10-6800k-and-a10-6700-richland-apus-reviewed/11

can TSMC 28nm HP deliver those speeds or switching frequencies ?

 

[mrmt]

its not just the size of the chip. its whether the desired clocks can be hit. Richland hits 4.4 Ghz with max Turbo. with voltage overclocking it can go upto 4.6 - 4.8 ghz. Kaveri needs to atleast be able to do similar clocks.
if Kaveri has a 20% higher IPC and 20% lower max clock compared to Richland , net CPU perf gain would be very close to nothing.

So do you think TSMC can scale up a 478mm^2 monster to 3.6GHz but wouldn't be able to scale a punny APU with a fifth of CPU die area to 4.4Ghz?

 

[NTMBK]

So do you think TSMC can scale up a 478mm^2 monster to 3.6GHz but wouldn't be able to scale a punny APU with a fifth of CPU die area to 4.4Ghz?

Hey, just because Intel can clock a Haswell at 3.9GHz doesn't mean that they can clock an Atom at 5GHz, does it?

 

[nehalem256]

its not just the size of the chip. its whether the desired clocks can be hit. Richland hits 4.4 Ghz with max Turbo. with voltage overclocking it can go upto 4.6 - 4.8 ghz. Kaveri needs to atleast be able to do similar clocks.
if Kaveri has a 20% higher IPC and 20% lower max clock compared to Richland , net CPU perf gain would be very close to nothing.

http://hexus.net/tech/reviews/cpu/56657-amd-a10-6800k-32nm-richland/?page=10

http://techreport.com/review/24954/amd-a10-6800k-and-a10-6700-richland-apus-reviewed/11

can TSMC 28nm HP deliver those speeds or switching frequencies ?

The notebook market is far more important than the desktop market at this point. Mobile Richland is a 2.5/3.5T part.

So if it could hit those clocks speeds with a 20% IPC boost it would be well worth it for mobile.

 

[mrmt]

Hey, just because Intel can clock a Haswell at 3.9GHz doesn't mean that they can clock an Atom at 5GHz, does it?

Probably no, but if they can reach 3.6GHz on that behemoth they should achieve at >>least<< 3.6GHz on Kaveri. Sure, it wouldn't blow away the desktop market but it would suit the mobile market very well.

 

[raghu78]

Probably no, but if they can reach 3.6GHz on that behemoth they should achieve at >>least<< 3.6GHz on Kaveri. Sure, it wouldn't blow away the desktop market but it would suit the mobile market very well.

the problem is Richland 2.5 Ghz (3.5 ghz max turbo) is a 35w product on GF 32 nm SOI. Kaveri needs to atleast be close to 2.5 Ghz at the same 35W TDP to make it a worthwhile upgrade. TSMC 28nm HP bulk could find it difficult to match the power and perf characteristics of the GF 32nm SOI process. in fact that could be the same problem for GF 28nm HPP.

 

[mrmt]

the problem is Richland 2.5 Ghz (3.5 ghz max turbo) is a 35w product on GF 32 nm SOI. Kaveri needs to atleast be close to 2.5 Ghz at the same 35W TDP to make it a worthwhile upgrade. TSMC 28nm HP bulk could find it difficult to match the power and perf characteristics of the GF 32nm SOI process. in fact that could be the same problem for GF 28nm HPP.

What you say here might be true, but this isn't what you said first:

TSMC has not manufactured high performance microprocessors at 4+ Ghz speeds ever. Intel and IBM are only the companies who have the processes suitable for high performance and only Intel with bulk technology. IBM used SOI to provide similar transistor performance as Intel's bulk at same node.

To say that GLF 32nm SOI *might* be better suited to Kaveri than TSMC 28nm bulk is one thing, to say that TSMC has no process suitable for AMD APUs is another, and you were saying the latter.