AMD to transition to 28nm bulk in 2013 (digitimes)

[ShintaiDK]

You don't have any road maps to verify this. Don't make claims that you cannot back up with facts.

I dont see how my post is invalid with the 2 added roadmaps.

Its no secret from AMDs CEO either that they gave up on performance parts. Its the cycle for AMD to replace VIA.

 

[Idontcare]

FD-SOI can be cheaper than bulk.

The Substrate cost (wafer) is higher than a bulk-Si but Litho and FEOL process steps are less making the FD-SOI cheaper overall.

Keyphrase in the first sentence is "can be".

Yes it can be cheaper if the engineers developing the bulk-Si process do silly things to make the bulk process much more expensive (no limit there) or if the engineers developing the FD-SOI process hit upon a stroke of genius and make their process cleverly lower in cost over that of the process developed by the competition.

This does happen by the way, there is fierce patent competition to develop integration of electrical components with the fewest mask-adders possible...developing a mimcap which is 1-mask adder versus say a mimcap that is 2-mask adder for example.

So the lead in process simplicity that the FD-SOI team has can easily be blown in terms of overall cost/wafer if the production fab itself is running at slightly slower cycle-time or has intrinsically slightly higher defect density (D0).

In the end, because of all the contributing factors that go into cost/wafer, we never have real-world examples of apples-to-apples comparisons except when a fab like IBM's fishkill runs both flavors of a node at the same time.

ARM Announces 45nm SOI Test Chip Results That Demonstrate Potential 40 Percent Power Savings Over Bulk Process

The silicon results show that 45nm high-performance SOI technology can provide up to 40 percent power savings and a 7 percent circuit area reduction compared to bulk CMOS low-power technology, operating at the same speed. This same implementation also demonstrated 20 percent higher operating frequency capability over bulk while saving 30 percent in total power in specific test applications.

The problem is that what we do have evidence of in the real-world is that (1) TSMC is the only foundry to have made money in the entire history foundry business and they have determined that SOI is not viable for their business model, and (2) the other major user of SOI was AMD and they too failed to financially capture success with their reliance on SOI whereas Intel's reliance on Bulk-Si certaintly did not lead to its financial undoing.

So this seems to be a case of theory versus practice. In theory there is no difference between theory and practice, but in practice there is.

It seams to me that the difference is only in the wafer. AMD can purchase ready FD-SOI wafers.
Can AMD or any other use the same 28nm process at GloFo but use FD-SOI wafers ???

Im asking because the transistors will be created over the FD-SOI Thin undoped body.

Yes, but...the but part comes in in that the layout of any given IC is going to be specifically optimized for the nuances and tradeoffs that come with the overall chip properties, both electrical and thermal.

So while AMD surely could in theory reuse the same maskset with an FDSOI wafer substrate as they would for a bulk-Si substrate, they would be fools to do so and not take advantage of any layout optimizations that ought to be done in porting one design for the other.

After all, the parametrics would change, electrostatics and all that, upon substitution of the FDSOI wafer for the bulk-Si wafer (if the parametrics did not change then why use the more expensive FDSOI wafer?). As soon as those parametrics change then so to do all the integrated circuit function, timings and so on.

You risk taking a design that yields at say 4GHz on bulk-Si and ending up with a design that has to be clocked at a lowly 1GHz on FDSOI for instance, but if you redo the maskset and layout work (a year minimum lead-time once you factor in validation) then you might get back to a 4GHz product but with lower power-consumption as the benefit.

Basically what this boils down to is that in practice, if you are going to use FDSOI at 14nm then you need to have already decided that today and be in the process of designing your chips with that expectation in mind.

 

[podspi]

The problem is that what we do have evidence of in the real-world is that (1) TSMC is the only foundry to have made money in the entire history foundry business and they have determined that SOI is not viable for their business model, and (2) the other major user of SOI was AMD and they too failed to financially capture success with their reliance on SOI whereas Intel's reliance on Bulk-Si certaintly did not lead to its financial undoing.

So this seems to be a case of theory versus practice. In theory there is no difference between theory and practice, but in practice there is.

I assume you mean pure for-hire foundries? Since I'm pretty sure that both Intel and Samsung fab things for 3rd parties AND make a profit :biggrin:

How many 'just fab' companies are actually in existence? I'm not an expert on the fab industry, the only two I can name are GF and TSMC.


Anyway, I agree that SOI is not the way to go. It is a useful technology if you are a lower-volume manufacturer, but AFAIK that is not one of GF's goals...

 

[Idontcare]

I assume you mean pure for-hire foundries? Since I'm pretty sure that both Intel and Samsung fab things for 3rd parties AND make a profit :biggrin:

How many 'just fab' companies are actually in existence? I'm not an expert on the fab industry, the only two I can name are GF and TSMC.


Anyway, I agree that SOI is not the way to go. It is a useful technology if you are a lower-volume manufacturer, but AFAIK that is not one of GF's goals...

Naturally I was speaking of pure-play foundries. We made a profit in the foundry business at Texas Instruments where we were a foundry for a number of IC designers (including Cyrix back in the day, as well as SUN for more than a decade) but it never really grew into enough business to take on a life of its own.

Fun fact - Morris Chang who founded TSMC actually came from TI. He was our VP in charge of semiconductor operations for some 20yrs before leaving to start TSMC

As for your question about other pure-play foundries, they come and go because no on else makes money, but the longest rival to TSMC has been UMC. Look near the bottom of this link for foundry rankings by revenue over the years.

At any rate, I know it sounds like hyperbole but I'm not making it up, there was financial analysis report that circulated a couple years ago (one of those that costs $3k to buy, I only had access because TI circulated it internally) that compiled all the revenue, profits, and losses for the entire pure-play foundry industry since its creation and TSMC was the only company to have profited from the industry. Everyone else that made profits at some point in time ceded them all back and then some in losses.

 

[Arachnotronic]

If by claiming that AMD is focusing on APUs means that they've given up the "performance" chips, does that mean Intel has too?

Intel's SNB/IVB is a direct competitor to LLano/Trinity. SNB-E/IVB-E is just the server/workstation stuff rebranded. I'm sure AMD could rebrand their ultra high core count server stuff as desktop...

 

[ShintaiDK]

If by claiming that AMD is focusing on APUs means that they've given up the "performance" chips, does that mean Intel has too?

Intel's SNB/IVB is a direct competitor to LLano/Trinity. SNB-E/IVB-E is just the server/workstation stuff rebranded. I'm sure AMD could rebrand their ultra high core count server stuff as desktop...

No, its no secret AMD cant keep up today. They are already 50-75% behind. Now its just official so to say. Just see today with i3s running around AMDs "highend/performance".

 

[Riek]

No, its no secret AMD cant keep up today. They are already 50-75% behind. Now its just official so to say. Just see today with i3s running around AMDs "highend/performance".

50-75% is that the performance difference multiplied with the power difference? or another number pulled out of thin air.
Also FX8 is alot faster than any i3... just not in 1/2 threaded applications..

AMd has its limitations, especially on very low threaded apps. But 50-75% is very exagurated.

 

[Homeles]

I dont see how my post is invalid with the 2 added roadmaps.

Its no secret from AMDs CEO either that they gave up on performance parts. Its the cycle for AMD to replace VIA.

You're misinterpreting AMD's pullout from the high end.

There's no new "performance parts" for 2013... they'll still be using Piledriver. But 2014? No one knows, so don't claim you do. Nowhere has it been stated that APUs will be all that AMD offers.

 

[ShintaiDK]

50-75% is that the performance difference multiplied with the power difference? or another number pulled out of thin air.
Also FX8 is alot faster than any i3... just not in 1/2 threaded applications..

AMd has its limitations, especially on very low threaded apps. But 50-75% is very exagurated.

Just look at gaming benches. 50-75% is normal.

Im sure you can always find some GPU limited benches or the usual zipping contest to prove me otherwise. besides the classic "We dont need more" statement. But that doesnt change reality.

AMD is the new VIA.

 

[frozentundra123456]

50-75% is that the performance difference multiplied with the power difference? or another number pulled out of thin air.
Also FX8 is alot faster than any i3... just not in 1/2 threaded applications..

AMd has its limitations, especially on very low threaded apps. But 50-75% is very exagurated.

But those are what most people use. I would say more like 40 to 50 percent slower ipc, which is not made up for by more cores except in very limited applications.

 

[ShintaiDK]

You're misinterpreting AMD's pullout from the high end.

There's no new "performance parts" for 2013... they'll still be using Piledriver. But 2014? No one knows, so don't claim you do. Nowhere has it been stated that APUs will be all that AMD offers.

Steamroller for APUs, Pilediver for the last FX. Not to mention the AM3 platform is dead in the serverspace, just like all other current AMD platforms. AMD just got nothing in the performance segment.

 

[Homeles]

Just look at gaming benches. 50-75% is normal.

Im sure you can always find some GPU limited benches or the usual zipping contest to prove me otherwise. besides the classic "We dont need more" statement. But that doesnt change reality.

AMD is the new VIA.

You claimed Bulldozer is 50-75% behind Intel, which implies slower. Not a single graph there verifies that claim -- at worst, it's 45% slower.

It's semantics, but they are two very different things. Intel being 50-75% ahead of AMD is very, very mathematically different than AMD being 50-75% behind Intel.

Steamroller for APUs, Pilediver for the last FX. Not to mention the AM3 platform is dead in the serverspace, just like all other current AMD platforms. AMD just got nothing in the performance segment.

You still don't have proof to back up these claims. I, in fact, have proof to verify the opposite.

New CPUs AND APUs coming to servers on 28nm. No release date... but AMD is not done in the performance segment. If AM3+ is dead, what will be its successor? Take a wild guess... and also take a wild guess on what type of memory it will use and when that memory becomes available...

 

[Riek]

No, its no secret AMD cant keep up today. They are already 50-75% behind. Now its just official so to say. Just see today with i3s running around AMDs "highend/performance".

Just look at gaming benches. 50-75% is normal.

Im sure you can always find some GPU limited benches or the usual zipping contest to prove me otherwise. besides the classic "We dont need more" statement. But that doesnt change reality.

AMD is the new VIA.

I don't see an i3 in any of those benches.
Gaming is also not mentionned in your post so i was going by overall performance and not very task specific performance.

@frozentundra123456:
That may very well be true, but those aren't the applications which will show the difference between cpu speeds either.

 

[ShintaiDK]

If AM3+ is dead, what will be its successor? Take a wild guess... and also take a wild guess on what type of memory it will use and when that memory becomes available...

The dualcore APUs we see today. So FMx platform replaces AM3+.

AM3+, C32 and G34 is dead as you show yourself.

 

[Homeles]

The dualcore APUs we see today. So FMx platform replaces AM3+.

AM3+, C32 and G34 is dead as you show yourself.

If AM2+ replaced AM2, AM3 replaced AM2+, and AM3+ replaced AM3, what comes next?

Also, what memory did those sockets use?

AMD's next server and high end desktop platform debuts in 2014:
http://www.realworldtech.com/page.cfm?ArticleID=RWT020712143904&p=2

 

[ShintaiDK]

If AM2+ replaced AM2, AM3 replaced AM2+, and AM3+ replaced AM3, what comes next?http://www.realworldtech.com/page.cfm?ArticleID=RWT020712143904&p=2

FM2 or FM3.

 

[Homeles]

FM2 or FM3.

Great, I'm arguing with a guy that can't count to 4.

 

[meloz]

Given that the process by which you made an FD-SOI requires you to start with a bulk-Si wafer and then do a bunch of processing steps to it to create the FD-SOI wafer, I'm not sure how an FD-SOI wafer could ever be the same cost as a bulk wafer.

<snip>

The loser in all this is SOITEC. Their existence depends on the niche market that was AMD's reliance on SOI, and being a middle-man their existence was a cost-adder. That cost-adder is being squeezed out of the equation.

Great post; you explained it in such a lucid way that everyone can understand.

 

[podspi]

Frankly, we don't know what AMD is going to do, although imho it would make sense for them to EOL desktop CPUs. Remember they are pushing Fusion? If they actually believe their own slogan, APU is the future. Now I know we all (or some of us anyway) want another AMD CPU that makes IVB or Haswell look like an embarrassment similar to what the Athlon did to the Pentium 4, but that is neither necessary or likely. Assuming AMD can keep up with Intel's mainstream, APU's are fine.


I would be very surprised if they didn't continue making CPUs for the server market, though.

 

[meloz]

Now I know we all (or some of us anyway) want another AMD CPU that makes IVB or Haswell look like an embarrassment similar to what the Athlon did to the Pentium 4, but that is neither necessary or likely.

If it is not necessary why do people (and business) buy the latest generation and faster processors? Why is the average age of a desktop computer in home or server in datacenter not 15 years; why is it not increasing dramatically even though we have had 'good enough' processors for at least 5 years?

It is not an either-or situation. AMD would be much happier if they could mate a CPU like IVB with their iGPUs, rather than their current underwhelming CPU lineup. AMD know they are weaker in the CPU area, so they prefer to talk about their strengths. Intel play the same game. As consumers we would be extremely naive to buy into such attempts to spin information by either manufacturer.

 

[lifeblood]

Great, I'm arguing with a guy that can't count to 4.

I think the point is that AMD is probably done making CPUs only, so the AM sockets are dead. The APU requires an FM type socket to support the additional pins used by the GPU. Hence I agree that the next desktop or enthusiast CPU (after PD) by AMD will require FM2 or FM3.

If HSA works then that won't be a bad thing, either. I'm not interested in a APU now because I have a discrete GPU and an iGPU is useless while still generating heat. But in the future that hopefully will change.

 

[Homeles]

I think the point is that AMD is probably done making CPUs only, so the AM sockets are dead.

Did you not see the roadmap I linked? CPUs are not dead yet.

 

[Phynaz]

Did you not see the roadmap I linked? CPUs are not dead yet.

I see a slide that runs until.....This year.

AMD has already stated that Piledriver is the end of AM3+. Just Google it.

 

[Homeles]

I see a slide that runs until.....This year.

AMD has already stated that Piledriver is the end of AM3+. Just Google it.

I am well aware of Piledriver being the last AM3+ processor. Please learn to read.

 

[Phynaz]

Oh, so now you're saying AM3 is dead. Okay.