TSMC and ARM Tape-Out First ARM Cortex-A57 Processor on 16 nm FinFET Technology

[SocketF]

2) They are going for GLF-like half-baked solutions, like their 14nm extreme lemon process.

Given TSMC track record, I can't see them doing neither 1 nor 2, meaning that 16nm will arrive in the 2015-2016 window.

Just read the full eetimes article, it is option 2 :

TSMC’s 16-nm FinFET process will be substantially similar to its 20-nm high-K metal gate SoC process in the back-end, said Cliff Hou, vice president of R&D at TSMC, in a conversation with EE Times after his talk here. Other companies are expected to take a similar approach of marrying 14- and 16-nm FinFET structures with their back-end 20- and 22-nm processes.

 

[mrmt]

Just read the full eetimes article, it is option 2 :

Then it won't be a process comparable to Intel's 14nm. Maybe in performance, but not in die size. And die size on this market means a lot.

 

[Fjodor2001]

For reference, Broadwell @ 14 nm was already taped out in summer 2012:

http://forums.anandtech.com/showthread.php?t=2274477

The article there is from September 2012, and says that Broadwell was taped out a few months before that.

It's expected to be available ~2 years after that.

 

[monstercameron]

There is endless roadmaps you can find to show neither TSMC or GloFo actually delivers what they promise on the fancy slides. 28nm for example for TSMC was a good 2 years late. GloFo doesnt even ship 28nm yet, even tho their roadmaps showed 2010.

I think they might have, isnt the rockchip 3188 28nm glofo?

 

[Madpacket]

Intel has a tough road ahead of them. I don't think they'll be able to for much longer but It'll be interesting if they can keep up in fabrication advances to stem off the ever increasing wave of ARM chips. I give Intel 3-4 years at most before their margins become so eroded by "fast enough" computing that they won't be able to keep dumping money into R&D for fabrication purposes, once they lose that edge it'll be tough times

 

[krumme]

Then it won't be a process comparable to Intel's 14nm. Maybe in performance, but not in die size. And die size on this market means a lot.

Yes they are not comparable. There is much more to the process than we can read. TSMC can get their 16nm to market fast because the transistors themselves, as i understand it, is not 16nm but 20nm. That sort of help, but you can debate if 16nm is the right denomination

I would guess TSMC is going for small size and design regulations that focus on it. Thats what they have done historically. TSMC process is really dense. There is a reason ooo bobcat was one third the size of atom, and the reason was just as well the dense 40nm TSMC process that was of newer tech than Intel 45nm and far denser. We have been over that years back. TSMC is good at their job.

Regarding 14nm i think Intel will focus on low power, more than density relative to TSMC. We have to remember ARM have this little-big concept that is kind of a brilliant way of tackling the problem of scalability. We have to see that in combination with TSMC design choice. I therefore asume 16nm will be dense, but dont have the power characteristics of Intel.

Arm bragging about A57 beeing 3 times as fast as A15 and using only 20% power.

http://www.gsmarena.com/16nm_arm_cortexa57_cpu_ready_to_hit_mass_production-news-5799.php

"The new chip utilizes TSMC's 16nm FinFET manufacturing process, even though the transistors are still 20nm. Still, ARM says the A57 outputs up to three times the CPU power of the Cortex-A15 without affecting battery life. And if its battery life you're after, the A57 can match the Cortex A15 performance while consuming up to five times less power."

Via engaget

 

[Idontcare]

Then it won't be a process comparable to Intel's 14nm. Maybe in performance, but not in die size. And die size on this market means a lot.

Its true. Sub-20nm, Intel is going to be in a realm all unto themselves when it comes to areal scaling. No one else has indicated they intend to follow Intel down the rabbit hole sub-20nm.

 

[krumme]

Intel has a tough road ahead of them. I don't think they'll be able to for much longer but It'll be interesting if they can keep up in fabrication advances to stem off the ever increasing wave of ARM chips. I give Intel 3-4 years at most before their margins become so eroded by "fast enough" computing that they won't be able to keep dumping money into R&D for fabrication purposes, once they lose that edge it'll be tough times

Well right now they dont sell any chips for the mobile market, and they still make a healty profit. I dont think that will change in the future but does it matter?

Arm might take the lowest end of servers, but that is hardly the most profitable anyway.

Intel have huge profit opportunities, why they want to compete with Samsung, TSMC, the ARM ecosystem, and little-big is beyond me. Perhaps its the typical growth greed. They should seek business where they are the best, and there is plenty of room for that.

 

[Exophase]

Arm bragging about A57 beeing 3 times as fast as A15 and using only 20% power.

Be very cautious of these "up to" statements. The claim sounds outrageous to me. I doubt an A57 on 16nm will be capable of delivering 3x the performance of an A15 on 28nm regardless of power consumption unless you're artificially limiting the A15 to some very low clock speed or a lower core count. Otherwise the A57 would have to be at something absurd like over 4GHz.

The idea of it delivering 5 times the perf/W sounds at least as absurd. I expect 2x perf/W improvement tops from the process and perhaps a small amount more from uarch refinement.

 

[zsdersw]

"Up to" is a weasel-word term like "practically" and "virtually" that are usually found in various product advertisements.

"Up to" is not a guarantee or proof of anything.

 

[Fjodor2001]

The idea of it delivering 5 times the perf/W sounds at least as absurd. I expect 2x perf/W improvement tops from the process and perhaps a small amount more from uarch refinement.

That is still impressive. I wonder how long ARM can sustain such improvements from one generation to the next?

Won't they eventually hit the same wall that Intel and AMD has?

 

[Exophase]

When I said improvement from the process I meant just that. Improvement purely from moving to the new nodes. This isn't one generation to the next because the 16nm part (regardless of what the geometry is actually like) isn't next generation. It's generation after next generation.

 

[zsdersw]

Won't they eventually hit the same wall that Intel and AMD has?

No, of course they won't. ARM is going to magically be able to overcome all of the obstacles they face and beat Intel at everything. /s

 

[Fjodor2001]

When I said improvement from the process I meant just that. Improvement purely from moving to the new nodes. This isn't one generation to the next because the 16nm part (regardless of what the geometry is actually like) isn't next generation. It's generation after next generation.

Ok, thanks for the clarification. I still think it's impressive though.

 

[SocketF]

Then it won't be a process comparable to Intel's 14nm. Maybe in performance, but not in die size. And die size on this market means a lot.

Yes but Atom was quite big compared to AMD's Bobcat. So maybe ARM's A57 will use less die size and both parameters are evened out in the end.

Question then is "only" performance/watt.

 

[Fjodor2001]

No, of course they won't. ARM is going to magically be able to overcome all of the obstacles they face and beat Intel at everything. /s

What I meant was more "isn't it likely that ARM soon will hit the same wall that Intel and AMD has". The question is when that will happen, how far off it is...

 

[krumme]

Be very cautious of these "up to" statements. The claim sounds outrageous to me. I doubt an A57 on 16nm will be capable of delivering 3x the performance of an A15 on 28nm regardless of power consumption unless you're artificially limiting the A15 to some very low clock speed or a lower core count. Otherwise the A57 would have to be at something absurd like over 4GHz.

The idea of it delivering 5 times the perf/W sounds at least as absurd. I expect 2x perf/W improvement tops from the process and perhaps a small amount more from uarch refinement.

Completely agree, its never going to happen for general usage. It is just bragging.

But they are quite confident. If i look at the performance gains from my s3 to the s4 they are really big. Not like 3 times for the cpu, but nearly in that order for the gpu part. Its 3 years since my nokia e72 had on 600Mhz Arm11. Its stone age now!

Arm went from 90nm A11 to 28nm S800 with integrated LTE

What happened on the x86 stage at the same time?
- AMD missed 28nm for bobcat, and then just introduced bobcat++ on 28nm.
- Intel upgraded Atom to Atom and went from 45nm to 32nm with GPU integrated, with hardly any driver support. Whoooaaa.

 

[ViRGE]

Just as a reminder guys, this is a fab thread. Please keep the Intel vs. ARM architecture discussion to the appropriate threads.

-ViRGE

 

[Ajay]

Its true. Sub-20nm, Intel is going to be in a realm all unto themselves when it comes to areal scaling. No one else has indicated they intend to follow Intel down the rabbit hole sub-20nm.

No one else can afford this kind of depth in R&D :http://www.eetimes.com/electronics-...ad-patterned-path-to-10-nm-chips?pageNumber=1

--------------------------------------

If TSMC is ahead of schedule for 20nm, then it's probably because Apple bought out the entire first quarter of non-risk wafer starts - and paid cash (I don't really know, just guessing). Heck, what's a paltry two billion dollars when you have Apple's bank :whiste:

 

[Exophase]

Just as a reminder guys, this is a fab thread. Please keep the Intel vs. ARM architecture discussion to the appropriate threads.

-ViRGE

Is this thread not also about Cortex-A57? I thought it being taped out for a test chip was significant in and of itself, regardless of manufacturing technology.. or was it already announced to have taped out on something else?

Completely agree, its never going to happen for general usage. It is just bragging.

But they are quite confident. If i look at the performance gains from my s3 to the s4 they are really big. Not like 3 times for the cpu, but nearly in that order for the gpu part. Its 3 years since my nokia e72 had on 600Mhz Arm11. Its stone age now!

Arm went from 90nm A11 to 28nm S800 with integrated LTE

What happened on the x86 stage at the same time?
- AMD missed 28nm for bobcat, and then just introduced bobcat++ on 28nm.
- Intel upgraded Atom to Atom and went from 45nm to 32nm with GPU integrated, with hardly any driver support. Whoooaaa.

The thing is, in the last few years the peak power budget of phones has raised. A lot. Mobile GPUs today are allowed to use vastly more power than they did a few years ago. Mind you, using an ARM11 in 2010 is like using a Cortex-A8 today, cutting edge it was not. You could already get 45nm ARM devices in 2010, including Cortex-A9 Tegra 2s. If you want to start with 90nm ARM11 go back to 2007.

The big increase in power consumption on phones is acceptable for two reasons. First, the phones have gotten a lot physically larger. A big part of facilitating this was the move to touch-only interfaces - you would not have flip phones with 5" screens. Larger phones means more thermal mass to take the heat and more space for higher capacity batteries, although it's partially offset by screens that use more power (mobile display technology seems to have been advancing perhaps more quickly than anything else but I don't know how much power consumption in particular has improved).

The other big reason is that a lot of improvement has been made in power management, meaning that the extra power consumption only happens when you really need it. In particular, a lot of software and hardware advancements have made basic tasks like talk, simple web browsing, video playback, and simple games sip power. These improvements happened across the board at the CPU, GPU, SoC, OS, etc levels. Really demanding stuff like high end 3D games will still kill a phone in a couple hours flat, much more quickly than anything less demanding.

But these improvements are starting to hit diminishing returns, and there's no signs that phones are going to keep getting larger. You can put in a CPU that only ramps up to a high clock speed for short bursts, therefore getting brief and intermittent tasks done quickly without impacting battery life - Intel is very big on this. But, there's a limit even to how much short term power consumption your cooling and power circuitry can take, and there are negative consequences for making a CPU that can clock really high but normally runs at a much lower clock speed (this is where I think ARM's big.LITTLE helps).

Bottom line, I think the days of huge increases in CPU and GPU power on phones are pretty much already over, and you can start getting used to much more modest increases that are driven by process improvements first and design upgrades second, while keeping similar peak power consumption. In other words, much like how things have been for desktop processors for the last several years. People talk about ARM quickly catching up, thinking they have a lot of easy wins to cover because their single threaded performance is way below. That isn't the point, it's about where they their power wall and since these devices are much smaller form factor they hit it at a much different place. ARM is pretty much already "caught up" in that regard (in that they compete decently with Intel who also caught up with power consumption with Atom)

You make it sound like Intel did nothing in the last several years because performance barely improved for Atom, but that's because they started well past the power wall. They just moved in a different direction and did a good job making the platform viable for mobile. There was a huge amount of work in that.

 

[Blandge]

For reference, Broadwell @ 14 nm was already taped out in summer 2012:

http://forums.anandtech.com/showthread.php?t=2274477

The article there is from September 2012, and says that Broadwell was taped out a few months before that.

It's expected to be available ~2 years after that.

I love how this was conveniently ignored by all who believe TSMC 16nm products will see light of day in 2014.

 

[Phynaz]

I love how this was conveniently ignored by all who believe TSMC 16nm products will see light of day in 2014.

I don't think anybody has ignored it. It's well known that tapeout to production silicon is two years - assuming everything goes well.

 

[Idontcare]

Just as a reminder guys, this is a fab thread. Please keep the Intel vs. ARM architecture discussion to the appropriate threads.

-ViRGE

Is this thread not also about Cortex-A57? I thought it being taped out for a test chip was significant in and of itself, regardless of manufacturing technology.. or was it already announced to have taped out on something else?

Just a friendly PSA, for Exophase and everyone else, please don't challenge/debate a moderator's stated position in this fashion.

The appropriate venue for requesting further clarification on a moderator's decision is through the Moderator Discussions subforum.

Moderator's can and do change their positions, they are open-minded, but don't derail a thread just to debate a moderator regarding their decisions handed down within the thread.

It is an asymmetric communication structure, reflecting the asymmetric power structure. A power structure that exists to keep the forum running smoothly.

So please don't debate the mods here. If you absolutely must debate them then take it to Mod Disc where it is appropriate.

 

[Fjodor2001]

Just wondering, does anyone know if it's takes less time from first tapeout to full production if the CPU core is small? E.g. is that time less for an ARM CPU than for a big Intel/AMD desktop CPU? And why?

 

[Idontcare]

Just wondering, does anyone know if it's takes less time from first tapeout to full production if the CPU core is small? E.g. is that time less for an ARM CPU than for a big Intel/AMD desktop CPU? And why?

You need to normalize a few things first before the question would begin to mean something.

For example - are the tapeout teams given the same budget? Are their respective IC designs of comparable complexity, one just consuming larger die area than the other?

Are the design teams equally resourced? (same headcount, comparable experience and education, comparable software/hardware validation support, etc)

Are their end-goals the same? Same margin targets for the ICs in question?

The time one takes to productize an IC is a factor that depends on many aspects of the overall project management triangle.

^ you get to pick and prioritize 2 of the 3, but not more than 2.