Samsung Exynos Thread (big.LITTLE Octa-core)

[witeken]

Great insight.

 

[witeken]

EE Times: Samsung Describes Road to 14nm

Random quote:

Samsung is also investing in 28nm FD-SOI to provide customers with a cheaper, fully depleted silicon-on-insulator architecture. The cost per transistor has increased in 14nm FinFETs and will continue to do so, Low said, so an alternative technology such as 28nm SOI is necessary.

 

[witeken]

Samsung already working on 1.5nm, New Patterning Paradigm?:

Meanwhile, according to the International Technology Roadmap for Semiconductors (ITRS), logic is supposed to scale to the 1.8nm node in the 2025 timeframe. It’s debatable if logic will extend that far, but some are working on it. “My target is to approach 1.5nm,” said E.S. Jung, executive vice president of the semiconductor R&D center at Samsung, at a recent event. “How can we make it happen? We need tools, materials and open innovation.”

At the [Intel] 1.8nm node (which the ITRS and SS call 1.5nm), reached in 2028 at the optimistic 2 year cadence (with 10nm placed in 2015 according to the ITRS'13 so make it 2030 or so :hmm, the logic half-pitch will be 7nm, 3D NAND will be even more disruptive with 200-400 layers at a 22nm half-pitch node. Fin width will be 5nm, just like fin half-pitch. SRAM 6T size will approach 3.000 nano².

ITRS Roadmap 2013 (last page)

 

[witeken]

Another interesting article: FinFET rollout slower than expected

Today, Samsung has roughly 11,000 wafer starts per month (wspm) of 14nm capacity, which represents about 10% of its total 300mm fab capacity, according to Pacific Crest Securities. Over time, Samsung is expected to convert some of its 28nm capacity, giving it a total of 46,000 wspm of 14nm capacity, according to the firm.

Not to be outdone, TSMC will begin volume production for its 16nm finFET process by the middle of this year. By the end of 2016, the company plans to have an installed capacity of 100,000 wspm for 16nm finFET technology, according to J.K. Wang, vice president of 300mm fab operations at TSMC.

How much 14nm capacity would Intel have?

 

[lopri]

Didn't they already say they had a custom microarch on the map to succeed A57?

I am sure they do but seeing the 7420 sweeping the performance floor now I doubt those custom cores (or even A72) will see the daylight. At least not this year. With nothing else coming close, I think the most we can expect for the Note 5 is a revised 7420.

.. GPU scaling going from MP6 to MP8 makes me wonder if Sammy intentionally capped its performance not to make S805 look too bad in comparison..

You mean S810? I suppose that is a possibility if Samsung thought there would be 2 variants of Galaxy S6s like the previous Galaxys. You cannot make one model perform drastically different than another under the same product name.

 

[Sweepr]

You mean S810? I suppose that is a possibility if Samsung thought there would be 2 variants of Galaxy S6s like the previous Galaxys. You cannot make one model perform drastically different than another under the same product name.

I mean Exynos 5433 vs Snapdragon 805.
Ditching Snapdragon was a bold move, but I wish they had done that earlier, by the time Galaxy Note 4 was launched. A 64-bit enabled Exynos 5433 with 25.6GB/s bandwidth would have given Samsung a competitive advantage last year. Heck, it still beats S810 in most CPU tasks.

 

[lopri]

Well, a 64-bit enabled 5433 will reportedly show up in tablets this year. (I think I've learned that from you?) A 64-bit enabled Note 4 would have incurred ire of users who purchased S805-based Note 4. Moreover, it is not clear whether Exynos 5433 with higher GPU performance could have maintained the same power profile as the one existing in the Note 4. Baseband is another matter.

I would not call ditching S810 this year was a "bold" move, though. It's more like a necessary move to remain competitive, and likely with mutual understanding from Qualcomm. It so happened that the stars aligned for Samsung this year.. (or maybe I should say Samsung's foresight is finally paying off)

 

[Idontcare]

Another interesting article: FinFET rollout slower than expected




How much 14nm capacity would Intel have?

That is a good article. Well worth the read, thanks for sharing!

 

[Sweepr]

Well, a 64-bit enabled 5433 will reportedly show up in tablets this year. (I think I've learned that from you?)

It's going to be used in the Galaxy Tab S 2, though it's not clear if Samsung will enable 64-bit or not. Wrong move if they don't IMHO.

A 64-bit enabled Note 4 would have incurred ire of users who purchased S805-based Note 4. Moreover, it is not clear whether Exynos 5433 with higher GPU performance could have maintained the same power profile as the one existing in the Note 4. Baseband is another matter.

That's my point, I think they could have launched a single Exynos version. Instead they choose to gimp Exynos 5433 by not enabling 64-bit to maintaing feature parity with S805.
Removing the BW bottleneck (giving it 25.6GB/s just like the S805) would probably be enough to overcome Adreno 420's minor advantage in synthetic benchmarks like GFXBench, no need to raise clocks (also Mali T760MP6 delivers comparable/better performance in actual games).
Anyway, I'm glad they are finally doing it now. A giant like Samsung can't rely on Qualcomm being able to deliver or not, especially if they want to beat Apple.

Ps: Some Exynos variants like N910C fully support ATT and US bands. Not sure about performance/power vs Qualcomm though.

 

[lopri]

EE Times: Samsung Describes Road to 14nm

I did not get a chance to read that article until now, and there are some interesting tidbits.

There are multiple designs already and some have taped out [on Samsung's 14nm], Low said. End application segments include “single hand-held mobile computing,” consumer applications, graphics, and compute/networking. “Because the technology has developed so much, the chip can fit into a lot of applications,” he noted. (emphasis mine)

Already taped out.. Graphics? Compute/networking?

“Another trend is there used to be a number of second source foundries where they were trying very hard to match their technology to the primary source. Going down below 28 we see these tasks as next to impossible,” he said. “It’s only more recently that we’re comfortable to bring [FinFets] to high manufacturing volume.”(emphasis mine)

If he is talking about outsourcing, his comment is in a stark contrast to what has been rumored so far with regard to A9 production split between Samsung/GF or Samsung/GF/TSMC.

 

[Idontcare]

If he is talking about outsourcing, his comment is in a stark contrast to what has been rumored so far with regard to A9 production split between Samsung/GF or Samsung/GF/TSMC.

Not talking about outsourcing (Samsung -> GF), but talking about a common practice in which lagging-edge foundries will parametric-match the process nodes of leading edge fabs so that it is very easy for customers of those leading-edge fabs to migrate their designs and dual/multi-source the production across multiple foundries.

For example, say it was 2011 and you had a 28nm design, you'd be producing it at TSMC. Fast forward to 2013 and you could, if you desired, have UMC produce the chip as well on their 28nm. It would perform electrically equivalent to your TSMC produced chips, but the cost structure might be different (yields will be different, negotiations will be different, etc).

You could do that, then, because UMC worked very hard for 2yrs to match the physical and electrical parametrics of TSMC's 28nm once TSMC's 28nm had been released to the public and they could determine what parametric values to replicate (metal pitch, drive currents, leakage specs, etc) so that TSMC's customers could port their 28nm designs with minimal hassle.

It is great for fabless customers because you can multi-source, but there is a lag of course. Not all the foundries can launch at the same time so you still have to evolve your product mix over time as foundries come available.

But the Samsung rep is saying that below 28nm that sort of multi-sourcing will be near impossible. Personally I don't agree, I think it is more wishful thinking on their part. Every foundry wants a captured customer base that can't migrate.

 

[Sweepr]

Here's an article on how Samsung catch up to TSCM so fast. It's a bit old but might be an interesting read.

http://english.cw.com.tw/article.do?action=show&id=14895

 

[krumme]

Tsmc opens new 10nm fab. Opens for volume production mid 2016
http://www.digitimes.com/news/a20150420PD207.html

 

[lopri]

@Idoncare: So what he says is essentially that 2nd-tier fabs will have a harder time to catch up ("next to impossible") as the manufacturing process advances beyond 28nm? Then his comment does not contradict the rumors surrounding the A9 production, which I mistakenly thought he was alluding to. Thank you for the clarification.

Here's an article on how Samsung catch up to TSCM so fast. It's a bit old but might be an interesting read.

http://english.cw.com.tw/article.do?action=show&id=14895

That is an emotionally charged narrative with more holes than Swiss cheese. I think actions speak louder in this case and the fact that TSMC did not bring lawsuits nor attempted counter-offensive against Samsung speaks volume.

On the other hand they are reportedly trying to bring their 16nm and 10nm processes ahead of schedule, which is a great news.

 

[kpkp]

ARM roadmap slide?

 

[Arachnotronic]

ARM roadmap slide?

Nice find!

 

[Idontcare]

@Idoncare: So what he says is essentially that 2nd-tier fabs will have a harder time to catch up ("next to impossible") as the manufacturing process advances beyond 28nm? Then his comment does not contradict the rumors surrounding the A9 production, which I mistakenly thought he was alluding to. Thank you for the clarification.

Well that is his claim, I was just trying to explain what he was saying.

Personally I don't believe he is correct. Foundries will do anything and everything to win customers, and sharing a customer is still better than not having a customer at all. So TSMC will do what it needs to do to make the electricals match Samsung if needed so that Samsung customers can also be TSMC customers, it is just the nature of business.

In fact that is exactly what has already happened which is why 16FF+, so in my view the Samsung guy is more just talking about what he hopes will happen versus what has happened at every node (including 14nm) thus far.

That is an emotionally charged narrative with more holes than Swiss cheese. I think actions speak louder in this case and the fact that TSMC did not bring lawsuits nor attempted counter-offensive against Samsung speaks volume.

It really is true though, believe it or not.

 

[Arachnotronic]

Well that is his claim, I was just trying to explain what he was saying.

Personally I don't believe he is correct. Foundries will do anything and everything to win customers, and sharing a customer is still better than not having a customer at all. So TSMC will do what it needs to do to make the electricals match Samsung if needed so that Samsung customers can also be TSMC customers, it is just the nature of business.

In fact that is exactly what has already happened which is why 16FF+, so in my view the Samsung guy is more just talking about what he hopes will happen versus what has happened at every node (including 14nm) thus far.


It really is true though, believe it or not.

Was TSMC's 16FF inferior electrically to Samsung 14nm LPE and/or LPP?

 

[Idontcare]

Was TSMC's 16FF inferior electrically to Samsung 14nm LPE and/or LPP?

to LPP, comparable to LPE

 

[Arachnotronic]

to LPP, comparable to LPE

Gotcha, makes sense then. TSMC said 16FF+ was ~10% faster than 14LPP. I don't know what their metric for speed is; maybe they took a common test structure (i.e. Cortex A57 core) , implemented on both processes, and compared frequency at given power consumption?

 

[Sweepr]

AnandTech: ARM Reveals Cortex-A72 Architecture Details

www.anandtech.com/show/9184/arm-reveals-cortex-a72-architecture-details

New article, thanks Andrei.

 

[Idontcare]

Gotcha, makes sense then. TSMC said 16FF+ was ~10% faster than 14LPP. I don't know what their metric for speed is; maybe they took a common test structure (i.e. Cortex A57 core) , implemented on both processes, and compared frequency at given power consumption?

Yeah that feedback would have been given to them by either their customers (who would be taping out designs for both foundries) or from ARM itself who naturally would have been busy validating IP blocks on both foundry's nodes. 10% faster still doesn't tell you all you need to know though, thermals are a big deal as are wafer costs.

It is no surprise that TSMC would have higher drive currents, just like Intel in the non-foundry space, in the foundry space TSMC has pretty much always delivered the highest drive currents on the subnodes that their customers desired. (which is why they've had the lionshare of the GPU business for the past decade)

 

[lopri]

It really is true though, believe it or not.

I am sure there are truths in there. But even given the most favorable reading, the article doesn't tell me what is to blame on whom. Lots of innuendos but few straight-up facts/details.

 

[lopri]

There is another graph from today's announcement.

Let's do a quick and dirty comparison. We see from single-threaded performance comparison, a 2.5 GHz A72 matches a 2.0 GHz Core-M in Geekbench. But since it falls behind in Specint and Specfp, it may be that the parity in Geekbench is owed to memory scores. We also see both the integer and floating-point performance of the A72 are not quite at 80% of Core-M. (but close) Let's say they are at 78% of Core-M @2.0 GHz. We can then say;

Core-M @2.0 GHz: single-threaded performance of P-1000
Cortex-A72 @2.5 GHz: single-threaded performance of P-780

Normalized performance (P) per clock (IPC)

Core-M: P-500/GHz
Cortex-A72: P-312/GHz

Thus Core-M is a 60% faster than Cortex-A72 in single-threaded workload at the same clock speed. (or Cortex-A72 is 38% slower than Core-M)

There is a question of power consumption, but they tried to play tricks - see how 4W, <1W are "footnoted" by the Geekbench comparison. Thus I do not think they are applicable to Specint/Specfp, which means the above IPC calculations cannot be used in conjunction with those power consumption figures.

 

[witeken]

A72 more efficient than BDW-Y? ARM has a track record of Intel FUD spreading, so we'll see. And as far as I know, the only A72 this year will be released at 28nm at the end of they; A72 has to compete versus SKL-Y and CNL-Y, so good luck with that.