Qualcomm SoC & ARMv8 custom core discussions

I am aware of the situation..see my post above. That is a very long report..I briefly skimmed through it but could not located the relevant section you refer to. Could you point me to it?

Yes..I have seen that..agreed that it may be more power efficient but we do not know to what degree. Besides even if a chip is more efficient, it can consume more power under load and it is the absolute power which will matter at the 28nm node..which is why we haven't seen any A57 implementations on it. Either ways..I'm not ruling it out..it's just unlikely in my opinion (i.e. that they will use A72 on a mid-range part and that it will be on 28nm)

 

On Page 31 they show a nice graph with cost per m gates. I like the report despite it comes from an investment company because it gives a nice overview about 20nm and 16/14"nm" Finfet.

Again based on this report/analysis the gain from going from 28nm to 16nm seems to be 54%. See page 17. So you could be correct that most of the quoted power reduction comes from the process ( I remembered this differently).

Regarding A72 on a midrange-product... well it is my suspicion that mediatek has a lot to do with this. Qualcomms latest lowend and midrange SoC seem to be subpar compared to the offerings from mediatek (MT6732/MT6752) regarding CPU muscle (which by the way caught me by complete surprise) so maybe the A72 is a way to make sure this happens not again.

 

As I previously wrote, this graph has been reproduced all over the web during the last months. Since it all comes from the same source, I really would like to read some other analysis not repeating IBS prediction.

 

That figure does not seem to be based on any analysis..it seems to be derived directly from TSMC's marketing figures. TSMC claims 20SoC is 30% more power efficient than 28nm and 16FF is 35% more power efficient than 20SoC. Therefore power consumed by 16FF = 1*0.7*0.65 = 0.455 = 46% of 28nm, i.e. a 54% reduction in power. But yes..as you note..all of the marketing figures from ARM do take into account process gains and are typically for "sustained performance".

Given that both Qualcomm and Mediatek have almost the exact same CPU configuration.i.e. 4xA53 and 8xA53..I would be surprised if there was a major difference in performance between them (taking frequency into account). I agree that Mediatek has really caught up to and is giving Qualcomm a run for their money in the mid-range, but Qualcomm still has superior modems..especially for LTE. And given that this segment is highly cost sensitive..going for A72 does not seem to be a very effective way to compete in my eyes. The average consumer barely knows the difference between an A5 and and A15..and I doubt its going to be any different for A53 and A72. The number of cores seem to matter more than the type so A72 does not seem to make much sense for a low to mid-range part.

Exactly..this is all from IBS and Handel Jones..who, by the way, also said that he did not think 20nm would not be in volume production in 2014 and that 14/16nm would be delayed to 2016. That prediction, along with the cost per gate graph is from the SEMI ISS Conference in January 2014. It may have been representative at the time but it is not a static figure, it is dynamic. A better graph would be the one from Nvidia (which I've posted on the previous page), which shows the trends and a cost per transistor crossover between 28/20nm around Q1'15. Maybe that has changed but the trends are broadly the same.

So yes, 28nm will remain the dominant node for 2015 and probably 2016 even..but towards the end of this year and into next year..the business case for 20nm in the mid-range SoC market starts to becomes stronger.

 

It seems that Qualcomm's new ARMv8-A core is called Mamba. There is also a cancelled 32-bit core called Taipan that was supposed to be in Snapdragon 810 and 808.

 

http://www.gizchina.com/2015/02/09/mediatek-mt6752-vs-snapdragon-615/

Well not the best comparison but it should show why I was so surprised about the performance of the MT6732/6752

 

Qualcomm have announced the Snapdragon 415, 425, 618 and 620.

415: 8xCortex A53 Up To 1.4Ghz, Adreno 405, LPDDR3 667mhz, Hexagon V50 DSP, 150Mbps Modem, 28nm
425: 8xCortex A53 Up To 1.7Ghz, Adreno 405, LPDDR3 933mhz, Hexagon V50 DSP, 300Mbps Modem, 28nm

618: 2xCortex A72 Up To 1.8Ghz&4xCortex A53 Up To 1.2Ghz, Next Generation Adreno, Dual Channel LPDDR3 933Mhz, Hexagon V56 DSP, 300Mbps Modem, 28nm
620: 4xCortex A72 Up To 1.8Ghz&4xCortex A53 Up To 1.2Ghz, Next Generation Adreno, Dual Channel LPDDR3 933Mhz, Hexagon V56 DSP, 300Mbps Modem, 28nm

https://www.qualcomm.com/products/snapdragon/processors
http://arstechnica.com/gadgets/2015...hips-bring-high-end-features-to-midrange-socs

 

These had been leaked before, so it's not a complete shock, but I have to say I'm surprised to see Qualcomm using A72 cores after relying on A57 ones for what seemed to be a stop-gap solution. What on Earth happened to Qualcomm's custom cores?

 

So we can expect some high end SOCs using Cortex A72 for the end of the year in 14/16nm ?

 

I'm a little shocked about the timing of the press release. The press release for the midrange 615 for example was in February 2014 and the first smartphones using the 615 where announced around September as far as I remember. Therefore we can expect the first smartphones with 618 and 620 in End of Q3/ Beginning of Q4 2015... that is really fast!

Compare that with the A15 or A57... when ARM had the press releases for this CPUs and when the first smartphones of tablets were released.

 

Makes me wonder what's going to be in the next 800s.

 

Octa A72?

 

Probably Qualcomm's own ARMv8?

 

Maybe, but it would be a tremendous leap to go from Krait to something that can offer competitive peak perf vs Cortex-A72 while maintaining better perf/W.

AFAIK 810 clocks at up to 2GHz. With the process improvement and uarch improvements they should be able to hit higher clocks in the same power budget, maybe as high as 2.4GHz. That alone could be enough of a step up in performance vs the 1.8GHz 620 to justify being 8xx series. Just look at the first 600 vs 800, which merely bumped the clock speed of the Krait 300 from up to 1.9GHz to up to 2.26GHz.

 

If the 32bit Krait successor had not been cancelled, the leap would probably be smaller.

 

Either way, it seems to me that ARM is iterating cores more aggressively than Qualcomm. In the rough time span of Scorpion and Krait ARM has released A8, A9, A15, A17, and A57, not to mention A5, A7, and A53. While Qualcomm did update Krait, the changes were pretty minor.

Maybe that does mean the next new thing was going to be a big jump, and maybe it just got delayed in favor of getting a 64-bit version out faster. But A72 showing up in their lineup isn't that promising, unless it's because their chip is huge and using A72 saves substantial die area over it. That's believable for ARM11, A5, A7, and A53, but difficult to for A72.

It also seems like a much poorer return on investment to make a CPU that's only used in high end devices, although at least it's probably better than whatever nVidia's getting out of Denver..

 

I don't know nothing about their (upcoming) in house high end CPU cores but it seems that Qualcomm is loosing design wins either to proprietary or Mediatek designs in the low-end at pretty fast pace.
There were a bunch of low-devices running on snapdragon 400, some running on the newer snapdragon 410, I see those devices replaced by non-qualcomm solutions in a lot of manufacturers line up(LG from the fino/bello/f60+ the freshly announced magma/spirit/etc, sony from Xperia e3 to e4, and a lot more manufacturers).

Decent Lte enable SoC are a commodity, worse they don't even need to produced on the newest lithography (/20nm) even less on upcoming ones.
Anytime I see a SOC, the CPU+GPU are taking a extremely limited amount of the die area, adding resources to the SOC is cheaper than adding RAM or storage. The marketing team are doing a good job at creating demand as truth be told I suspect no average users would be able to tell apart to identical phone, one powered by a snapdragon 400 and the other by a snapdragon 200. Performances are already good enough for most average use as far as the SoC are concerned. Differentiation though extra power (or something perceived as such) is cheap, it seems most of the manufacturers can afford to add 2, 4 or even 6 cores for that sake.

I wonder if Qualcomm should consider changing its business plan, the marking is on the wall a lot of their business is going away. Not an easy decision to take and there are no easy answer to the challenges they are facing.

Edit I forgot quite a few words and corrected some mistakes/misspelling.

 

I've looked at nearly 30 SoCs in the last couple of years, a mix of low, mid and high-end, and the average area dedicated to CPU and GPU is just under 20% each.

 

Qualcomm Kryo announced.:
http://arstechnica.com/gadgets/2015...ill-feature-kryo-cpu-kraits-64-bit-successor/
For instance. Information seems rather sparse.