ARM announces new Cortex A72 core

anexanhume

Veteran
Just saw this come up on twitter:

https://twitter.com/ARMMobile/status/562678216679493635

First performance claim: https://twitter.com/ARMMobile/status/562679255436636163

edit: This performance claim has since been shown to be at some constant power envelope not disclosed.
#CortexA72 at 16nm is 3.5x #CortexA15 @28nm performance + 75% less energy for workloads! @nayampally_ARM #Mobile2016

https://twitter.com/JohnHeinleinARM/status/562678111985467393
New POP IP for TSMC 16FF+ enables up to 2.5GHz for #CortexA72 in mobile configs = faster time to market #Mobile2016


They also announced the new Mali T880 GPU: https://twitter.com/JohnHeinleinARM/status/562677951402348547
New #MaliT880 delivers 1.8x perf vs #MaliT760 shipping in devices today #Mobile2016
 
Last edited:
3.5x Cortex A15?
Isn't that within Broadwell's territory on performance alone?
 
3.5x Cortex A15?
Isn't that within Broadwell's territory on performance alone?

Are there numbers for Broadwell in DMIPs?

Using 3.5 DMIPs/MHz for A15 and assuming a 1.5GHz to 2.5 GHz scaling factor, that puts A72 around 7.35 DMIPs/MHz (assuming ARM is again using DMIPs as their standard). That would be around a 50% jump over A57, not accounting for clock differences. Sounds fair.
 
First off..the naming convention seems a bit odd..what are they going to name the new LITTLE core now? Was hoping to see a significant update to the A53..it's performance compared to the A7 is a bit of a let down IMHO. I guess that's going to be announced another day.

They also announced a new CCI-500 interconnect btw - https://twitter.com/ARMMobile/status/562680394534092802
RT @JohnHeinleinARM: New Corelink CCI-500 interconnect 2x peak system bandwith, optimized for #bigLITTLE #Mobile2016



They say 1.8x performance compared to T760. Given that T860 is basically the same as T760, wonder if T880 is based on a new architecture..or is still on Midgard.
Are there numbers for Broadwell in DMIPs?

Using 3.5 DMIPs/MHz for A15 and assuming a 1.5GHz to 2.5 GHz scaling factor, that puts A72 around 7.35 DMIPs/MHz (assuming ARM is again using DMIPs as their standard). That would be around a 50% jump over A57, not accounting for clock differences. Sounds fair.

Did they state that the A15 was at 1.5 Ghz? I am not able to find any reference to that figure. A72 being ~50% faster than the A57 sounds a bit high..but presumably the comparison is a A72 on 16FF+ vs A57 on 20SoC, so there should be a significant difference in clocks).
 
Last edited:
Full press release is here: http://www.arm.com/about/newsroom/arm-sets-new-standard-for-the-premium-mobile-experience.php

Did they state that the A15 was at 1.5 Ghz? I am not able to find any reference to that figure. A72 being ~50% faster than the A57 sounds a bit high..but presumably the comparison is a A72 on 16FF+ vs A57 on 20SoC, so there should be a significant difference in clocks).

No, I was making a guess based on common frequencies for that core, and one that would paint them in the most favorable light (hence the 3.5 rather than 4.0 DMIPs assumption too).
 
http://semiaccurate.com/2015/02/03/arm-outs-a57-successor-maya-core-cortex-a72/

The new interconnect is what scares me actually as it opens a whole new can of worms...

What can of worms?

Sidenote..the new interconnect will allow upto 4 clusters so Mediatek can make a 16 core A53 SoC. Oh Joy...
Full press release is here: http://www.arm.com/about/newsroom/arm-sets-new-standard-for-the-premium-mobile-experience.php

No, I was making a guess based on common frequencies for that core, and one that would paint them in the most favorable light (hence the 3.5 rather than 4.0 DMIPs assumption too).

Thanks for the link..the press release has a lot of good information.

But wouldn't the most common frequencies be 1.7-1.8 Ghz? In the press release ARM says - "3.5X the performance of 2014 devices based on the Cortex-A15 processor". Most 2014 devices have been higher than 1.5 Ghz to my recollection. If we take 1.8 Ghz..then the A72 would be about 8.82 DMIPS/Mhz, so about 85% faster than the A57.

Also..check out the A72 product page and click on the performance tab - http://www.arm.com/products/processors/cortex-a/cortex-a72-processor.php

ARM says that 20nm A57 is 1.9x the performance of 28nm A15 and 16nm A72 is 3.5x faster than 28nm A15. So if we compare A72 to A57..i.e. 3.5/1.9, it gives us 84%..which corroborates the 85% figure. Quite impressive if true.


Another line from the press release -
Extended performance and efficiency when the Cortex-A72 CPU is combined with a Cortex-A53 CPU in ARM big.LITTLE™ processor configurations

Seems strange not to introduce a new LITTLE core along with the A72 and to again mention the A53..but maybe I'm reading too much into it.
 
Seems strange not to introduce a new LITTLE core along with the A72 and to again mention the A53..but maybe I'm reading too much into it.

One possibility may be that something delayed whatever the A53's replacement would have been, but there could be other reasons.
The transition to FinFETs changed a number of things for Intel, and ARM POP cores are likely seeing some of them. Leakage is significantly cut back, thanks to the improved control over the channel, and we see that Intel's high-end cores have a very significant dynamic range.
It may very well be that thanks to the much better physical properties of the A72's implementation (possibly better DVFS as well?) that it can go significantly lower in the power/performance graph that was used to illustrate where the LITTLE range ended and the big one began.
If that's the case, maybe the 16nm version of the A53 POP is better off not being expanded, but being more heavily optimized physically to fit below the A72, relying on the general uplift of the process to handle the rest.

I'm curious if there is some import to being able to run with the if this A53 if its ISA support is unchanged. ARM V8 has one 0.1 revision tacked on already, and I'm not clear on whether that's enough to change the core name.
 
1. All performance figures talk about sustained performance, not peak performance. Due to process and efficiency gains we have no idea what actual the actual IPC improvements are. Be very careful with these marketing figures.
2. The new interconnect is great. I'm more excited about that than the A72.
3. The A72 is still used in a bL config, paired with A53's. I think their new little core is still some time off.

We'll have more info on the A72 arch in the coming months.
 
@16nm those Cortex A53 replacement cores would have been so tiny that engineers decided to integrate them into the big cores. Just kidding.

Other than that W&S.
 
Are there numbers for Broadwell in DMIPs?

Using 3.5 DMIPs/MHz for A15 and assuming a 1.5GHz to 2.5 GHz scaling factor, that puts A72 around 7.35 DMIPs/MHz (assuming ARM is again using DMIPs as their standard). That would be around a 50% jump over A57, not accounting for clock differences. Sounds fair.
How would the A72 compare to Apple's Cyclone in terms of clock for clock performance?
 
No mention of die size for the A72, that I've seen? I'd imagine this indicates it is substantially larger than the A57, which is understandable when you consider the purported increase in performance and the size of the Cyclone and Denver cores in comparison to A57.

Perhaps ARM have decided that A53 provides good enough performance for now when die size is taken into comparison and a larger A7x version wouldn't be worth producing at present?
 
could it be that the leak from mediatek about the 10 and 12 core SoC has something to do with the CCI-500:

http://blog.gsmarena.com/mediatek-may-planning-develop-cpus-12-cores/

That could mean that we will see the new technology a little earlier than expected.

[edit]
Using 3.5 DMIPs/MHz for A15 and assuming a 1.5GHz to 2.5 GHz scaling factor, that puts A72 around 7.35 DMIPs/MHz (assuming ARM is again using DMIPs as their standard). That would be around a 50% jump over A57, not accounting for clock differences. Sounds fair.

I wouldn't expect such a large increase in DMIPS at all. ARM talks about "sustained" performance in the comparison. The increase from the A57 @20nm and A72 @16nm is 84% (3,5/1,9). The process chance contributes around 20% to the increase according to a analyst PDF from BNP Paribas. Therefore at the most "only" 53% can be attributed to the SoC improvements. From this the CCI-500 contributes ~30% which leaves only 18% for the improvement of the core itself.

According to this page the A57 reaches around 4.8 DMIPS/MHz:
http://us.hardware.info/reviews/538...id-smartphones-to-be-expected-android-64-bitn

Therefore I would expect (as a very wild guess of course) ~~ 5,7 DMIPS/MHz for the A72 "only"
 
Last edited:
It's 3.5x the power efficiency at cell phone power levels.

Look here in the performance tab. There is no mention of outright performance.


You're right. I assumed they were talking about absolute performance because the same tweet mentioned a performance comparison first and a power consumption comparison afterwards.
 
You're right. I assumed they were talking about absolute performance because the same tweet mentioned a performance comparison first and a power consumption comparison afterwards.

So did I.

This tweet is outright misleading. "Performance" here means power efficiency at some arbitrary power level.

Cheers
 
Back
Top