Apple A11 SoC

[Gubbi]

The overhead is always greater than the performance gain, I don't think there's any case where SMT improves perf/W.

I think there is, once you max all cores. Running two threads/core would probably yield higher perf/power. That is a non-existing use-case in the mobile space though.

If you're not at max, it's a non-starter. Say SMT buys you 20% increased throughput at a given frequency, you can run two threads on one SMT core at 2.5GHz or run them on two cores at 1.5GHz; The latter wins because of (very) non-linear DVFS scaling.

Cheers

 

[AlNom]

Would any more cache be of use to phone/tablet workloads?

 

[Dayman1225]

I would direct link, but I cannot due to being new

 

[3dilettante]

I think there is, once you max all cores. Running two threads/core would probably yield higher perf/power. That is a non-existing use-case in the mobile space though.

If you're not at max, it's a non-starter. Say SMT buys you 20% increased throughput at a given frequency, you can run two threads on one SMT core at 2.5GHz or run them on two cores at 1.5GHz; The latter wins because of (very) non-linear DVFS scaling.

Cheers

The perf/watt trade-off with SMT may also be changed by the more rigid limits to absolute power, which may not play well with SMT's general motivation of increasing silicon utilization. With utilized silicon being powered and/or active silicon, SMT's filling of pipeline bubbles or utilizing of parallel resources can fragment idle periods below the threshold that the DVFS hardware can shift, or the core effectively engage clock or power gating.

The burstiness of the periods where the wide and fast performance cores are engaged likely increases the premium on the power they consume in terms of leakage and the amount of silicon left marginally active while in the unsustainable power regime. If not well-coordinated, having two threads threatens to leave that core's static power footprint wider thanks to extended periods of wakefullness for the overall core and sub-units, based on the granularity of each level's control. The responsiveness goal of the active portion can be constrained by this extra static cost, or the big core's incremental throughput bubbles being filled when they are at their most expensive.

There are other forms of multithreading that might avoid some of this, although the big cores seem less suited. Switch on Event multithreading could reduce the unpredictability at the pipeline level, and reduce the chance of threads interfering with one another. It might not work with the overall goal of putting a big core to sleep as soon as possible, but perhaps it can work in areas where the power and performance points aren't as variable.
That might allow a smaller subset of low-power cores being kept un-gated while offering just enough responsiveness for tasks that do have spotty utilization, while not being consistent enough to provide a net benefit from cycling between gated and un-gated states.

 

[Grall]

Be careful what you wish for, you'll be loading webpages that max 8 CPU cores, because some developer don't know what they are doing.

Yeah, umm... That'd perhaps be like, opening a literal pandora's box...

Also, all massively multicored CPUs today turbo like crazy, exactly to accomodate the single thread scenario.

Yes, if there aren't already cores loaded by other software. Also, Intel's (and AMDs) highest-cored CPUs don't turbo as high as their lower-core count brethren.

 

[Dayman1225]

https://browser.geekbench.com/v4/cpu/3967901

Supposed A11 Geekbench score.

 

[Arnold Beckenbauer]

2+4 design and Apple GPU.

 

[roninja]

So is this a totally new GPU and non PowerVR derived?

 

[Entropy]

So is then GPU totally new and non PowerVR derived?

That's what they say.
They also say it's 30% faster at game code. They had bigger fish to fry tonight than expanding on the capabilities of the new GPU other than that, unfortunately.

 

[Shifty Geezer]

But as Pcchen says, multi threaded workloads are rare on phones;

Chicken and egg though. If phones supported SMT, software could target it. As you say, maximum utilisation is unwanted, and hiding behind peak core and clock numbers is better than improved efficiency. Wouldn't efficient 2 core designs be better than four cores, if actually doing work rather than sitting idle?

 

[wco81]

https://browser.geekbench.com/v4/cpu/3967901

Supposed A11 Geekbench score.

And here's a iPhone 7 Plus for comparison. Not much higher than the 7 Plus for single core but looks like 2/3 higher multicore score?

Is iOS optimized for multi-core operations?

 

[tangey]

It might have gone under the radar a bit, but Apple also called out the video encoder as being their design, img previously supplied the video encoder and decoder

 

[Grall]

Is iOS optimized for multi-core operations?

It's been said to be poorly optimized, but then, at that time it only supported dual-core devices at most. We've never had a chip like A11 before... Maybe iOS 11 remedies that, or at least partially perhaps.

We'll know in a few weeks at most.

 

[french toast]

4k-60 recording, that's a first on a phone if i remember correctly, multi core performance is pretty amazing for a phone, apple now has fastest phone across 99℅ of scenarios- impressive.
That new gpu looks suspicious though to say the least, there is no way in hell that is a totally brand new gpu architecture whilst able to run all the heavily optimised iOS software and metal APIs, must have stepped on imagination IP somewhere.
interesting how this will play out.
only 3GB ram? eek. £1000 phone and they couldn't put 6GB in? admittedly it would be no faster than 3gb in 90℅ of scenarios- but value for money and all as well as future proofing dictates it should be in there.
one plus 5 manages 8GB.

 

[Grall]

6-8GB RAM would encourage software bloat and would risk make apps run like crap on just about all other iOS devices.

Besides, refreshing DRAM that hardly ever goes used for most people is a drain on battery power.

 

[mfaisalkemal]

maybe anyone with knowledge can explain the architecture of A11 GPU(still TBDR)
https://developer.apple.com/videos/play/fall2017/602/
https://developer.apple.com/documentation/metal/about_gpu_family_4

 

[wco81]

4k-60 recording, that's a first on a phone if i remember correctly, multi core performance is pretty amazing for a phone, apple now has fastest phone across 99℅ of scenarios- impressive.
That new gpu looks suspicious though to say the least, there is no way in hell that is a totally brand new gpu architecture whilst able to run all the heavily optimised iOS software and metal APIs, must have stepped on imagination IP somewhere.
interesting how this will play out.
only 3GB ram? eek. £1000 phone and they couldn't put 6GB in? admittedly it would be no faster than 3gb in 90℅ of scenarios- but value for money and all as well as future proofing dictates it should be in there.
one plus 5 manages 8GB.

I really wonder about that 4K video recording capability.

I haven't used it at all on my 6S Plus because that was the first implementation of 4K video on iOS.

Now if I record 4K60 videos, will I be able to Airplay those videos to Apple TV 5 (announced today) connected to a 4K TV?

But does it even record HDR? Actually only iPhone X has HDR rendering (not recording) while iPhone 8, which has the same A11 SOC, does not render or record HDR?

 

[mfaisalkemal]

from this video https://developer.apple.com/videos/play/fall2017/604/ A11 GPU texture filtering rate 2 time greater than A10. PowerVR Furian maybe?

 

[wco81]

Isn't it suppose to be an Apple designed GPU, not PowerVR?

 

[tangey]

maybe anyone with knowledge can explain the architecture of A11 GPU(still TBDR)
https://developer.apple.com/videos/play/fall2017/602/
https://developer.apple.com/documentation/metal/about_gpu_family_4

I note from those tables that the new GPU still supports Powervr texture compression. So either this is no longer protected under patent (how long ago was this patented), or surely it can't be that IMG were silly enough several years ago to give Apple a perpetual license for this?