Arm a72?

[monstercameron]

It seems no one is talking about it? it is already commercially available in the mediatek soc of the AFTV2015. Is there a write up that I missed? Has anyone spoken about IPC and other aspect of this core yet?

 

[jhu]

It seems no one is talking about it? it is already commercially available in the mediatek soc of the AFTV2015. Is there a write up that I missed? Has anyone spoken about IPC and other aspect of this core yet?

What products are available with it right now?

 

[geoxile]

What products are available with it right now?

Doesn't the new Amazon Fire TV have it?

 

[monstercameron]

What products are available with it right now?

amazon firetv /aftv 2015. mediatek mt8173

 

[Mondozei]

IIRC Huawei's own SoC will have it for their premium models in the next few months.

 

[Nothingness]

IIRC Huawei's own SoC will have it for their premium models in the next few months.

Cortex-A72 is in Kirin 950 which is supposed to be in Huawei Mate 8, but the latter hasn't been released yet.

 

[jhu]

amazon firetv /aftv 2015. mediatek mt8173

Oh snap! If it's at least rootable, I'd be all over that (would be fairly trivial to install Debian).

 

[Nothingness]

Oh snap! If it's at least rootable, I'd be all over that (would be fairly trivial to install Debian).

Quickly looking with Google, it seems like soldering is required

 

[Andrei.]

Quickly looking with Google, it seems like soldering is required

That's for the 2014 variant.

In any case it's a bit too much of a pain to work with otherwise I would have done a piece with it.

 

[jhu]

That's for the 2014 variant.

In any case it's a bit too much of a pain to work with otherwise I would have done a piece with it.

And it comes with a Snapdragon 600. There are other devices with this CPU that are easier to root and/or bootloader unlock.

 

[Andrei.]

There are other devices with this CPU that are easier to root and/or bootloader unlock.

The CPU/SoC has nothing to do with how a device gets rooted.

 

[monstercameron]

The CPU/SoC has nothing to do with how a device gets rooted.

are you sure about that? I've seen root unlocks across multiple devices using the same hardware via similar methods.

 

[podspi]

I think it depends on exactly how the devices gets rooted.

I remember that devices with TI SoCs could be easily rooted (and other fun things) with factory cables.

 

[paffinity]

I'm guessing Apple A7 will still be faster than A72. Single core performance, I mean.

 

[monstercameron]

I'm guessing Apple A7 will still be faster than A72. Single core performance, I mean.

maybe but i dont believe so, arm has stated iirc, 1.1-1.5x perf uplift vs a57.

how was a7 compared to a57?

 

[Nothingness]

I'm guessing Apple A7 will still be faster than A72. Single core performance, I mean.

On Geekbench Cortex-A72 is significantly better than Apple A7.

 

[Arachnotronic]

On Geekbench Cortex-A72 is significantly better than Apple A7.

Not in floating point perf...

 

[Nothingness]

Not in floating point perf...

It still leads

Looking at the individual results it looks like Apple wins on vectorized benchmarks, which points at wider SIMD units than A72.

 

[paffinity]

Oh nice! Didn't realize there were benchmarks out yet. I don't think any of the geekbench benchmarks are vectorized, as they have to run on all hardware (no-neon, etc). If that were the case the A7 FP numbers would be much higher.

 

[paffinity]

On Geekbench Cortex-A72 is significantly better than Apple A7.

I think it's actually running at 2.0 GHz, not the 1.57 GHz listed. The MediaTek part has 2 A72 cores @ 2.0 GHz and 2 A53 @ 1.6 GHz. I really doubt these results are on the A53 cores.

 

[NTMBK]

Oh nice! Didn't realize there were benchmarks out yet. I don't think any of the geekbench benchmarks are vectorized, as they have to run on all hardware (no-neon, etc). If that were the case the A7 FP numbers would be much higher.

I thought Neon was compulsory with ARMv8?

 

[Tuna-Fish]

On Geekbench Cortex-A72 is significantly better than Apple A7.

As it's not a battery-powered device and is not a handheld, it might have significantly higher power limits.

We'll have to see A72 in a battery-powered device for real comparisons.

I thought Neon was compulsory with ARMv8?

It is.

 

[paffinity]

Sorry, was thinking about 32-bit code. Yes, standard v8 does require Neon.

If it is vectorized, it's very poorly done. A7 with the Accelerate library can push over 8 GFlops/s DGEMM per core, while geekbench reports less than 2. Obviously these libraries aren't available for Android, but the level of vectorization in these tests are very low. Looking at the A72 numbers, DGEMM: 1.35 GFlops/s is less than 1 flop per cycle and SGEMM: 3.01 GF/s is ~1.5 f/c.

 

[StrangerGuy]

It still leads

Looking at the individual results it looks like Apple wins on vectorized benchmarks, which points at wider SIMD units than A72.

The slight margin the A72 wins over the A7 in both ST/MT scores is dreadful for a SoC that still doesn't hit retail yet and the A7 is already over 2 years old by this point while only being dual-core.

BTW my Xiaomi phone with Snapdragon 400 feels amazingly quick despite only a ST/MT score of ~450/1500 pts. Software is a much more potent bottleneck than the SoC in Android.

 

[Mondozei]

Huawei is about to unveil it's Kirin 950 which has 4 A72's in it.