Samsung SoC & ARMv8 discussions

[Nebuchadnezzar]

The 1.5 GHz is for the little cores (core0), apps are still too stupid to handle and properly read multiple clock planes and people and sites always jump to conclusions.

From what I know the chip is clocked at 2.1GHz on the A57's.

 

[Erinyes]

First 7420 numbers

http://browser.primatelabs.com/geekbench3/compare/1714816?baseline=1780313

Performance looks very good...especially the memory performance (assuming that they have the same LPDDR4 memory speed). Also..any idea why the AES and SHA single core scores are so much higher for the Exynos (200% and 50% higher respectively)?

the single-score values are IMHO rather high. Could that be true? Was the CPU really clocked at only 1.5GHz?

If I compare the results with the Nexus9 and the Apple6 SoC's then ARM seems to have improved the single-core values quite a lot with the A57. Link http://www.phonearena.com/news/Nexu...gra-K1-outperforms-Apple-iPhone-6s-A8_id61825

If true, then the Single-Score value of the Samsung 7420 is clock-for-clock higher than for Nvidias Denver core which could be the reason why Nvidia will not use the Denver-cores for the next gen.

That 1.5 Ghz clock should be for the A53s. The A57s should be at ~2-2.2 Ghz. So clock-for-clock, Denver still beats A57..but Denver's problem is power consumption and not performance. Also FWIW, rumours say Nvidia will use Denver in its Parker SoC.

 

[Gubbi]

Also..any idea why the AES and SHA single core scores are so much higher for the Exynos (200% and 50% higher respectively)?

Because it is easy to add crypto-hash execution resources to ensure Geekbench superiority.

Cheers

 

[Nebuchadnezzar]

Because it is easy to add crypto-hash execution resources to ensure Geekbench superiority.

Cheers

I'm not aware of any intermediary resource allocation in the ARM IP beyond either having the crypto accelerators or not. It just looks like throttling on the S810, the 5433 performs better.

 

[Alexko]

I'm not aware of any intermediary resource allocation in the ARM IP beyond either having the crypto accelerators or not. It just looks like throttling on the S810, the 5433 performs better.

By a factor of 2.83 for AES?

 

[juicytuna]

A lot of talk that this will be on Samsung's 14nm process. Isn't it a bit soon for that? I was under the impression that 14nm won't be ready till Q3.

 

[Nebuchadnezzar]

By a factor of 2.83 for AES?

We'll have to see once actual devices are out. Qualcomm has its issues with the S810 and we don't know if they're fixed in the Flex2 or not.

A lot of talk that this will be on Samsung's 14nm process. Isn't it a bit soon for that? I was under the impression that 14nm won't be ready till Q3.

It's on 14nm, I think the Q3 date was for external customers such as Apple.

 

[juicytuna]

I'm quite excited for the S6 all of a sudden. Hope they go all out on the GPU. MP8 + higher clocks at the very least.

 

[Erinyes]

Because it is easy to add crypto-hash execution resources to ensure Geekbench superiority.

Cheers

As Nebu said..I am not aware of it either. And the multi-core scores are much closer so it has to be something else.

By a factor of 2.83 for AES?

Exactly what I was wondering. And the multi-core score is only 1.39x so there is something odd going on. Even weirder is the fact that SHA single core is ~50% faster, but multi-core is ~10% slower :S

It's on 14nm, I think the Q3 date was for external customers such as Apple.

Bit of a surprise for me..was not expecting it to be on 14nm TBH. Would they be able to ship products in volume in Q2?

I'm quite excited for the S6 all of a sudden. Hope they go all out on the GPU. MP8 + higher clocks at the very least.

MP8 + higher clocks was my speculation as well..and if its on 14nm..should be quite possible.

 

[Gubbi]

As Nebu said..I am not aware of it either. And the multi-core scores are much closer so it has to be something else.

The multi-threaded scores are almost exactly four times the single thread scores. Either it runs with only four cores at a time (big.LITTLE ?) or each pair of cores share crypto hash resources.

Cheers

 

[Nebuchadnezzar]

Bit of a surprise for me..was not expecting it to be on 14nm TBH. Would they be able to ship products in volume in Q2?.

Given that's it's in the S6, I sure hope so.

The multi-threaded scores are almost exactly four times the single thread scores. Either it runs with only four cores at a time (big.LITTLE ?) or each pair of cores share crypto hash resources.

GeekBench is a 4-thread benchmark even on 8 core systems. ARMv8 crypto units are part of the CPU core itself.

 

[Laurent06]

GeekBench is a 4-thread benchmark even on 8 core systems.

Are you sure? Because both LGE and 7420 integer scores you linked have > x4 integer multicore results, x5 for the LGE and x4.4 for the Samsung.

ARMv8 crypto units are part of the CPU core itself.

Definitely correct.

 

[Gubbi]

Are you sure? Because both LGE and 7420 integer scores you linked have > x4 integer multicore results, x5 for the LGE and x4.4 for the Samsung.

The highest multi-core geekbench score is by a 60 core Xeon E7-4880 v2

AES multi score is 8.2 x single threaded on the LG.

The other multi threaded scores are in the 4-5 x single thread performance range which does look like heavy throttling. I just found the multi threaded AES and SHA scores being exactly four times the single thread scores odd.

Cheers

 

[Nebuchadnezzar]

Hrm. It may depend on the test. It's not first time I see a benchmark not dispatch full amount of threads on all of its subtests.

 

[Laurent06]

The highest multi-core geekbench score is by a 60 core Xeon E7-4880 v2

Yes, but here it's slightly different as we're talking about heterogeneous multi CPU. Also for Geekbench to detect the 8 CPU the 2 clusters should be publicized by the OS (which wasn't the case for the first big.LITTLE implementations).

AES multi score is 8.2 x single threaded on the LG.

The single core score is plain wrong (look at the Samsung results, or any other A57), either it was throttling, or the process scheduler put the process on one of the Cortex-A53 (this would also mean the 8 cores are available).

 

[Nebuchadnezzar]

The single core score is plain wrong (look at the Samsung results, or any other A57), either it was throttling, or the process scheduler put the process on one of the Cortex-A53 (this would also mean the 8 cores are available).

All cores are available, only Nvidia remains as last vendor with a cluster migration scheme where only 4 cores are visible.

Qualcomm uses a totally custom scheduler for big.LITTLE, so you might be onto something suggesting it may sit on the A53 cores in the single-thread benchmark, I didn't think of that (that's a bad thing).

 

[tuna]

All cores are available, only Nvidia remains as last vendor with a cluster migration scheme where only 4 cores are visible.

Qualcomm uses a totally custom scheduler for big.LITTLE, so you might be onto something suggesting it may sit on the A53 cores in the single-thread benchmark, I didn't think of that (that's a bad thing).

Where is the code for scheduler?

 

[Nebuchadnezzar]

https://www.codeaurora.org/cgit/quic/la/kernel/msm-3.10/log/kernel/sched?h=LA.BF.2.1_rb1.10
Though you want to read the dev branches for QuiC https://www.codeaurora.org/cgit/ but their branches are always a mess if you don't have a proper overview.

 

[Gubbi]

That's what you get when you take a NP-hard problem and make it harder.

Cheers

 

[Erinyes]

The multi-threaded scores are almost exactly four times the single thread scores. Either it runs with only four cores at a time (big.LITTLE ?) or each pair of cores share crypto hash resources.

Cheers

I was referring to the fact that while the Exynos's single core score was 2.83x the S810, the multi-core was only 1.39x.

An A57 and A53 core cant possibly be sharing crypto has resources. They both have their own independent resources built into the core. Now whether they're the same speed on both the A57 and the A53 is another question.

Given that's it's in the S6, I sure hope so.

I'm skeptical as well but they announced the start of volume production in early December. Yields probably aren't great but if they have enough wafers in production in time for Q2..it could still work. AFAIK this is the first time ever that Samsung has beaten TSMC to a node. After how far behind they were on 28nm..it is quite a surprise.