Re: ARM 64?

[Michael Tremer <michael.tremer@ipfire.org>]

[-- Attachment #1: Type: text/plain, Size: 3227 bytes --]

Hey Matt,

On Mon, 2018-05-28 at 20:32 +1000, Mathew McBride wrote:
> Hi Michael,
> 
> Just in response to your questions:
> On 25/5/18, 11:10 pm, "Michael Tremer" <michael.tremer(a)ipfire.org> wrote:
>     
>     
>     I think you hardware is good enough for a builder. But I still am not sure
> what
>     to expect from the CPU. It will be faster than a Raspberry Pi, but not a
>     Mustang.
>     
> We did some benchmarks with the Phoronix test suite a while ago, this will
> give you an idea:
> http://openbenchmarking.org/result/1708303-TR-
> 1703199RI93&obr_hgv=Traverse+LS1043+Prototype

I had a look at that. And yes indeed, it is a bit hard to figure out the
performance by the CPU name alone for most ARM SoCs. There is no branding in
order of performance (or similar) like Intel has.

That might actually turn out to be a bigger marketing problem, but we will see
that in the future.

> To give an idea of the Cortex (ARM designed)-based core performance:
> 
> The LS1043 has the same A53 cores as the RPi3, but performs better due to
> having more cache, DDR4 etc (and higher clock).

Performance is also coming from the rest of the periphery. The RPi has a slow
and not very stable USB bus to talk to the network to and SD card storage. Even
with a faster CPU it might very often just wait for data.

We have been trying to tell people that they should look out for some specific
features like cache and good single-core performance.

> A72 is about double A53 in performance (and power consumption!) per MHz, as
> A72 is a modern out-of-order speculative core (it did get hit with the
> Meltdown/Spectre issue).

Yes, wouldn't mind to have some systems based on that one since the A53 will be
too slow for really large enterprise deployments.

> The latest gen of ARM64 server cores would all be well above A72, your Mustang
> is probably around the A72 level.
> 
> In general, ARM network SoCs try to work 'smarter' instead of 'harder', so the
> high network performance comes from having very good network silicon, taking
> advantage of crypto accelerators etc.

I prefer the NICs in the SoC which gives great performance. The disadvantage
only is that they sometimes to odd configurations like 5x 1G and 1x 10G in this
case which I don't really understand. The only use-case that makes sense to me
is a server but for that the CPU is too slow and people would probably go for a
A72-class CPU.

>     > There is a TrustZone firmware running in the ring/EL above the OS, for
> the NXP
>     > Layerscape/QorIQ SoC's this firmware is open source, and not strictly
> required
>     > to run the system (it gets loaded by u-boot after power on).
>     
>     What does the firmware do?
> It implements some vendor-specific power-management extensions (PSCI), as well
> as some TPM-like functions.
> NXP provides a good overview: https://github.com/qoriq-open-source/ppa-generic
> /blob/integration/ReleaseNotes.txt
> I am not a security expert, but it could be a good test environment for secure
> boot, private key storage and other things.

Great that this is entirely open.

-Michael

> 
>     
> Cheers,
> Matt
>  
> 
>