[LLVMdev] Build times on ARM

C Bergström cbergstrom at pathscale.com
Wed Jun 17 02:06:05 PDT 2015


On Wed, Jun 17, 2015 at 1:26 PM, Renato Golin <renato.golin at linaro.org> wrote:
> On 17 June 2015 at 04:29, Rafael Espíndola <rafael.espindola at gmail.com> wrote:
>> Does anyone know of a even faster ARM dev board?
>
> Hi Rafael,
>
> That's as fast as ARMv7 will ever get, I believe. I'm using Chromebook
> 2's quad-A15+2GB and it's about the same speed. Qualcomm's IFC6410 are
> meant to be as fast as Chr2s, but the kernel is currently not
> scheduling frequency due to power management failure.
>
> On the AArch64 side, Chris is right that the options are far better!
> Cavium's offering may be fast, but are not available yet. And when
> they do, they're going to be *very* expensive.

Sorry, but I don't believe this is true. It's my understanding that
most ARM partners plan to price their chips and server competitively
against market leader (Intel). So unless you think Intel servers are
*very* expensive, then it's not fair to make that statement - unless
you have specific pricing information.

While not publicly available - Early access partners and potential
customers do have systems. (Now and more being delivered soon) -
Developer kits almost always cost more unless the manufacturer
subsidizes or sells as a loss leader.


>
> Junos are fast-ish, AMDs are faster but the APMs are the fastest
> boards you can buy today, and you can run on AArch32 mode and build
> ARMv7 binaries with it. Though, you'll have to shell out a couple
> thousand dollars for it.

APM's are fastest at what? Single core performance, memory bandwidth,
total socket performance, floating point.. I agree they are a
respectable chip, but when only a tiny amount of ARMv8 chips can be
purchased today - it's like saying fastest of 1 out of 2. If you
factor into chips which are available to developers - then they are
not the fastest in most areas. It's not meant to be though - look at
how much power they consume. I don't mean to take this discussion the
wrong way. I just get itchy when statements are made without
supporting evidence. I have 1st hand access to all mentioned systems
and if it wasn't for NDA's - I'd disclose details.

>
> If this is for personal use, I'd stick to either the Chromebook2 or
> the Snapdragon (or the NVidia). If you need a bit more oomph and
> stability, APMs are the ones to go.

Personally I'd hold out for the NVIDIA Shield tv to get ported or
other parts targeting the higher side of server market. With the
Shield tv you also get a GPU that could be fun to play with...

>
> cheers,
> --renato
>
> Ref:
>
> IFC 6410 - ARMv7 Quad-Krait 2GB
> http://www.inforcecomputing.com/products/single-board-computers/6410-single-board-computer-sbc
> Available, $150
> (*) Kernel needs fixing, hopefully soon (months)
>
> Chromebook2 - ARMv7 Quad-Exynos5 2GB
> http://www.samsung.com/us/computer/chrome-os-devices/XE503C32-K01US
> Available, $380
>
> ARM Juno - AArch64 Hexa-A53/A57 up-to 8GB
> http://www.arm.com/products/tools/development-boards/versatile-express/juno-arm-development-platform.php
> Hard to get, $100s to $1000s
>
> APM X-Gene - AArch64 Octa-A57, lots of RAM
> https://www.apm.com/products/data-center/x-gene-family/x-gene/
> Hard to get, $100s to $1000s
>
> AMD Seattle - AArch64 Quad~Octa, lots of RAM
> http://www.amd.com/en-us/press-releases/Pages/64-bit-developer-kit-2014jul30.aspx
> Not available, likely many $1000s
>
> Cavium Thunder-X - AArch64 48/96 cores, lots of RAM
> http://www.cavium.com/ThunderX_ARM_Processors.html
> Not available, likely many many $1000s

I wouldn't speculate pricing without 1st hand details. Lets be honest.
On the ARMv8 side - NVIDIA Sheild tv (sorry to repeat this), Nexus 9??
and Apple iOS devices are the only ones I know which are probably easy
to get your hands on. Everything else is pretty much given to
developers and partners 1st. (APM's may fall into a gray area..)
------------
I think this thread digressed - you said you wanted to know what was
faster.. That's really all I was trying to relay.




More information about the llvm-dev mailing list