[Beowulf] ARM cpu's and development boards and research

[Vincent Diepeveen]

On Nov 27, 2012, at 9:05 PM, Bill Broadley wrote:

> On 11/27/2012 07:46 AM, Vincent Diepeveen wrote:
>> i dug around in price of ARMs and development boards.
>>
>> If you just buy a handful most interesting offer seems to be
>>
>> http://www.hardkernel.com/renewal_2011/products/prdt_info.php?
>> g_code=G133999328931
>>
>> it's $129 and has a quad core ARM cortex A9 at 1.4Ghz and a mali gpu
>> on it.
>>
>> So from raw cpu cores this is by far best offer on the market,
>> nothing even gets close to it,
>
> Maybe if you want the maximum cores per $ regardless of performance.
>
> Atom boards are likely to give you:
> * substantially better CPU performance
> * 10x the network bandwidth
> * 4x the memory size
>
> Like say the $70 board with a D2500 (1.86GHz dual core) or the $80  
> D2700
> (2.1 GHz dual core/4 thread).

Atom D2700 is 10 watt TDP and with intel you never know whether they  
over that or not.

That's really too much man for a robot that runs on a battery and  
wants to get a beer for you from the fridge while you watch
your favourite (sport)show on TV.

Also it's price of $52 a piece is a tad high.

>
>> The number of quadcore offers is pretty limited in ARM world - it's
>> nearly all dual cores there.
>
> Umm, you seem to be ignoring quite a few quad core arm chips.  Like  
> say
> the nexu s4, nexus 7, nexus 10, and many others.  Both the tegra quad
> core and the Snapdragon quad core seem rather common.
>

it's about the price of the entire kit. This thing is $129 including  
everything.

If you go negotiate with intel you won't get a cheaper price simply  
than the $52 price they sell it for, if you buy a 1000.

If you go toy yourself with quadcore ARM you can get them a lot cheaper

I have no idea which of the 2 cpu's is faster for artificial  
intelligence software / navigation codes.

That's not bandwidth limited code, it's branch mispredicted penalty  
limited code and in general it's
lots of codelines, so a small L1i is a big disadvantage as well.

The real problem with intel is not their slowness - they're fast  
usually.

It's the price of the cpu's. If you buy a 100, you still are nowhere  
and pay a $100 or so.

If we pick an atom of the same TDP : http://ark.intel.com/products/ 
58916/Intel-Atom-Processor-N2600-1M-Cache-1_6-GHz

We get to the 3.2 watt TDP N2600.

http://www.ebay.com/itm/Jetway-NF9C-2600-Intel-Atom-N2600-Low-Profile- 
Mini-ITX-MB-w-Onboard-Power-HDMI-/110821875068? 
pt=Motherboards&hash=item19cd7f9d7c

On ebay $150 for a board.  Seems it's without RAM yet.

Even if the board would not draw too much power in total under full  
load, we still have to face that this is unsellable
as a robot product.

$200 in total for motherboard+cpu+adapter.

That's nuts as that means a customer price of $800 just based upon  
the CPU and
there is more components than this...


>> This CPU should be eating at 1.0Ghz exactly 3 watt under full load
>> and a tad more at 1.4Ghz,
>> though the "1.4Ghz" clock i couldn't confirm in an easy manner. Could
>> be the quadcore version is 1.0
>> Ghz after all.
>
> Sounds like hand waving to me, have you actually measured this?  Were
> all cores and network in use?  Was it diskless?  ssd?  sata?

Maximum power draw (measured) of the 1 Ghz quadcore ARMs is 2.9 watt.

That's already a lot for battery driven equipment.

Most of the handheld devices the cpu is under 1 watt.

>
>> The cortex A15 that goes in production in 2013 maybe, that one is 64
>
> Er, the cortex-a15 is 32 bit, it has a PAE like extension to allow  
> bank
> swapping.  You probably mean the cortex a57 which is 64 bit (and not
> shipping)

Thanks for the information.

Note for AI that 32 bits or 64 bits is no big deal.

(for specific chessprograms it is - but realize they didn't do effort  
to optimize for 32 bits then usually
and just cut'n pasted codes from others whereas my move generator  
which i threw in GPL,
and which is fastest move generator on the planet in software,  
they're too lazy to use it as then they
would really need to code up themselves an entire program).

>
> The cortex a15 however shipping in tablets (the nexus 10), laptops  
> (the
> chrome book), and motherboards (the Arndale).
>
> The A15 benchmarks look pretty good compared to the A9, TI has an OMAP
> 5, Samsung has the Exynos 5250.
>
> My interest in Arm has been the flip side of balancing flops to  
> network
> bandwidth.  A standard dual socket (AMD or Intel) can trivially  
> saturate
> GigE.  One option for improving the flops/network balance is to add
> network bandwidth with Infiniband.  Another is a slower, cheaper,  
> cooler
> CPU and GigE.
>
>> A big hurdle always is the many dual core ARMS out there which do not
>> have fast RAM yet which only have
>
> I believe the cortex a15 substantially improves memory bandwidth.

Memory latency always is important.
DDR3 always is a huge improvement over DDR2.

>
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin  
> Computing
> To change your subscription (digest mode or unsubscribe) visit  
> http://www.beowulf.org/mailman/listinfo/beowulf