On 11/11/13 13:32, Matt Doidge wrote:
> On 11/11/2013 12:49 PM, Stephen Jones wrote:
>> On 11/07/2013 08:45 AM, Jensen, Jens (STFC,RAL,SC) wrote:
>>> They get fewer events per minute per core than on Xeons, but more
>>> events per minute per Watt.
>>
>> I wonder how XEONs and Pi's stack up. It would be good to compare these.
>> Have you any typical events per minute per watt values to plug in?
>>
>> If we convert the ARM's events per minute per watt to events per KWH, we
>> can use the cost of 1 KWH (£0.10) to get "events per £" or the energy
>> cost
>> of one event etc. It would be good to know if there is a huge difference.
>
> *Spoiler alert for those who haven't read the paper[1] yet - all I'm
> doing is regurgitating[2] the numbers here.*
>
> The article has figures of 1.14 events/minute/watt for arm CPUs, and
> 0.21 events/minute/watt for hexacore xeons. The power consumption is
> estimated from the TDP of the CPUs, so isn't fully indicative of total
> power draw of the servers doing their thing (particularly spinning all
> those disks), but it's a good place to start.
>
> Converting to KWH to gives 68,400 arm events per KWH, and 12,600 Xeon
> events/KWH.
>
> Of course, the ARM events/minute/core were roughly a third that of the
> Xeons, and time is money, but this argument could be counter-countered
> by just how many ARM chips manufacturers could shoehorn into specialist
> boxes.
>
> Definitely worth keeping an eye on.
>
I don't know the status of 64 bit ARM, but that may be particularly
interesting.
I shall be chatting with some guys at ARM tomorrow and over the weekend.
Is there anything people want me to ask?
Chris
> Cheers,
> Matt
>
> [1]http://xxx.soton.ac.uk/abs/1311.0269
> [2]I also apologise for using the word regurgitate. It's not a pleasant
> word.
>
>>
>> Cheers,
>>
>> Steve
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>> Events per minute per watt is a measure of how much work a watt can do
>> in a minute (or how much work a minute can do with a watt!)
>>
>>
>>
>>
>>
>>
|