Edward Ned Harvey <[email protected]> writes:

> We're talking a whole new scale here.  512 atom processors in a 10u
> formfactor, consuming 2kw, for $200 each.  Looking quickly at dell, it's not
> difficult to get down to $200 per core, 4 cores in 1U.  But you're going to
> fit at most 40 cores into 10u, and it will still consume 2kw of power at
> that scale.  


http://www.supermicro.com/Aplus/system/1U/1042/AS-1042G-TF.cfm

that thing supports 4 12 core opterons in 1u.   48 cores per U.  and something
like 6.6 watts per core, assuming the 80 watt ACP.  if the atom is 4 watts per 
core, I bet you are getting more compute per watt out of those opterons.  

It's not expensive, either.  I have a similar system right now.  Now,
I can't get very much more than 3Kw usable per rack around here anyhow,
so density doesn't matter that much for me, and I'm fairly I/O bound,
so I like having lots of disks, so I opted for a dual-socket system with
8 disks:

http://www.supermicro.com/Aplus/system/2U/2022/AS-2022G-URF.cfm

Now I'm a cheap bastard, so I found some of those chassis used, and
I bought a new power supply, new motherboard, and new backplane.  Total
cost for those parts?  under six hundred bucks.  add in the cost of ram,
reg. ecc ddr3 is around $130 for a 4GiB stick on newegg business right now,
and CPU, the 12 core seems to start around $700 each, and the 8 core, 
2Ghz start around $300, and you can see that my costs blow seamicro
away, even if I charge $500 for the hour and a half it takes me to 
put this together and setup/deal with burn in.  (btw, if anyone wants 
me to assemble them one, order me the parts, and I'll do it for $500.)


> The root concept is:  Break away from the assumption of xeon or equivalent
> amd processors.  Jump down to the super small, super low-power, super cheap
> class of processors, atom, arm, etc... and use them to beat the xeons for
> some situations, such as distributed work load, or server virtualization.

This is a good evaluation to step back and make every few years.  
It's especially important for me, because my customers will actually
pay me more for a dedicated server than for a virtual server of the
same specifications.  

There certainly is an 'optimal size' for servers, and if you go over
that or under that you end up paying more per compute resource than
you would otherwise, so if you are as cost conscious as I am, it's 
/very important/ to remain aware of where this optimal size is.

This optimal size does vary a lot if you are say, cpu bound vs. ram
bound, vs I/O bound.  But, I think the atoms don't really compete
with the opterons in any way (unless you are optimizing for /isolation/
rather than performance.) 

> The cost to purchase is approximately equal (as far as I can tell) but the
> power and density are improved by an order of magnitude each, compared
> against "standard" servers.

The capital cost of the atoms only look good if you think one atom
core is worth one xeon or one opteron core, which is simply not the case.

If someone comes out with an ARM based board with socketed ram, I'll
re-run my evaluation.  But for now, as far as I can tell, AMD has
the atom whooped in terms of capital costs, and competes in terms of
power costs.   (Xeon, as usual, takes the crown if you have lots of money,
power, and you care mostly about per-core performance.) 
_______________________________________________
Tech mailing list
[email protected]
http://lopsa.org/cgi-bin/mailman/listinfo/tech
This list provided by the League of Professional System Administrators
 http://lopsa.org/

Reply via email to