> SuperMicro sells a 2U machine with 4 full sized computers in it with 6 drives. You can put 12 cores in that
i don't think AMZN or Google put 4 sockets per 1 U. While obviously it is possible, it would be pretty dense for normally powered chips. My understanding is that 2 socket/1U is more typical. At 100W/chip it would be a difference between 8KW and 16KW heat from a rack. But anyway, you're right - AMZN probably gets more than $1/1U. Their cluster 2-socket (16 real, 32 v-cores) instances go for $3.5/hour. Starting from that number we can adjust it down somewhat for pre-paid/reserved/spot pricing. Anyway, lets hope fuzziness in one number is compensated (not amplified :) by fuzziness in another ... Actually from the 2013 link you posted in the response below (that one - https://storageservers.wordpress.com/2013/07/17/facts-and-st...):
"Amazon data center- Amazon has around 450,000 servers in its data centers sited in 7 locations in the world. "
So adjusting for 1+ years of growth (of at least 10-20% as they had 49% revenue increase) my original estimate was less than 2 times wrong :)
Edit: actually adding storage servers - may be up to 4 times wrong if they have all 1U servers, so the most probable parameter to adjust would be $/1U to something like $2+/1U.
With the photos released by Google/MSFT/FB and OpenCompute specs they are putting ~60-96 "1U equivalent" boxes per rack (each with 2x E5-2600 class CPUs). Getting datacenter densities to 25kW/rack requires a bit of extra work (Chimneys, containment, etc) but the big guys do this already.
It's totally possible to get to 80 general purpose compute servers to a rack, with a power draw of around 15kW. I prefer the Supermicro FatTwin, but there are other options with similar density.
It usually isn't worth going to a mega dense rack (25kW or whatever), because unless you built the entire site with this in mind, you'll probably run out of power before you run out of space. Then you spent extra on these extreme density solutions for no real return. It can still make sense, if the entire solution is thought out.
Please don't mistake my engagement for dismay. I enjoy your analysis.
>> SuperMicro sells a 2U machine with 4 full sized computers in it with 6 drives.
>You can put 12 cores in that
In fact, the quad node contains 4 motherboards. Each motherboard supports 2 processors. So that's 8 processors in 2U, or 4 processors per U.
Each E5-2600 processor could support 8 'real' cores, or 16 Amazon Virtual cores. 4 * 16 = 64 Amazon cores in 1RU.
I agree that our math won't be right, and I think you're in the ballpark. But, the densities can be 'hella' higher than one things with modern gear. It's only 70W for the 8 core processor at 1.8ghz, 135W for the 2.9Ghz version, so if we do 4 processors per rack unit, that's "only" 540W per rack unit. 60 Rack units is 32KW/RU - which is a lot but not impossible. I'm gently simplifying by assuming 0 watts for the rest of the chassis, which is approximately reasonable if you're building a CPU farm.
i don't think AMZN or Google put 4 sockets per 1 U. While obviously it is possible, it would be pretty dense for normally powered chips. My understanding is that 2 socket/1U is more typical. At 100W/chip it would be a difference between 8KW and 16KW heat from a rack. But anyway, you're right - AMZN probably gets more than $1/1U. Their cluster 2-socket (16 real, 32 v-cores) instances go for $3.5/hour. Starting from that number we can adjust it down somewhat for pre-paid/reserved/spot pricing. Anyway, lets hope fuzziness in one number is compensated (not amplified :) by fuzziness in another ... Actually from the 2013 link you posted in the response below (that one - https://storageservers.wordpress.com/2013/07/17/facts-and-st...):
"Amazon data center- Amazon has around 450,000 servers in its data centers sited in 7 locations in the world. "
So adjusting for 1+ years of growth (of at least 10-20% as they had 49% revenue increase) my original estimate was less than 2 times wrong :) Edit: actually adding storage servers - may be up to 4 times wrong if they have all 1U servers, so the most probable parameter to adjust would be $/1U to something like $2+/1U.