Intel Edison is the size of a cracker with 2 CPUs, solid state storage and a gig of RAM. They size up to 3.5 x 2.5 x 0.4 cm and use <1W. Compared to an HP Proliant server at 68.20 x 44.80 x 4.32 cm, one could layer some Edisons 19 x 17 x 10 all sandwiched together in the same space making for 3230 units, or 6460 CPUs, 3.2 TB RAM and ~13TB of storage space, completed modularly and distributed. Of course, a Poweredge would only use 500W while this setup would use over 3kW. But imagine even 500 to get keep wattage in check: that's still 2TB storage, 1000 CPU cores and 500GB RAM.
To compare more fairly, a Xeon chip of a Proliant server would be about 53kMIPS, and at 615MIPS per Edison, that's 307kMIPS. That's right: 6x the the processing power...and 4x the cost.
The glory here is that unlike other SOC's, the Edison isn't ARM, but x86, meaning you can run a lot more linux applications (or Docker containers?) than one could with ARM. The performance seems to be a lot better, but that may be because of OS optimizations since x86 linux has been around and under development longer. The idea here is: the modular hardware makes it easy to distribute storage and processing power, and replace bad $39 units as they burn out with minimal impact to the overall cluster. Provided tasks and services can be written for a distributed system, it would be quite the cheap setup (and silent as it has no fans).
 |
| Edison Beowulf http://lcamtuf.coredump.cx/edison_fuzz/ |
To compare more fairly, a Xeon chip of a Proliant server would be about 53kMIPS, and at 615MIPS per Edison, that's 307kMIPS. That's right: 6x the the processing power...and 4x the cost.
The glory here is that unlike other SOC's, the Edison isn't ARM, but x86, meaning you can run a lot more linux applications (or Docker containers?) than one could with ARM. The performance seems to be a lot better, but that may be because of OS optimizations since x86 linux has been around and under development longer. The idea here is: the modular hardware makes it easy to distribute storage and processing power, and replace bad $39 units as they burn out with minimal impact to the overall cluster. Provided tasks and services can be written for a distributed system, it would be quite the cheap setup (and silent as it has no fans).
