SMP Linux

[darence]

Hi guys,

Based on the information I could find about Parallella, I infer that the 16 cores are not able to run a single OS instance (SMP Linux for example), but each core will run its own. Inter-core communication is then mainly through the on-chip memory, abstracted away using established interfaces (such as MPI). In a nutshell, the way the system is programmed very much resembles the Intel SCC.

Can you please confirm (or refute) this?

[Gravis]

it's possible to have one core route to all the other cores (typically how SMP works) but do remember that the cores have strict on-chip memory limitations which would cause a sizable penalty for memory IO compared to recent processors with MBs of on-chip memory.

that aside, the highlight of the core design is it's fast floating point instructions. the highlights of the processor as a whole is it's ability to have many programs operate completely independently but also able to communicate between cores at an extremely fast rate. you can use this to run many independent programs or coordinate them to accelerate algorithms similar to how GPUs work. a huge advantage over GPUs is that branching is not a huge bottleneck which will do wonders for computer vision. i think it's most suited for processing streams of data in realtime which is something typical processors struggle with.

[darence]

[CIB]

[darence]

[9600]

[CIB]

[Gravis]

[darence]

[Gravis]