3. DESIGN
Our design—the NoCMsg layer—is driven by the objective to create a close to baremetal
NoC-level messaging protocol. Since this requires low-level NoC capabilities
to be used, one becomes exposed to wormhole routing problems, such as potential
deadlocks. Our NoCMsg design ensures absence of deadlocks with reduced flow control
to lower overheads within its protocol layer.We assume a generic, generalized 2D mesh
NoC switching architecture similar to existing fabricated designs with high core counts,
which is a viable solution for future microprocessor design. Notice that even 3D stacking
of memories still assumes a single silicon layer of processing cores at the top of the stack
due to thermal constraints, likelywith a 2D NoC mesh to ensure scalability. Each core is
composed of a compute core, network switch, and local caches. Communication over the
NoC is assumed to be reliable, lossless, and without duplication. We assume wormhole
Manhattan-path (XY dimension-ordered) routing over the NoC, but the method is
applicable to any static routing scheme. (Dynamic routing is not likely to become
feasible for NoCs due to single cycle routing requirements.) We discuss its relation to
our NoCMsg layer design and describe constraints of such an architecture next.
3. DESIGNOur design—the NoCMsg layer—is driven by the objective to create a close to baremetalNoC-level messaging protocol. Since this requires low-level NoC capabilitiesto be used, one becomes exposed to wormhole routing problems, such as potentialdeadlocks. Our NoCMsg design ensures absence of deadlocks with reduced flow controlto lower overheads within its protocol layer.We assume a generic, generalized 2D meshNoC switching architecture similar to existing fabricated designs with high core counts,which is a viable solution for future microprocessor design. Notice that even 3D stackingof memories still assumes a single silicon layer of processing cores at the top of the stackdue to thermal constraints, likelywith a 2D NoC mesh to ensure scalability. Each core iscomposed of a compute core, network switch, and local caches. Communication over theNoC is assumed to be reliable, lossless, and without duplication. We assume wormholeManhattan-path (XY dimension-ordered) routing over the NoC, but the method isapplicable to any static routing scheme. (Dynamic routing is not likely to becomefeasible for NoCs due to single cycle routing requirements.) We discuss its relation toour NoCMsg layer design and describe constraints of such an architecture next.
การแปล กรุณารอสักครู่..