Dimensionally-Decomposed Router for 3D-NoC*

1. Dimensionally-Decomposed Router for 3D-NoC* J. Kim et al., ISCA 2007 ECE 284 5/14/2013 Presenter: Hyein Lee *J. Kim et al., "A Novel Dimensionally-Decomposed Router for On-Chip Communication in 3D Architectures”,"Proc. ISCA,pp. 138-149, June 2007.

2. Outline • Introduction • Current 3D NoC Architecture • Dimensionally-Decomposed Router • Performance Results

3. Introduction: NoC for 3D Chips • 3D chip technology • Reduce interconnect delay by stacking multiple layers • NoC for 3D chips • Interconnect architecture design across the layers should be carefully designed • Needs an integrated approach for the network in 2D plane and the vertical interconnect  The interconnect delay of inter-layer is much smaller than that of intra-layer (inherent asymmetry) • This work • A new 3D dimensionally-decomposed router architecture

4. Current 3D NoC Architecture • 3D Symmetric NoC Architecture • 3D NoC-Bus Hybrid Architecture • True 3D NoC Router A Generic NoC Router

5. Current 3D NoC Architecture • The 3D Symmetric NoC Architecture • Simple extension to 3D: 5x5 crossbar in typical 2D router + inter-layer movement (hop-by-hop traversal) • Does not exploit the low inter-layer distances • Requires larger crossbars (7x7)  2X power of 5x5 crossbar A 3D Symmetric NoC Network PE : Processing Elements (CPUs, DSPs, etc.)

6. Current 3D NoC Architecture • Area and power comparison of various crossbar types

7. Current 3D NoC Architecture • The 3D NoC-Bus Hybrid Architecture • Uses a bus link in the vertical dimension • However, it does not allow concurrent communication in the vertical dimension • Bus is a shared medium; can be used by a single flit only

8. Current 3D NoC Architecture • A True 3D NoC Router • Add vertical links to allow flexible flit traversal within the 3D crossbar • Dedicated connection box(CB)s at each layer • 3D connection box can facilitate linkage between vertical and horizontal channel Connection Box

9. Current 3D NoC Architecture • A True 3D NoC Router • Increase path diversity  good for optimization but hard to control • Requires excessive control signals Impact of the number of vertical bundles on performance  use of 4 vertical pillars is not efficient Number of minimal paths between A and B 3x3x3 crossbar, k=90 4x4x4 crossbar, k=1680

10. Dimensionally-Decomposed Router: 2D RoCo • 2D Row-Column (RoCo) Decoupled Router • Incoming traffic is decomposed into two independent stream (Guided Fit Queuing*) • X dimension (East-West traffic) and Y dimension (North-South traffic) • Use two smaller 2x2 crossbars instead of 5x5 *J. Kim, et al., “A Gracefully Degrading and Energy-Efficient Modular Router Architecture for On-Chip Networks”, Proc. ISCA, 2006

11. Dimensionally-Decomposed Router: 3D DimDe • 2D RoCo + vertical links • Segmented vertical links are integrated into row/column module

12. Dimensionally-Decomposed Router: 3D DimDe • Vertical link arbitration • 2-stage arbitration (in deterministic XYZ routing) • Local arbitration (intra-layer)  global arbitration (inter-layer) • Concurrent communication is needed to increase the bandwidth • 3D global arbitration request scenarios (1), (5), (9) combination can allow full concurrent communication

13. Dimensionally-Decomposed Router: 3D DimDe • Guided Flit Switching • Decompose incoming flits into x, y, and z direction • Early Ejection Mechanism • Enable flits to bypass the destination router and be ejected to the NIC • Vertical module • Two bidirectional vertical bundles

14. Performance Results • The proposed method remains within 5% on average of the performance of the full 3D crossbar (ideal) • 20% improvement in latency over other 3D crossbars (without Full 3D); 26% reduction in Energy-Delay Product

15. Thank you!