Stacking Signal TSV for Thermal Dissipation in Global Routing for 3D IC National Tsing Hua University Po-Yang Hsu,Hsien-Te Chen, TingTing Hwang ASPDAC’13.

Stacking Signal TSV for Thermal Dissipation in Global

Routingfor 3D IC

National Tsing Hua University

Po-Yang Hsu,Hsien-Te Chen,

TingTing Hwang

ASPDAC’13

Introduction

Motivation

Signal TSV Assignment and Relocation for Thermal Dissipation

Experimental Result

Conclusion

Outline

2

Introduction

Three dimensional (3D) chip stacking by Through-Silicon-Via (TSV) has been identified as an effective way to achieve better performance in speed and power [2, 3].

However, such solution inevitably encounters challenges in thermal dissipation since stacked dies generate significant amount of heat per unit volume.

3

Introduction Temperature aware 3D global routing algorithm by

inserting ”thermal vias” and ”thermal wires” to lower the thermal resistance[4]

Reduces the temperature at the cost of extra area of ”thermal vias”[1,6-10]

Performance and thermalaware Steiner routing algorithm to place signal TSVs to reduce temperature.[11]

Does not fully utilize the outstanding thermal conductance of TSV in thermal dissipation.

[12] proposed a stacked-TSV power network structure to improve thermal dissipation by fully utilizing TSVs in power network.

only employs stacked-TSV structure in power network.4

Motivation- Thermal model

5

The lateral thermal resistors Rlateral are determined by heat conductance of device material

Motivation

6

20um

Relationship between temperature and distance of stacked signal TSV to heat source

Motivation

7

Signal TSV Assignment and Relocation for Thermal Dissipation

8 Overall flow of placing signal TSVs in

global routing

Initial TSV Assignment

9

Initial TSV Assignment

10

PowDensityi,j,k : power density in grid (i,j,k) where i, j, k denotes coordinates of the grid node in x, y, z axis direction

high lumped power density grid needs more signal TSVs to dissipate its heat.

n : number of tiers in the design.

TSVNumi,j,k : number of signal TSVs in grid (i,j,k).

Initial TSV Assignment

11

SDi,j is defined as the stacking degree in grid (i,j), which is computed as the number of TSV stacking at grid position (i,j).

Larger Gain value means higher power density, less TSVs, and more stacking signal TSVs.

Initial TSV Assignment

12

Stacked-TSV Relocation Stage

13

Hotspot grids Identification

14

Hotspot grid is identified by the top 10% highest thermal criticality grids.

define a circle region to find its saver net.

Hotspot grids Identification

15

Use a matching algorithm to find the overall best solution.

GridDist is the summation of distance from hotspot grid to the nearest TSV of the saver net n in all tiers.

wiring overhead if we stack the TSVs of saver net n close to the grid g.

Weighted graph G = ( H∪S, E)

H S

Hotspot grids Identification

16

Use a matching algorithm to find the overall best solution.

StackingDegree is the number of tiers that a saver net crosses.

heat dissipation ability

Weighted graph G = ( H∪S, E)

H S

Determination of Stacking Grid

Based on the matching solution, TSV of a saver net will be relocated near the hotspot grid.

However, there are other factors to determine if a grid location is the best choice.

Define candidate target grids which are hotspot grids and the adjacent grids nearby them to determine the best target grid location for moving signal TSV.

17

Determination of Stacking Grid Gain function to select our target grid to place stacked

signal TSV at grid (i, j) is defined as

Consider

Distance between candidate target grid and hotspot grid

Power density

Number of TSVs

Whitespace

Wirelength

18

Determination of Stacking Grid Gain function to select our target grid to place stacked

signal TSV at grid (i, j) is defined as

Consider

Distance between candidate target grid and hotspot grid

The larger DSST the closer the distance between stacking location to the hotspot grid. 19

Determination of Stacking Grid Gain function to select our target grid to place stacked

signal TSV at grid (i, j) is defined as

Consider

Power density

High power density grid needs more stacked signal TSV to dissipate its heat.

20

Determination of Stacking Grid Gain function to select our target grid to place stacked

signal TSV at grid (i, j) is defined as

Consider

Number of TSVs

When TSVi,j,k is larger, fewer number of TSVs is in grid (i,j,k).

21

Determination of Stacking Grid Gain function to select our target grid to place stacked

signal TSV at grid (i, j) is defined as

Consider

Whitespace

22

Determination of Stacking Grid Gain function to select our target grid to place stacked

signal TSV at grid (i, j) is defined as

Consider

Wirelength

Wirelengthi,j,k is the wirelength overhead in tier k if stacking location is at grid (i,j).

smaller value of WL denotes higher wiring overhead.23

Move signal TSVs to the same 2D location across all tiers will change the routing topology and increase wiring overhead.

Experimental Result

2005 IWLS benchmarks [20] and industrial circuits.

3D placement results are produced by a partitioning driven placement for 3D ICs [5].

minimize the total wirelength and signal-TSV count

24

Experimental Result

25

S.TSV : Total # of Stacked TSV

Extra hardware overhead !!!

Conclusion

A new integrated architecture, stacked signal TSV, was developed to dissipate heat.

Based on this structure, a two-stage TSV locating algorithm has been proposed to construct the stacked signal TSVs and fully utilize the TSV thermal conductance to optimize the chip temperature.

Compared to previous thermal-TSV insertion method, our proposed algorithm has zero hardware overhead incurred by thermal-TSV.

26