JP7239099B2 - TSV Error Tolerant Router Device for 3D Network-on-Chip - Google Patents

Description

The present invention relates to a TSV error tolerant router device for 3D network-on-chip.

Through Silicon Vias (TSVs) serve as inter-layer wires in 3D networks on chips (3D-NoCs) as well as 3D-integrated circuits (3D-ICs). A TSV is established by forming a via, thinning the wafer, performing thermal compression, and connecting the two wafers through the TSV. TSVs are usually formed into groups, or clusters, on a regular basis, or irregularly in random locations.

Although TSV offers many advantages over 3D-NoCs, one of its major drawbacks is reliability. The productivity of 3D-NoCs with TSVs is considered a critical factor due to imperfections in the manufacturing process.

The defect rate of TSV is reported to be 0.63% (Non-Patent Document 1). In addition, 3D-NoCs are subject to stress based on differences in thermal expansion coefficients between materials of manufacture. It has been reported that the temperature change between the two layers can reach 10°C, adversely affecting Time Dependent Dielectric Breakdown and Thermal Cycling. It should also be remembered that electric field transfer is also of major interest.

As a result, TSVs in 3D-NoCs are more prone to defects not only during manufacturing, but also during operation. Therefore, the need for defect tolerance in 3D-NoCs during and after fabrication is inevitable.

TSV defects can be classified into three types: open (empty), bridging and stuck-at faults. An open defect occurs when a TSV is broken and both ends are electrically disconnected. A bridging defect appears when two or more TSVs are connected. As a result, these TSVs cannot convey different values. A stuck-at fault shorts TSV to ground or Vdd where the output is always '1' or '0'. If the TSV is partially defective, excessive delay will occur and the time requirement will not be met.

Previous studies have addressed the high defect rate of TSVs with different approaches: manufacturing process refinement for reliability (2); accounting of potential defects during the design stage (3). ); support circuitry [4, 5]; correction of defective TSVs using redundant [5] or error code codes [6]; and avoiding defective TSV channels using alternate channels. (For example, the use of defect-tolerant routes in 3D-NoCs (Non-Patent Document 7)).

FIG. 1 illustrates a conventional method of using redundancy. To handle N defective TSVs, we need at least N redundant TSVs. In FIG. 1, (a) is for one defective TSV; D, (b) is for two defective TSVs; 2D, and (c) is for four defective TSVs; 4D. . We need as many redundant TSVs as there are defective TSVs to be detected.

JP 2014-179433 A JP 2016-082018 A WO2012/140810 WO2013/037048 U.S. Pat. No. 8,384,417

U. Kang et al., "8Gb 3D DDR3 DRAM using through-silicon-via technology", in IEEE International Solid-State Circuits Conference, 2009, pp. 130-131. J.U. Knickerbocker et al., "Three-dimensional silicon integration", IBM Journal of Research and Development, vol.52, no.6, pp.553-569, 2008. T. Zhang et al., “Temperature-aware routing in 3D ICs,” in Asia and South Pacific Conf. on Design Automation, Jan 2006, pp.309-314. M. Cho et al., “Design method and test structure to characterize and repair TSV defect induced signal degradation in 3D system”, in Proc. Int. Conf. on Computer-Aided Design, 2010, pp.694-697. Y. Zhao et al., “Online Fault Tolerance Technique for TSV-Based 3-D-IC,” IEEE Trans. Very Large Scale Integr.(VLSI) Syst, vol.23, no. 8, pp. 1567-1571, 2015 . D. Bertozzi, L. Benini, and G. De Micheli, “Error control schemes for on-chip communication links: the energy-reliability tradeoff,” Computer-Aided Design of Integrated Circuits and Systems, IEEE Transactions on, vol. 24, pp. 818-831, June 2005 A. Ben Ahmed, A. Ben Abdallah. “Graceful Deadlock-Free Fault-Tolerant Routing Algorithm for 3D Network-on-Chip Architectures”. I. Loi et al., "A low-overhead fault tolerance scheme for TSV-based 3D network on chip links", in Proc.2008 IEEE/ACM Int. Conf. on Computer-Aided Design, 2008, pp.598-602 . L. Jiang, Q. Xu, and B. Eklow, “On effective through-silicon via repair for 3-D-stacked ICs,” IEEE Trans. Comput.-Aided Design Integr. Circuits Syst, vol. 32, no. 4 , pp. 559-571, 2013. Y. Zhao et al., “Cost-effective TSV grouping for yield improvement of 3D-ICs,” in Asian Test Symp. IEEE, 2011, pp. 201-206. F. Ye and K. Chakrabarty, “TSV open defects in 3D integrated circuits: Characterization, test, and optimal spare allocation,” in Proc. 49th Annual Design Automation Conf. ACM, 2012, pp. 1024-1030. Y. J. Hwang et al., “3D Network-on-Chip system communication using minimum number of TSVs,” in 2011 Int. Conf. on ICT Convergence. IEEE, 2011, pp. 517-522. A. Kologeski et al., “Combining fault tolerance and serialization effort to improve yield in 3d networks-on-chip,” in 2013 IEEE 20th Int. Conf. on Electronics, Circuits, and Systems, Dec 2013, pp. 125-128 . Z. Qian and C. Y. Tsui, “A thermal-aware application specific routing algorithm for Network-on-Chip design”, in 16th Asia and South Pacific Design Automation Conf., Jan 2011, pp.449-454. M. Palesi et al., "Application-specific routing algorithms for networks on chip", IEEE Trans. Parallel Distrib. Syst., vol.20, no. 3, pp.316-330, Mar 2009. Y. Ghidini et al., “Lasio 3D NoC vertical links serialization: Evaluation of latency and buffer occupancy,” in 26th Symp. On Integrated Circuits and Systems Design, Sep 2013, pp.1-6.

Although the above previous work has impressively expanded the reliability of TSV-based systems, problems in fault allocation still exist. Most of the first studies performed refer to random allocation (8). And cluster defect allocation (Non-Patent Document 9) is currently considered as the most realistic one.

To deal with defects in cluster TSVs, many studies choose the optimal grouping configuration (Non-Patent Document 10) and place TSVs in different positions (Non-Patent Document 11) or increase the redundancy correction rate (Non-Patent Document 11). Reference 9) is aimed at.

While these methods can improve system reliability, adding significant redundancy and complex coordination results in penalties in area cost, wire delay and power consumption. As shown in Figure 1, when a cluster defect occurs, all TSVs of a group (4 TSVs) become defective. Therefore, a fault tolerant system requires twice as many TSVs to ensure connectivity. Furthermore, if the number of defective TSVs exceeds the number of allocated redundant TSVs, the cascade will be broken.

Better management solutions are therefore likely to help address this problem, especially for 3D-NoCs, where low utilization of TSVs has been reported (12, 13).

SUMMARY OF THE INVENTION It is therefore an object of the present invention to provide a TSV fault tolerant router device for 3D network-on-chip that maintains inter-layer connectivity with smart management and avoids adding large redundant TSVs.

One aspect of the present invention is a TSV fault tolerant router device for a 3D network-on-chip having a plurality of routers arranged in each of a plurality of layers, the routers between layers being connected by through silicon vias, , each of the through-silicon vias belongs to one corresponding router and has a plurality of clusters around the corresponding router, and when one of the plurality of clusters belonging to the corresponding router is defective, the corresponding neighboring routers are selected and one cluster of the selected routers is replaced in place of the defective cluster to maintain connectivity between the layers.

An example of a TSV error tolerant router device for a 3D network-on-chip according to one aspect of the present invention, wherein weights are respectively set for the plurality of routers, and selection of the adjacent router is assigned to the plurality of routers. It is characterized in that it is executed according to the weight.

A further example of a TSV error tolerant router device for a 3D network-on-chip according to one aspect of the present invention is characterized in that the selected neighboring router has a weight greater than the weight of the router whose cluster is defective. do.

A further example of the TSV error tolerant router device for a 3D network-on-chip according to one aspect of the present invention is characterized in that the weights set in the plurality of routers are such that a router in the center of one layer has the largest weight , the weights decrease and are minimized at the edge of the layer.

1 illustrates a conventional method of using redundancy; FIG. 1 is a diagram illustrating the principle of a fault tolerant router device for 3D network-on-chip according to the present invention; FIG. FIG. 2 is a diagram showing phase-to-phase connections of routers; Fig. 2 shows a simplified layout of a TSV-based 3x3x3 3D-NoC device according to the present invention; Fig. 3 shows a wrapper for the 3D router; FIG. 4 is a diagram showing connections between two layers; FIG. 4 is a flow diagram illustrating a proposed algorithm for the sharing mechanism according to the invention; Fig. 2 is a flow chart diagram of a cluster TSV error resilience technique;

An embodiment of the present invention will be described below with reference to the drawings. The examples are for a better understanding of the invention, and therefore the application of the invention is not limited to the examples. The scope of the invention extends to what is recited in the claims and their equivalents.

FIG. 2 is a diagram illustrating the principle of a fault tolerant router device for 3D network on chip according to the present invention. When all TSVs in group A are defective, the device will process the data by neighboring group B. Smart management allows the device to maintain both connections and avoid adding a large amount of redundant TSVs.

Faults in TSVs are detected by Fault Detection, Diagnosis and Recovery Algorithms and the like.

To deal with the shortcomings of TSV clusters in 3D-NoCs, our solution is to share TSVs among neighboring routers. Therefore, if the TSV cluster fails, the router can borrow a healthy cluster from one of its neighbors to maintain connectivity.

FIG. 3 shows the inter-layer interconnection of routers R1-R3 and routers R2-R4 adjacent to routers R1-R3, illustrating TSV cluster tolerance according to the present invention.

As shown in FIG. 3(b), if one cluster DC in the connection of routers R1-R3 is defective, one healthy cluster HC is borrowed from the connection of routers R2-R4 to maintain the connection. to maintain As shown in Figure 3(c), if there are no healthy clusters to borrow, it can be changed to defective cluster DC serialization mode (for example, if there are only two healthy clusters, the connection of router R1-R3 uses a 1:2 serial).

Now consider a simple layout example of a three-layer 3D-NoC with layers 0, 1 and 2 using TSVs for easier understanding of the invention. FIG. 4 shows a simplified layout of a TSV-based 3×3×3 3D-NoC according to the present invention. In FIG. 4, 40 is a router, 41 is a TSV landing pad, 42 is a TSV cluster, and 43 is a TSV common area. Since the routers 40 are arranged in a 3×3×3 array, they are assigned coordinates of (x, y, z).

For each cascade connection, one router needs a set of TSVs. Instead of grouping all TSVs together, they are divided into four groups. As a result, router 40 has four TSV clusters 41 centered on said router 40, and further TSV clusters belonging to up to four adjacent routers. If one TSV cluster of a router is defective, the router does not need to have redundant TSV clusters and can alternatively select one of the clusters belonging to the four neighboring routers.

To meet the time constraint, the router selects the closest TSV cluster among the neighboring clusters. To avoid the long wires required to establish a connection, it is not conceivable to consider further TSV clusters. By configuring TSVs into four clusters for each router, the size of 3D-NoCs is preserved and long wire delays are avoided.

A router needs a support module to borrow a TSV cluster from a neighboring router. FIG. 5 shows a wrapper for the 3D router. The 3D router of Fig. 5 is located at coordinates (1,1,1) and comprises additional shared circuits (S-UP and S-Down) 50, 51 which are coordinators managing the proposed algorithm according to the invention. .

Two identical shared modules (S-UP and S-Down) 50, 51 are provided for the two vertical and vertical connections, each using a tristate gate. There are two control registers CR52U, 52D (CR53U, 53D) for the input and output ports of the control circuits 54,55. The control registers CR52U, 52D (CR53U, 53D) store CR values that allow router TSV access. The CR value is read by control circuits 54,55.

As previously shown in FIG. 4, router R(1,1,1) has four neighboring routers R(1,1,0), R(1,1,2), R(1,0,1 ) and R(1,2,1) share the TSV cluster.

FIG. 6 is a diagram showing connections between two layers. In FIG. 6, "w" means the flit width. The input of this TSV cluster is shared between R(2,1,0) and R(2,1,1) on layer 2, and the output of the TSV cluster is R(1,1,1) on layer 1. 1) and R(1,1,0).

If this TSV cluster is defective or borrowed, the data is sent using one of the four adjacent clusters 60 . Based on the 6-bit CR value in little-endian format (represented as CR[5:0] for 6 bits) set in the control registers CR52U, 52D (53U, 53D), the input/output port of the control circuit 54 is , (1) the data can be selected from the original TSV cluster (first bit), or (2) from one of four adjacent clusters (second bit), or (3) the data is unconnected ( Alternate clusters are indicated by one of the remaining four bits).

As shown in FIG. 6, the output data from R (2, 1, 1) is in the TSV cluster when the least significant bit (Least Significant Bit) is “1” (CR[0]=1) Sent. Disconnect the original TSV cluster from R(2,1,1) by setting the least significant bit to "0" (CR[0]=0) .

When the second bit is set to "1" (CR[1]=1) , the neighbor router R(2,1,0) accesses the cluster. If the original TSV is defective or replaced, the router replaces one of the neighboring routers and maintains the connection based on the remaining 4 bits of CR (CR[5:1]). do. At the receiving 2 router R(1,1,1), a similar CR is used to maintain the connection. The value of this CR is the same as the CR of the sending router.

Since CR only manages connections, its value must be set carefully to avoid possible conflicts in the use of TSV clusters and to obtain optimal performance.

For this purpose an adaptive sharing algorithm is needed.

Above, we described how routers can use neighboring TSV clusters to maintain connectivity and operations on layers. CR values need to be configured to handle TSV defects. A simple method for this purpose is to go online and fuse the configuration to the TSV group [9].

However, fixed connections have two drawbacks. First, recovery of a newly defective TSV requires stopping the system and re-performing the mapping. Second, each application has a different distribution in the vertical connection and has variations depending on task execution that are not optimized for online mapping.

Therefore, we aim to perform mapping online so that the system can immediately react to novel defective TSV clusters and consider the connectivity of 3D-NoC systems. Thus, an online algorithm for shared TSVs is provided that can be implemented in the system.

FIG. 7 is a flow diagram illustrating a proposed sharing algorithm for the sharing mechanism according to the invention. Each router is assigned a weight for each cascade connection. This weight determines the priority when sharing/borrowing. Weights can be assigned at design time or updated by modules determined.

You can change the router weights to generate different mappings. At the initial stage, all routers in the network exchange their weights and TSV cluster states with neighboring routers (steps S1, Yes to S3). In the next step the algorithm performs the mapping process. If the TSV cluster is defective (step S4, Yes), the corresponding router finds possible candidates from neighboring routers depending on the following conditions (step S5).

• The weight of the candidate must be less than the current router.
• Candidate TSV clusters must be healthy and not borrowed.
• The weight of the final candidate should be the smallest among all possible candidates.

At the end of the algorithm, the router finds possible candidates to borrow (step S6). If no candidate is found, cascade connection of routers is disabled (steps S7 and S8). If there is a candidate, the router sends a request to the borrowing router to use that TSV cluster as the replacement for the defective TSV cluster. A router with a borrowed TSV cluster looks for a replacement of one of its neighbors (step S9). Using a weighting system, disabled TSV clusters focus on routers with lower weights.

Here, one of the most important parameters in the shared algorithm is the router weight value. Weights help the algorithm decide which routers are eligible to borrow. Routers with small weights are disabled after the shared chain is established. Since weights determine the router's priority in shared processes, weights need to be optimized for maximum system performance. To do so, the best solution is to use a statistical solution where cascade priority depends on communication traffic [14, 15].

In other words, higher weights are assigned to cascades with more data communication. Otherwise a small weight is assigned. A simple method is that the router in the middle of the layer has the highest weight as an example of the invention. The weight of the router is reduced and is lowest at the edge of the layer. Equation 1 shows this assignment of weight values used.
Weight _router (x, y) = min(x, cols - x) + min (y, rows - y) + 1

When a TSV is borrowed, it is managed by the borrowing router. However, if the borrowing router is layer disabled, the borrowed cluster must be released and returned to the original router. As a result, when a borrowed TSV cluster forms a borrowing chain, an undoing chain is generated.

A router's connection weight is designed to deprioritize a router's cascade connection over other connections. The following three reasons are presented.

(1) Ease of coordination; there is a need for a coordination mechanism between two routers. Unlike other devices/systems, routers in a network operate in parallel without knowledge of master/slave status. Having a weight helps the router know the capacity it borrows, in order to let the router interpret the accessibility to the TSV cluster.

On the other hand, without weights, routers require handshake functions, requiring complex architectures and costly uptime. Weights allow routers to compare their own weights with other weights and decide whether or not they can be borrowed. These weights ensure absolutely no collisions between routers. Requests to borrow are not checked at the receiving router due to the security of protocol issue requests. As a result, the borrowing state does not change until the weights are updated. And there is no race condition (the forever lend-and-borrow case), which is the most famous bug in multi-threaded and distributed systems/programs.

(2) Shift disabled routers to the same region; using weights, routers with higher weights are more likely to have active clusters. On the other hand, disabled clusters are shifted to lower weight routers. Without the weight system, it is not possible to create a focal point for routers incapable of sharing/borrowing processing.

(3) Accommodates different configurations; depending on the application, the designer can use different configurations. As a result, the same architecture and algorithm can be applied to multiple applications without modification, just by adjusting the weights. Furthermore, the application is open to online application systems. Adjustments in the algorithm are needed to optimize the configuration.

After applying the sharing mechanism, disabled TSV clusters are shifted to regions with lower weight routers. However, an incapable router may borrow from a higher weight neighbor to get full connectivity. Since the borrowing process is inhibited (lower weight routers borrow from higher weight routers), weights are adjusted after a router fails. After a router's weight is reduced, the router with lower weight can borrow a new cluster to get a vertical connection.

By renting a TSV cluster that operates around a defective cluster, without adding redundancy, some routers will be reduced to four accessible clusters. As a result, these routers are disabled for communication. To address this problem, the original solution uses a fault-tolerant routing algorithm to divert packets to neighboring routers. As mentioned above, this solution leads to non-shortest routes and conflicts in the network. Therefore, we propose a virtual TSV, which is a TSV that these routers temporarily borrow to maintain connectivity without using fault-tolerant routing algorithms. If the virtual TSV is infeasible, a serialization technique is implemented to establish only one or two TSV clusters by cascading.

When a router is not allowed access to four TSV clusters, it becomes disabled. However, if the number of nearby TSVs, which is sufficient to maintain longitudinal communication, is greater than or equal to 4, it will be used to establish the connection. A possible connection that requires four TSV clusters requires clusters belonging to neighboring routers. If these routers do not use the cluster, incapable routers will borrow them for a short time to establish communication.

A high priority router may occupy the TSV for a long communication time, so that a low weight router cannot access the TSV to establish a connection. Moreover, with a large failure rate, routers with small weights fail to find good candidates for virtual TSVs. To solve these problems, the inventors adapt the continuous technique [6] to maintain the connection.

Virtual TSV can help disabled routers maintain cascading connections, but there are two situations in which virtual TSV is not implemented:
(a) There are four healthy TSV clusters. (b) The router with the higher priority TSV of the candidate is regularly occupied.

To solve these cases, a serial technique [16] is used to keep the connection alive. For continuous technology, a router needs at least one TSV cluster to maintain connectivity. With one possible cluster, 1:4 continuity is used, and with two possible clusters, 1:2 continuity is established. The up and down cluster outputs are stored in registers and the successive modules send flits on the remaining clusters.

FIG. 8 is a diagram showing a flow chart of TSV error resilience from TSV error detection to substitution. The system starts with the initialization of weights and TSV states. When a new defect is detected (step S10), the system runs the shared algorithm described in FIG. 7 to determine router priority by weight (step S11). Once the network is established, there is no sharing/borrowing (step S12 ) and the system optimizes the TSV sharing algorithm (step S13). The system then exits and stores its configuration in the CR register (step S14). Based on the value of the CR register (see Figure 5), routers can implement specific sharing, virtualization, or serialization methods.

Claims (3)

In a TSV fault tolerant router device for a 3D network-on-chip having a plurality of routers arranged in each of a plurality of layers, the routers between layers being connected by through silicon vias,
each of the through silicon vias belongs to one corresponding router and has a plurality of clusters around the corresponding router;
when one of a plurality of clusters belonging to the corresponding router is defective, a router adjacent to the corresponding router is selected, and one cluster of the selected router is replaced in place of the defective cluster; maintain connectivity between layers, and
A weight is set for each of the plurality of routers, and selection of the neighboring router is performed according to the weight assigned to the plurality of routers.
A TSV fault tolerant router device for a 3D network on chip, characterized by: In claim 1 ,
the selected neighboring router has a weight less than the weight of the router whose cluster is defective;
A TSV fault tolerant router device for a 3D network on chip, characterized by: In claim 1 ,
The weights set for the plurality of routers are such that a router at the center of one layer has the highest weight, the weight decreases, and the weight is lowest at the edge of the layer.
A TSV fault tolerant router device for a 3D network on chip, characterized by:

Images