ICCAD 2015 Contest
3D-ICON: 3D Interlayer Cooling Optimized Network

Mohamed M. Sabry1, Arvind Sridhar2, and David Atienza3
1Stanford University, 2IBM Zurich, 3ESL-EPFL

I. INTRODUCTION

Interlayer single-phase liquid cooling has been proposed as an effective cooling mechanism for the high-heat dissipated in high-performance 3D-stacked processing architectures. Single-phase fluid, such as water, is injected into micro-scale channels (also called micro-channels) that are etched between two consecutive vertical tiers to carry the heat out from different layers in the 3D stack.  This cooling mechanism has already been proven to be much more effective than conventional air-cooling in order to remove very high on-chip temperature and cool the computing system much faster (several orders of magnitude of improvement).

A key component in inter-layer cooling that needs careful consideration is the efficiency of the cooling network. Indeed a one-design-fits-all approach of using uniform micro-channels can easily lead to overcooling, which is not an energy-efficient operation, or non-uniform cooling that increases on-chip thermal gradients giving rise to reliability issues and shorter lifetimes.  Thus, it is fundamental to optimize several parameters, such as the pressure drop across the channels, the number of fluid inlet and outlet ports, and the fluid network, to achieve optimized cooling in terms of energy efficiency, thermal gradient and peak temperature.

 

II. CONTEST OBJECTIVES

The goal of this ICCAD 2015 problem is to evaluate the impact of different cooling networks on different computing architectures floorplans, when subjected to specific thermal performance requirements. With the CMOS process that is used for etching micro-channels on the back-side of a chip, the possibilities of creating microfluidic networks for customized cooling is nearly endless (subject to system requirements such as allowing TSVs placement and pitch constraints for inter-layer communication). However, this aspect of microfluidic cooling has not been fully explored and exploited in the current research on the field of liquid cooling and thermal packaging. We hope solving this problem would bridge this gap, motivate further research insights in liquid-based cooling, and potentially attract even more industrial and academic interest in this field.

 

 

III. FORMAL PROBLEM DEFINITION

The main objective of this problem is to find an optimized cooling network that minimizes a specific cost function while subjected to design and physical constraints. Before unfolding the target problem, we need to define several terms to avoid ambiguity. These terms are the following:

Term

Definition

fi

Floorplan of layer i in the target 3D stacked architecture

mji

Architectural module j in the ith layer

cni

Cooling network layer i in the target 3D stack

T(mji)

The peak temperature observed in module mji

DP

The pressure drop observed in the entire cooling network

Qin

Inlet flow rate

 

To reflect various operating conditions we have created a set of problems, which we elaborate in the following sections. As general definition, given a certain N-layer 3D-stacked computing architecture defined by floorplan F={f1, f2, …, fN}, where each layer i Î [1, N]  consists of a set of j modules Mi={m1i, m2i, …, mji}, provide the set of N-1 cooling network layers CN={CN1, CN2, …, CNN-1} (cooling network layer CNi is placed between stacked layers i and i+1) that is a solution of:

a)    MinT: minimizing the peak temperature (argmaxi Î [1, N] T(mij) ) and thermal gradient (ÑT), subject to pressure drop constrain (DP £ Pmax). 

b)    MinCE: minimizing the cooling energy (CE º DP. Qin), subject to peak temperature (argmaxi Î [1, N] T(mij) £ Tmax) and thermal gradient constraints (ÑT £ Tgradmax).

Based on the problem definition, we will provide different case studies, where each case study will have the input defined as follows:

·        Number of layers in the targeted 3D stack (i.e., the number N)

·        The floorplan of each layer in the targeted 3D stack (F)

·        The power traces of each component in the 3D stack (M={M1, M2, …, MN})

·        The list of interlayer floorplan (FIL={fil1, fil2, …, filN-1}), where each element of this list shows the locations where a microchannel placement is infeasible

In addition to the mentioned inputs, we fix the channel width to the following value for all test cases:

·        Channel width= 100um

Various teams will use the mentioned inputs to provide the following list of outputs, which are used for evaluation:

·        The applied pressure drop to inject the fluid to the cooling network (DP)

·        The layout of the different cooling networks in the 3D stack (CN)

·        The location of inlet and outlet ports

Case Study: 2-tier stack

Macintosh HD:Users:mohamedmostafasabryaly:Documents:2_tier.pngAs an illustrative example, we provide the inputs of a 2-tier stack, which is depicted in Fig. 1.

The inputs for this example is as follows:

N=2

F={f1,f2}

f1={m11: [(0,0), (L,W)]}

f2={m12: [(0,0), (L/2,W)],  m22:[(L/2,0), (L/W/2)], m32:[(L/2,W/2), [L,W)]}

Figure 1. Schematic 3D diagram of the example 2-tier stackP(M1)={0,0,P,P,P/2}

P(M2)={(0;0;0),(P1:P2;P3),(P1;P2/2;P3),(P1/2;0;P3/2),(0;0;0)}

FIL={fil1}

Fil1={[L1,W1],[L2,W2],…, [Lk,Wk]}

IV. EVALUATION METHODOLOGY

The proposed solutions by the different teams working in the competition, which is defined by the output list in the previous section, will be assessed based on:

·        Finding a feasible solution

·        Meeting the problem constraints

·        Minimizing the relevant metric to each problem

Detailed rules regarding the format of the problem statements provided, the mode of solving and the solutions obtained are as follows:

Figure 2: Template of a 11X11 Fluid-Network: 0 = silicon, 1= Fluid, -1=TSV (fixed), 2=inlet, 3=outlet.

Rules

Rule #1:

The description of the problem 3D stack, the floorplan and the heat inputs will be given in the same format as the 3D-ICE input files (“.stk” and “.flp”). Please refer to the 3D-ICE User Guide available following this link here. Divide your cavity layer into square cells (from top-view) of 100um. For a chip size of 10.1mmX10.1mm, this would give a grid of 101X101 cells. Note that some of the channel description part of the .stk file will be skipped in the input definition

as it is part of the solution that you must compute.

Rule #2:

Specify your solution fluid-network map solution in an excel file using an array of 0’s and 1’s. 0 implies there is silicon in the cell, and 1 implies channel/liquid. For the purpose of illustration, a 11X11 grid example template for a fluid-network representing a cavity layer is shown below in Fig. 2.

Rule #3:

A 2D array of alternating cells cannot have fluid and have a -1 by default- to account for TSV fabrication. This is specified in the template using bold -1’s in cells filled with brown background (Fig. 1).

Rule #4:

The table in the excel file will be visualized in the same orientation as the top-view of the physical device. That is, the bottom-left (south-west) corner of the table will correspond to origin. The origin point and the cardinal directions considered are shown in Fig. 2.  This is the same convention that is used for the input heat-flux map provided in the problem, and in the 3D-ICE model.

Rule #5:

Your solution must have at least one inlet and one outlet. The inlet and outlets can only be at the edges of the microchannel layer. The edge fluid cells corresponding to inlet and outlet are specified using 2’s and 3’s respectively (see Examples 1-4, Fig. 3-6).

Rule #6:

There can be multiple inlets and outlets along all four edges of your solution. However, to reduce the complexity of the packaging, for each side, there can be at most one “continuous” inlet and outlet, as illustrated in Example 3. There should be at least one traceable fluidic path from the inlet to the outlet of the system. If there are multiple channel layers in the problem, the inlet and outlet patterns across the multiple layers must be identical. The individual fluid networks within each channel layer can however be different.

Rule #7:

Provide your final solution excel file for each problem (separately), after you name the file in the following format:

<<TeamID>>_ProblemA_<<testcase#>>.xls

where

<<TeamID>> refers to your Team Identification as you have been assigned by the ICCAD 2015 registration process

<<testcase#>> refers to the serial number of the problem test case.

If there are multiple channel layers in a given test case, provide them as separate .xls files in the following format (channel layer counted from bottom to top):

<<TeamID>>_ProblemA_<<testcase#>>_<<channellayer#>>.xls

Rule #8:

Along with your solution excel file, you must also provide ONE value of pressure drop between the inlet and outlet of your system (even if you have multiple inlets and outlets in the cavity, the pressure drop between each of them must be identical).

Important: Solutions that do not adhere to the above rules will be disqualified.

Rule #9: To all teams: please provide all your excel sheet submissions (named according to the rules) for all test cases in a SINGLE folder and all the pressure drops corresponding to each test case in a SINGLE text file inside the same folder. 

 

 

Hint:

1.         Use the solution fluid network template file Fluid_Network_Solution_Template_101X101.xls that has been provided to construct your final solution files.

2.         First, You must solve the flow rates into and out of each fluid cell from/to neighboring fluid cells based on the discretization above, and using the formula below (for simplicity of modeling, we assume Darcy-Weisbach friction factor for developed laminar flow):
     (1)

Here, subscript i,i+1 refer to the pressure drop between two neighboring cells with indices i and i+1, and the flow rate from cell i to the neighboring cell i+1 (please note that the indices are only illustrative, and do not refer to which “direction” we are moving from the current cell to the neighboring cell- could be north, south, east or west). The value of the k or the method to compute is provided for each test case separately. Please see the test case description pdfs.

Essentially, you must construct and solve a simple “pressure-flow network” problem which looks like a resistive network where Pressures are voltage and flow rates are currents, according to the above formula and get all local flow rate for each fluid-fluid interface of each fluid cell in your network (see Examples 1-3). Describing an arbitrary network of microchannels here can be easily accomplished using Graph Theory, by building an "Directed Incidence matrix" Z where the rows indicate all the liquid cells/nodes and the columns containing +1/-1/0 indicate the presence/absence of a connection between 2 nodes (channel) and the assumed direction of flow (https://reference.wolfram.com/language/ref/IncidenceMatrix.html). This matrix Z would essentially give the relationship between the pressures in each cell/node (with respect to the reference pressure at the inlet) and the pressure drops across each edges/channel segments as follows:

   (2)

By combining this matrix Z with the known linear relationship between pressure drop across each "edge" and the flow rate between them (Equation 1 above),  you can easily solve the flow rates in the entire network using sparse matrix inversion.

3.         Once the local fluid flow rates and directions are computed, you could tweak the 3D-ICE (4RM) model based on the same discretization of 100um X 100um for all layers. This will require some minor coding of the 3D-ICE program (if you plan to use it). For each fluid-solid interface in your problem, use the following formula to calculate the local convective conductance term in the equivalent thermal RC grid construction (see [1] and http://esl.epfl.ch/3D-ICE  for more details):

    (3)

Ainterface  refers to the area of the wall interface of the fluid cell towards a particular neighboring solid cell. The value of/the method to compute the heat transfer coefficient hconv will be provided for each test case separately. For simplicity of modeling, we assume Shah-London heat transfer correlation for fully developed laminar flow in all our problems.

4.         Once you build the channel network flow rate and the heat transfer models as described above, your system model is ready. Now you use this model to search the design space of channel networks (by respecting the various constraints) using any method you like (Gradient descent methods, Monte-Carlo, Simulated Annealing, Genetic algorithms etc.) for an optimal design.

5.         As a first attempt, if building the channel network model is too complex,  you could simplify in various ways. For example, in the new heat transfer model, you could only consider the vertical thermal resistances and neglect the North-South-East-West resistances in the first pass, since bulk of the heat enters the channels from the Top and the Bottom. The optimal channel solution you obtain would be pretty close to the one you would find if you included these lateral resistances in a network. Once you find this approximate solution, you can refine your search to find the exact one.

6.         There have been works done in the past on finding optimal microfluidic networks to minimize various cost functions, similar to this problem. You can find them here:

[i] Van Oevelen, Tijs, and Martine Baelmans. "Numerical topology optimization of heat sinks." Proceedings of the 15th International Heat Transfer Conference. 2014.

[ii] Zhang, Yongcun, and Shutian Liu. "Design of conducting paths based on topology optimization." Heat and Mass Transfer 44.10 (2008): 1217-1227.

[iii] Xu, Peng, and Boming Yu. "The scaling laws of transport properties for fractal-like tree networks." Journal of applied physics 100.10 (2006): 104906.

You could adapt the principles described in these papers to find the optimal design for this problem.

Figure 3: Fluid-Network Example 1 and the resulting Pressure-Flow Diagram

Figure 4: Fluid-Network Example 2 and the resulting Pressure-Flow Diagram

Figure 5: Fluid-Network Example 3 and the resulting Pressure-Flow Diagram

Figure 6: Fluid-Network Example 4 and the resulting Pressure-Flow Diagram

 

References

[1] Sridhar, A. Vincenzi, D. Atienza Alonso and T. Brunschwiler. 3D-ICE: a Compact Thermal Model for Early-Stage Design of Liquid-Cooled ICs, in IEEE Transactions on Computers, vol. 63, num. 10, p. 2576-2589, 2014.

[2] M. M. Sabry, A. Sridhar, J. Meng, A. K. Coskun and D. Atienza. GreenCool: An Energy-Efficient Liquid Cooling Design Technique for 3-D MPSoCs Via Channel Width Modulation, in Ieee Transactions On Computer-Aided Design Of Integrated Circuits And Systems, vol. 32, num. 4, p. 524-537, 2013.

 

Test Cases

l  Testcase01  Testcase01(update)

l  Testcase02  Testcase02(update)

l  Testcase03  Testcase03(update)

l  Testcase04  Testcase04(update)

l  Testcase05  Testcase05(update)

l  ReadMe.txt ReadMe.txt(update)

 

 

General feedback

1.         How do we rate the submissions: since these problems are extremely challenging without any clear analytical solutions, we will consider this to be a "race" between the teams- which ever team minimizes the cost function the most (while respecting the constraints) gets the best score. We will give appropriate weights to the final cost function and the observance/violations of the constraints where applicable. The priority of the constraints, as described in the problem statements, is in the order in which they are listed in each case.

2.         Clarification regarding the "temperature gradient" constraint in the various test cases. It is defined as follows:
cid:01919F9901AE4E2B8FF8EC2183D7F471@shhpc
That is, first for each "junction" layer or source layer in a given test case, find the Tmax-Tmin. This is the per-die temperature gradient. Then find the maximum of these temperature gradients across all dies in the 3D stack in the test case. This is the "temperature gradient" of the 3D stack, which must be maintained below the specified value in each test case problem. This has been updated in the attached Alpha Test cases. (please download them from the contest website).

3.         As mentioned in the rules of the problem statement in the website, the channel layers are numbered from bottom to top- that is the same as what you use in naming your excel file.
For eg. <<TeamName>>_ProblemA_testcase04_channel01.xls refers to the bottom channel layer and <<TeamName>>_ProblemA_testcase04_channel02.xls refers to the top channel layer in Test Case 04.

4.         As already described in the description of Test Case 04, both the channel layers must share the SAME inlet and outlet "lines" or ports (even if their individual channel networks are different), which means that the pressure drops across all inlets and all outlets in all layers must be the SAME. Hence, you must only give ONE pressure drop value for both the channel layers.

5.         There are some are some typos in some of the .stk files in the original Alpha test cases. They have been corrected in the newly updated Alpha test cases. Please download them from the contest website

 

6.         The teams are free to continue to revise their submissions for the existing test cases for Problem A until the final submission deadline

 

7.         The teams must submit a brief abstract describing their methodology in solving the Problem A. The abstract (in .pdf or .doc format) should be written in a single A4/US letter page layout, and should not be longer than 1200 characters including spaces, containing a maximum of two figures. Reports not adhering to these rules will be rejected.

 

8.         The teams must follow the instructions on how the excel files and the text document for pressure drop must be organized before the final submission. Remember to put all excel files in the same folder and the pressure drops in in a single text file in the same folder - just submit ONE folder without any sub folders. The submissions of teams not adhering to these rules will be disqualified.

 

9.         Only the latest solutions submitted on/before the final submission deadline will be evaluated. All previous versions/submissions will be discarded.

 



FAQ

1.         For micro-channels with bended structure, should we modify the codes in 3D-ICE by using the ways discussed in the paperhttp://infoscience.epfl.ch/record/186785 if we need to use the state-of-art thermal simulator in the optimizing flows for the microchannel network design? 

ANS: Yes, you are free to 3D-ICE model based on the paper.

2.         If question 1 is acceptable, could we know that how the evaluation of the results is going on for the bended structure? That is, should we need to upload the codes about the modified codes of 3D-ICEs or do other things (such as follows the format of the given design in the future on website)? 

ANS: All the above points about constraints and evaluation procedure will be given later, during the final problem announcement. 

3.         For the given inputs, it seems no constraint about the structure (only maximum temperature ant thermal gradient) during design iterations. Could we know that microchannel widgets/heights can be changed (along with the microchannel length, like the proposed method in GreenCoolhttp://infoscience.epfl.ch/record/186491/files/TCAD2013-04-06480864.pdf?version=1 ); pitch located between microchannels be constrained or not; and how thermal gradient is defined?

ANS: All the above points about constraints and evaluation procedure will be given later, during the final problem announcement. 

4.         I am very new to the research area, so my apologies if the question is trivial. The problem asks us to provide a layout of the cooling networks given a set of constraints. The tool to be used to evaluate the solution is 3D-ICE. I have been unable to locate where are the cavity geometries specified in 3D-ICE. As far as I have been able to tell, it assumes uniform cavities. Please suggest.

ANS: “D-ICE” is not mandatory to be used, it is highly recommended though. The 3D-ICE framework can be modified similar to the approach shown in this link http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=6532287. Thus some coding effort is required. I hope this clarifies the confusion and addresses the question below.

5.         How is the definition of thermal gradient?

ANS: Thermal gradient is defined as Tmax-Tmin in a given active layer (or across different layers)

6.         In the microchannel design, is the flow rate of each microchannel grid all the same?

(It seems if the microchannel grid is with three input and one output, the flow rate is not the same? So it could be needed to solve mass balance equation?)

ANS: You must solve the flow rates into and out of each fluid cell from/to neighboring fluid cells based on the discretization above, and using the formula below (for simplicity of modeling, we assume Darcy-Weisbach friction factor for developed laminar flow):
  
     Here, subscript i,i+1 refer to the pressure drop between two neighboring cells with indices i and i+1, and the flow rate from cell i to the neighboring cell i+1 (please note that the indices are only illustrative, and do not refer to which “direction” we are moving from the current cell to the neighboring cell- could be north, south, east or west). The value of the k or the method to compute is provided for each test case separately.
     You must construct and solve a “pressure-flow network” problem based on the above formula and get all local flow rate for each fluid-fluid interface of each fluid cell in your network (see Examples 1-3).

7.         For the question 1. “How is the definition of thermal gradient?” I’m still confused. The example seems to be designed for transient state power waveform? If yes, how is the definitions of thermal gradient in transient state?

ANS: The definition of the thermal gradient is as I have stated before. If there is a case where transient simulations are needed (highly unlikely), the specific definition will be given on a case- by-case basis.

 

8.         I have some question about the case study: 2-tier stack on the website. It is shown as follows:

Case Study: 2-tier stack

As an illustrative example, we provide the inputs of a 2-tier stack, which is depicted in Fig. 1.

The inputs for this example is as follows:

N=2

F={f1,f2}

f1={m11: [(0,0), (L,W)]}

f2={m12: [(0,0), (L/2,W)],  m22:[(L/2,0), (L/W/2)], m32:[(L/2,W/2), [L,W)]}

P(M1)={0,0,P,P,P/2}

P(M2)={(0;0;0),(P1:P2;P3),(P1;P2/2;P3),(P1/2;0;P3/2),(0;0;0)}

FIL={fil1}

Fil1={[L1,W1],[L2,W2],…, [Lk,Wk]}

Are P(M1) and P(M2) represented transient power waveform (with five time steps on each macro) or other explanations on those?

ANS: The power values will be given as steady-state values but at different running scenarios. If for any reason there is a transient component to the power problem, we will explicitly specify it.

 

9.         The 3D-ice can only deal with one type of cooling network(parallel form). So if we want to check the perform of our designed cooling network, we have to modify the 3D-ICE tool. But how will the result be graded? Is this possible that you already have a tool to compute the efficiency of network? (But can't offer the tool to participants) Or maybe I have misunderstand some part of the problem?

ANS: 3D-ICE can deal with parallel and bent channels (check the TCAD paper: http://infoscience.epfl.ch/record/186785/files/TCOM2014-3D-ICE.pdf?version=1)
              So the participants have to slightly modify 3D-ICE to accept different cooling network topologies, but the tool kernel can handle it.

10.     How important of the MinT and MinCE in the formal problem definition? Is there any percentage between them? (e.g, MinT occupied 35% of total points and MinCE occupied the left 65%)

ANS: Both problems are equally important, which can be seen as 50% each

11.     Can we know how 3D-ICE tool work in details in the coming announcement? Will the official release more explicit problem information? (After surveying the reference papers and downloading the 3D-ICE tool then tracing 3D-ICE's source code, we found out that it seems that we now still don't know how 3D-ICE work in reality since we didn't see any equation or assumption in the codes. So we want to inquire about the assumptions or the equations used in 3D-ICE since the evaluation tool is also the 3D-ICE.)

ANS: 3D-ICE has a special documentation on how it works, and the two publications (ICCAD’10 and TCAD’14) explain the theoretical foundations well.
               Please note that 3D-ICE has been used numerously in other research efforts. Thus the tool can be seen as self-contained.

12.     After tracing 3D-ICE's codes, we found out that 3D-ICE seems to package some information and then send it to server side to be computed. Then server side send back the computed results to client side(user, we). So can we modify the real computation codes in 3D-ICE?

ANS: 3D-ICE includes both server-client and standalone modes. So you can only use the standalone model.

 

13.     When can I get the problem stack .stk file?

ANS: we will post the test case project files soon . Please keep checking the problem website

Do I have to assume that, the outside environment is ambient or should I assume that it is insulated?

ANS: It will be specified for each test case separately.

And regarding the problem A, do I have to submit the code or only the generated .xls files?

ANS: No, only the .xls file is necessary.

 

14.     The content of problem A mention that we need to refer to the user guide and we should compile the SuperLU and 3D-ICE before we begin, but When we enter the command line $make superlulib, it present ar: /edahome/cad2015/Codes/SuperLU_4.3/lib/libsuperlu_4.3.a: No such file or directory and we check the directory and then we found that we don't have the file libsuperLU_4.3.a in our directory. We check the compressed file superlu_4.3.tar.gz, find there are no library in this file, the "lib" folder is empty. So we cannot successfully construct the experimental environment.

ANS: Did you follow the process of compiling superlu correctly? See the 3D-ICE user guide page 6 for the pre-requisites... 

Before compiling 3D-ICE, you must compile the SuperLU library by executing the following commands: 

$ wget http://crd.lbl.gov/~xiaoye/SuperLU/superlu_4.3.tar.gz 

$ tar xvfz superlu_4.3.tar.gz 

$ cd SuperLU_4.3/ 

$ cp MAKE_INC/make.linux make.inc 

Next, check and edit the SuperLUroot variable in ./make.inc and select the blas library before compiling.

You can either use a blaslibrary installed on you system or theblaslibrary supplied by the authors of SuperLU (see the README file).

If you decide to use the former then thevariable BLASDEFmust be set to -DUSE_VENDOR_BLASand BLASLIB must point to your blas library. Then compile SuperLU with

 $ makesuperlulib 

For the latter, BLASDEF must be unset and BLASLIB must point to the libblas.a archive.

To compile the blas library included in the sources of SuperLU the Fortran compiler g77must be available on the system. Then compile SuperLU with 

$ make blaslib 

$ makesuprlulib 

These are the operations that can be done when compiling SuperLU on a generic Linux platform. In case of a different architecture, please reference to the README file. 

15.     We need Test Cases:

(1) We need Test Cases as inputs for our alpha test, which we could not find in CAD Problem A website links. Please tell us how to get the inputs and input formats.
      (2) Whether "stk" and "flp" files are part of the inputs? Whether exist Pmax and

Tmax, the feasible constraints, if so, which part of inputs

for us to allocate those constraints?

ANS:The test cases will be posted soon. Please keep checking the website.

 

16.     What are the relationships between Problem A and 3D-ICE tool?

(1) Whether Problem A gets the same inputs of stk and flp, so that Problem A could be compared with 3D-ICE tool? That is, Problem A is the similar tool with 3D-ICE?
     (2) Whether the inputs of stk & flp in 3D-ICE are part of the inputs of Problem

A? That is, we need the complete description of input file.

(3) Whether 3D-ICE is an evaluation platform for Problem A's results to execute to get the results? 

ANS: stk and flp files will be provided for Problem A. Please see 3D-ICE documentation for their definitions. You don't have to use 3D-ICE for your problem solving.
                But you can if you want (you will have to modify the code to your needs).

 

17.     What are the measurement metrics and their cooresponding percentage in the final score?

(1) Fluid Network: What are the exact measurement metrics/criteria with their coorespnding percentage: CE (Cooling Engergy), Peak Temperature, Gradient of Temperature,number of inputs, number of outputs, the total router length of cooler

paths, the maximum router length of cooler paths?

(2)Pressure Flow Diagram: What are the metrics for Pressure Flow Diagram and their percentage? (3) CPU Time? Memory Usage? And other metrics?

ANS: Each test case will have its own cost function to minimize and that will be the metric used to compare different teams.

 

18.    MinT: minimizing the peak temperature (argmaxi Î [1, N] T(mij) ) and thermal gradient (ÑT), subject to pressure drop constrain (DP £ Pmax). 

b) MinCE: minimizing the cooling energy (CE º DP. Qin), subject to peak temperature (argmaxi Î [1, N] T(mij) £ Tmax) and thermal gradient constraints (ÑT £Tgradmax).

 Is it two problems that need to provide two solutions or one problem that the solution satisfy the two requests at the same time?

ANS: Each test case will have its own cost function to minimize. Please wait for the details.

19.     If we only need to submit the .xsl file, how can you know the 'stk' file which we implemented? And we read the user guide book about 'stk' description, there is only the model of straight line channel, such as page 18

But in fact, we can also implement the model of L-shape channel (turn-left or turn-right)

We don't know how to describe this kind of model in 'stk' file, can you give us some examples?

ANS: The purpose of the contest is to design new channel networks. It is up to the participants to model these different structures on the top of existing 3D-ICE. You must provide your design in the .xls file.

 

20.     And in 3D-ICE emulator, it only need to feed the 'stk' and 'flr' file to check our performance, is there any way to translate the result into Excel '.xls' file ? or we need to type the '.xls' manually ?

ANS:Yes, you must generate the .xls file yourself and program these aspects.

21.     In test case 1, we have some constraint about temperature,

How can we know the temperature is not out of the constraint when we are designing the channel in our program ? Or the only way to know the temperature drop is to run the emulator to get the results?

ANS: Yes you must run the simulator.

22.     What's the unit of '.xls' file and the dimensions in 'stk' file ?

The size of '.xls' file is a 101*101

but the size of 'stk' file is 10100*10100

ANS: One is a number while the other is dimension in um. As mentioned in the test cases, the dimension is 100um X 100um

23.     For test cases with #Dies > 2, should the pressure drop between the inlet and outlet of the system be found separately? (i.e. P_channel_01_layer=300kpa and P_hannel02_layer=400kpa)

ANS: The two layers of channel share the same inlet and outlet port lines. So they must have the same pressure drops across them. Sorry, if this wasn't clear in the problem definition.

 

24.     The problem asks for a cooling network and the pressure drop between the inlet and outlet of the system​. Is the pressure drop a single value or a vector (e.g. a value for each inlet or outlet) ?

ANS:The pressure drop is a single value from the inlet(s) to outlet(s). Please note that despite there can be several inlet ports, they are all connected to the same source (same case in outlet). Thus all inlets will have the fluids at the same input pressure, and at the outlet(s) will have the same pressure (to avoid any back flow of fluids).

 

25.     Would you explain more about the evaluation procedure?

ANS:Each time will submit their cooling network(s), based on the optimization problems and test cases. We will then use these inputs to see if indeed the objective functions have been met or not.

 

26.     If we do not need to submit any code or executable, does that mean there will be no hidden benchmark? Will it be possible for the participants to exhaust all reasonable networks and submit the best one as a single excel file?

ANS:For the time being there are no hidden benchmarks. The whole purpose of the problem is two provide the optimal network based on the problem definitions. you are free in your choice to do an exhaustive search algorithm.

 

27.     According to the announcement, we can use CIC's machine to build our source code. So we can co-build our projects also on CIC's machine? (I mean that whole team use that account to code) And does it equals to that we can also build up "3D-ICE" on CIC's machine? If it is ok, but I found out that "3D-ICE" need to be compiled under C11 standard, but CIC's machine's gcc version seems to be not enough?

ANS:In the problem A, you only need to submit excel files. Therefore, you can install and run 3D-ICE on your machines. So far, we do not install C11 standard in CIC machines. We need to discuss with CIC if it is possible to install C11 standard in CIC machines.

 

28.     According to the pictures attached to this email, what should we submit to alpha version test, beta version test and final submission? Due to the pictures, it seems that we need to submit our program's execution file and README.txt, but the official website(FAQ part, question 13) said that "And regarding the problem A, do I have to submit the code or only the generated .xls files? ANS: No, only the .xls file is necessary.". So what should we really submit?  Only ".xls" file? Or also need ".exe" file? Or something else?

ANS: In the problem A, you only need to submit excel files.

 

29.     If I did not submit files in alpha test, may I still submit files for beta test?

ANS: Yes. You are encouraged to submit files for beta test and final test.

 

30.     In the hint 2 of question A, it mentions that: "You must solve the flow rates into and out of each fluid cell from/to neighboring fluid cells based on the discretization above." But I don't know how to calculate flow rate of each cells by using input informations, such as number of layers, floorplan and power trace.

ANS: The method to solve is explained in the problem description file for each test case (and also in the website)

 

31.     The solution .xls file, is the color needed? Would we just submit the .xls file with numbers and no colors

ANS: Numbers are just fine.

 

32.     Would we have to use our program to get .xls file, or could we get a solution but .txt file by our program and get .xls file by hand?

ANS: It is up to you how you generate the excel file.

 

33.     What is the unit of the value of pressure drop? According to the formula, we transform all the parameters to SI unit and get the unit s^-1. We don't know the correct unit of pressure drop.

ANS: The unit of the pressure drop can be either in Pascals or Bar (1 Bar=100 Kilo Pascals)

 

34.     About Thermal gradient, In the website, it's said That is, first for each "junction" layer or source layer in a given test case, find the Tmax-Tmin. This is the per-die temperature gradient. Is it needed to account for the Tmax-Tmin in Channel layer?

ANS: The optimization problem accounts for the change of the die temperature gradient. If for any reason they need to account for the channel thermal gradient then they should define the thermal gradient of the channel layer as Tmax-Tmin of the fluid for each channel layer.

 

35.     I want to confirm that is the junction layer refers to the layer that the channel exists?

ANS: The junction layer is the layer where the active components reside (i.e., the layers with the digital components generating heat).

 

36.     Basically, we have implemented our own evaluation platform, but are not sure about the consistency between ours and yours. If you can kindly give us the result values of our beta submission, we can use it to check our evaluation platform and need not worry about doing a misguided design.

ANS: There are easy ways to test the accuracy of the models you build: simulate an existing conventional straight channel example using the default 3D-ICE code and then again using your new modified model. The comparison of the results must give you an indication whether your implementation is correct or not.