Thermo-mechanical reliability prediction for copper pillar 3D IC devices

3DIC technology has enabled scaling beyond the Moore's Law to achieve higher transistor count, increased functionality and superior performance. Additionally, this technology allows integrating heterogeneous components such as Processor, FPGA, GPU, Memory, Serdes, etc. on the same interposer die enabling faster computing through reduced latency. The yields on the 3DIC technology have matured and are equivalent monolithic flip chip products. All of the above has made 3DIC technology, the key driver for some of the high end computing applications such as Data centers. Any technology is viable only if the end product is reliable and manufacturable with high yields. Understanding the reliability margin of 3D IC package is essential for making them commercially successful. This paper has focused on predicting the thermo-mechanical reliability of 3D IC devices. The device used for this study is a 20nm 3D IC device with 2 FPGA slices stacked on a passive interposer. Both ubump and C4 bumps use tin capped copper pillar bumps. The FPGA and interposer stack are mounted on a 45mm substrate with lead free BGA. Experimental data is collected using standard component level thermal cycling and board level reliability to determine the cycles for first failure. Additionally a board level power cycling test vehicle has been used to mimic real end use condition. The power cycling test vehicle uses self heating elements using the logic elements in the FPGA such as LUTS and Flip Flops. Toggling the logic circuits at a higher rate increases the dynamic power consumption of the FPGA thereby creating more heat. Finite element analysis has been used to predict the thermo-mechanical fatigue life of C4 and BGA solder joints. The accumulated strain energy density was used as damage indicator. A quarter symmetry model is used to increase the computation efficiency. The predicted stain energy density has been correlated with experimental data to derive the prediction constants. The derived prediction constants are then used to predict the life of the C4 bump in field use condition.