The paper shows IBM’s road map extending from embedded DRAM to various 3-D stacks with and without interposers using face-to-face (see diagram below) and back-to-back stacks. In a separate paper, IBM disclosed the top two layers of metal in its new 22-nm process are optimized for use with through silicon vias needed for 3-D stacks.
“Embedded DRAM, 3-D stacking, interposers and wafer-to-wafer integration are intrinsic to this [IBM] roadmap,” Iyer said..
In one example, a 45-nm SOI-based processor with eDRAM is stacked with two SiGe BiCMOS transceiver chips. “The bandwidth between chips exceeds 2 Tbits/s, far greater than achievable by either ceramic or organic interposers,” Iyer wrote.
“This interposer technology which can also include decoupling capacitors for mid frequency power supply noise reduction is a key feature of the next step of the evolution of the embedded memory roadmap,” he added.
IBM was quick to caution that “die-to-die 3-D integration is a tedious process and is limited in the ability to achieve dense inter-die connections…but multi GByte [memory] stacks can be integrated with the processor cache stack on a silicon interposer.”
Iyer said IBM is using so-called fat copper wires for TSVs. Although nthey have a four-fold difference with silicon in their thermal coefficient of expansion “control of Cu microstructure can minimize the deleterious effects of this large mismatch,” Iyer wrote.
“The stacking process is very sensitive to die warpage and the handling of thin die and controlling their warpage are key challenges,” he added.
For some time IBM has not invested in leading edge Fabs . So they have a reason to pursue TSV based 3D stacking, partitioning of functions ( SRAM, I/O etc. ). The thermal and stress issues of 3D stacking w/ TSVs are only beginning to be studied and both routing and performance would be affected by these factors.
Integration at the Package level w/o compromising performance too much requires fine - pitch thin film interconnects and soon drilling holes ( TSVs ) in live Si. This is by no means cheap. By not building Fabs of the latest node IBM saves on Capital but their unit cost goes up because of expensive packaging. But they make large and expensive systems so additional part costs get buried.
Companies that sell less than million units of high priced chips ( e,g. FPGA ) are next in line for integration at Package level.
This is not the case for Consumer systems where the massive volume enables at least the leaders to build the latest Fabs and integrate everything ( as in a SoC ) on a single small chip. As some of the leading Fabless companies who have recently dabbled in 3D stacking etc. have found out, cost would be a big deterrent.
So for them its back to the Intel single chip approach ( but built at offshore Foundries )
IBM has been filling B323A in East Fishkill N.Y. for years with tooling for advanced nodes. It seems to have been their plan all along. They are running 22nm and 14nm development in pretty good volumes on the latest immersion tools. I don't understand the reference to " has not invested in a leading edge Fab" for quite some time , there is still room to expand in the current building, and as far as I know, they can support at least the next two nodes there.