Monolithic 3D Shows Promise, Challenges

　　Monolithic 3D integration shows promise as a way to create faster, cheaper, smaller chips. But despite interest from several quarters, most efforts are still in a research stage with significant work ahead in proving the technology and building an ecosystem for it.

　　That was the take away from the annual S3S conference herewhere ARM, CEA-Leti, DARPA, Mentor Graphics and Qualcomm presented work in the field. M3D aims to carve diverse functions into blocks stacked vertically in one or more die, however most approaches have not yet demonstrated commercial viability.

　　The Leti European research insitute has worked for several years on Cool Cube, an approach for vertically stacking transistors. IBM, Qualcomm and STMicroelectronics are among Leti’s partners on the project. While it holds promise, it has yet to show how it can scale to problems such as global routing in a full M3D chip.

　　A Qualcomm research director co-hosted a half-day program on Cool Cube here. His enthusiasm for the area which he named 3D VLSI was clear, as was the vigor with which he is still exploring many shortcomings of current approaches.

　　Years ago, Qualcomm was equally vigorous exploring through silicon vias (TSVs), an approach many thought would be used for mobile application processors. But today, Qualcomm has largely dropped that effort because the thermal and cost challenges for TSVs in such chips remain largely unsolved. Meanwhile TSMC and others have made advances in wafer-level fan-out packaging now used in Apple’s iPhone processors

　　Startup MonolithIC3D has been championing M3D with its CEO, semiconductor veteran Zvi Or-Bach, acting as an organizer for the event. The five-person company with Or-Bach as its only full-time employee is said to have compelling ideas and IP for M3D, but it has yet to find a partner to test them in silicon.

　　DARPA sees M3D as one of many promising efforts to reinvigorate the chip industry at a time when it is seeing slowing gains and rising costs. One of six DARPA programs on next-generation silicon design will explore chip stacks, and some companies have said they will proposed M3D projects for it.

　　Top challenges for M3Ds include:

• Getting EDA giants to build design tools for it

• Protecting components from high temperatures required to make some elements in a device

• Aligning circuits with transistor-level accuracy

• Routing dozens if not hundreds of elements

• Reducing costs of layering memory and logic in a single device

　　Still M3D has potential, said Nathan Brookwood, market watcher at Insight64 who attended a Cool Cube workshop here.

　　“It’s like going upstairs and downstairs rather than from one end of a building to the other on the same floor. Instead of travelling tens of millimeters across a chip from a cache to a computing element, you could just go up one or two millimeters,” he said.

　　Leti and others point to exiting CMOS image sensors from Sony as a first example of commercial silicon using M3D techniques to link logic and memory. In January, Sony presented a paper on a three-layer stack with help from Samsung.

　　An Intel keynoter suggested the company is using a form of M3D to pack peripheral and memory circuits on its 3D XPoint and 3D NAND chips. Earlier this year, Samsung said it will put peripheral circuits on a future 3D NAND chip, suggesting an M3D-like approach.

　　To drive its Cool Cube approach forward, Leti is developing a PDK for a multi-project wafer so partners can test the technology, probably in late 2018. It is also working on a handful of chip-on-wafer and wafer-on-wafer stacking technologies.

　　Leti has demoed Cool Cube packing 2x107 vias/mm2. It also showed work managing thermal issues and precisely aligning Cool Cube structures. “We are convinced this is manufacturable,” said Olivier Faynot, a Leti section manager.

　　Or-Bach showed ideas for creating stacks of at least four devices using a combination of existing fab techniques he said were relatively low cost. He showed cutting out components with grinding processes and binding them face-to-face, using SiGe guide layers that are later etched away.

　　Existing bonders that can deliver alignment within 50nm tolerances were one element of his approach. He suggested his concepts could be used to build DRAMs that integrate memory and peripheral circuits in a single die.

　　Yang Du, a senior director at Qualcomm Research said M3D could be applied to a wide range of products. The technology could enable mating a memory array and an array of multiply-accumulate units to create an ideal neural network accelerator, an example shared by several speakers at the event.

　　In the short term, one of the top hurdles seems to be getting more EDA vendors involved in the work. A representative of Mentor Graphics showed the beginnings of efforts in parasitic extraction and verification for M3D structures. Qualcomm and ARM have worked with universities such as Georgia Tech on M3D design tools.

　　“There’s still a long way to go…We now have fine-grained 3D technology that looks very promising, and there are many challenges but they are not insurmountable,” said Mudit Bhargava, a representative of ARM Research.