The main problem in wireless is, as metioned, the manufactring of the coils. However difficult it may sound, if the cost and complexity is less in wireless stacking then TSV will be difficult to implement. Stacking seems like the only way to keep Moore's law in relevance.
Die are stacked for desnity reasons, not for price. Sometimes you just need more memory than will fit on a board, or on the DIMMs in your already full DIMM sockets. If these DIMMs already have the densest die, then you are out of luck--unless a vendor offers stacked die.
This is also used by FPGA companies to increase density or to add analog functions (ultra high-speed SerDes) to a digital chip.
Regarding the queston: "Can someone explain how 3-D stacking saves money, because the silicon die cost remains the same?" Your assumption about "silicon die cost reamains the same" is incorrect for some stacking approaches. Thru Silicon Vias (TSV) offer nice improvements in reducing power compared to wire-bond, for example, but it requires a non-standard CMOS process that can cost substantially more, so die costs for TSV do not reamain the same. TSV is also hard to rework yield defects, which further increases costs. All of these challenges to build TSV combine to increase costs compared to the original non-TSV die typically by 1.4x to 2x. Ask your fab for actual costs for TSV. This often makes the TSV approach too expensive compared to the technical benefit. By using ThruChip's near-field inductive coupling approach to data communication between stacked chips, the die costs could essentially remain the same, as no new semiconductor process is required. With the ThruChip stacking approach, similar benefits to TSV can be achieved without die cost increase. Reductions in I/O circuitry and lower power with ThruChip could lead to cost reductiions as well. -- ThruChip CEO Dave
It is NOT just a matter of silicon cost, but cost at the system level. And even more so, even a 'feasibility' of certain applications. Besides the manufacturing headaches in executing a true 3D structure one is also faced with several others that may be the 'brick wall': yield, thermal, reliability, even FA.
There should not be any difficulty in "manufacturing of the coils" as the thruchip coils are made from normal wires. Don't think "wireless" as in radio, instead realize that this is just simple near-field inductive coupling. All designers have to worry about coupling between wires on their chips, but in this case professor Kuroda realized that by making a few simple turns of a standard wire, with diameter about 3x the distance to be travelled, that the coupling was so strong that it could be made as reliable as a wired connection. As you can see in the diagram, the circuits driving and receiving from the coils are relatively simple digital circuits, and will scale with future technology scaling as long as Moore's law continues. As there are no special CMOS process change required, in many cases there would be no increase to the cost of each die, which is not the case with other technologies like TSV. Completely agree with your point that this would act like an accelerator to Moore's law, allowing N times the number of transistors to be placed in the same area when stacking N chips. -- ThruChip CEO Dave
Great approach ! If it works then certainly beats doing root canal ( TSV ) on finished chips. But I am wondering if the chip to chip data transmission is going to be parallel or serial ? If it is the former then how about SSN and Signal Integrity ? If serial then will it be like SerDes and require adding the necessary circuitry to ea. chip ( in addition to the pulse creating and sensing circuits ) ? In either case what physics would limit the Bandwidth ?
If SerDes are needed it will dramatically increase the amount of circuitry required. Perhaps a better approach would be to use something like a memory array of inductors. Anyone else remember core memory?
@chipmonk0: On page two of this article, there is a table comparing these two technologies TSV vs TCI, which also mentions about the number of channels, area/channel and data rate. Does this help to answer your questions to some extent? I understand the communication would be using serial channels (Tx/Rx). Not sure about the pros/cons of increased speed upto >5Gbps though.
As replies in my earlier comment, it will not require a SerDes as it purely creates a Digital Coupling Link, serialization is not essentially required. But SSN and Signal Integrity will be a major issue near and needs detailed study, I think there will be surely some papers describing details of it, also the article says that there is an example of thousand such link on a single chip so it will be discussing about SSN and Signal Integrity in detail.
Hmmm. It has been many decades since EE school, but if I remember EE101 correctly - you need a change in current on one side to induce a change in voltage on the other side. A SERDES is one way of doing this, but since we have been told (in this article and subsequent comments) that it is a simple circuit to do this, I had to dig deeper.
Then I found a paper in ISCC 2009 titled "Wireless DC Voltage Transmission Using Inductive-
Coupling Channel for Highly-Parallel Wafer-Level Testing".
This paper explains how it can be done. The application in this case was for wafer-test, but the principles are the same.
Of course there are no silicon lasers. But there are modulators (need CW light to power them) some of them literally micron size and very fast photodiodes. For 1500 nm light to propagate through that ~ 50 um thich wafer one of stack is easy. Light will go right through and do not wreck any havoc inside the die (1 layer). The dimensions of needed diffraction gratings will be way smaller than sizes of RF links and silicon photonics can offer WDM on one beam. Given speed of those structures and very small size (and a few lasers bonded to top of 3-D structure) the throughput looks better than is wireless connection. Power consumption poses no problem. The crosstalk problem is way more easier to solve. Besides light beam goes through another and they are preserving their parameters. All these CMOS compatible. Thank you.
It would be a far more appealing demo to have the power signals, not just data, transmitted wirelessly. There might even be an energy harvesting opportunity here. Other questions still need to be answered as well. Do resistive or reactive losses in the coils limit the bandwidth? Do the wireless links allow conventional thickness silicon? What are the multiplexing options?
1. Yes, it is appealing to demo wirelessly sending power as well as signals. This has been demonstrated in several published papers already, but works best when the power consumption is very low. For many cases, simple wire bonding for power/ground is sufficiently good and low cost that you would have to very carefully examine the tradeoffs to determine when to use wireless power distribution vs wired.
2. According to professor Kuroda, resistive and reactive losses are not big effects. Even using what is normally highly resistive metal0 available in some processes seems to work fine. Bigger issue with bandwidth is simply the quality of tranistor, which is better in high speed digital processes and not as good in processes for DRAM or NAND FLASH.
3. Yes, conventional thickness silicon can be supported, so long as the diameter of the coils is about 3x the distance to be travelled wirelessly. But anyone serious about 3D stacking would probably use die thinning, which can result in die an order of magnitude thinner die than standard wafer thickness. See the slide titled "Range Extended by Enlarging Coil Size" and you can trade off coil size with distance and bandwidth.
4. You can put many channels in parallel, or use a Ser/Des to reduce the number of channels between die.
Yes, the coil impedance losses and required dimensions are nontrivial considerations in the design. On the other hand, TSVs are not a walk in the park either. I think the potential to NOT continually thin the silicon is quite promising. Not to mention the fragility or increased thermal risk, but there is less competitive difference from monolithic approaches.
I wonder about the "potential" for external electromagnetic interference. Slight offsets in positioning might also result in changes of signal amplitude with the risk of digitizing errors. Would it be better to use light communication between the chips (optoisolators)? When encased, there would be no risk of interference, slight offsets or changes in separation would not be an issue, and very high frequency signals could be conveyed without any inductance issues.