Common Platform – questions for keynote speakers
January 18, 2011, Common Platform Technology Forum, Santa Clara, CA—the three keynote speakers—Gary Patton of IBM, Chia Song Hwee of Global Foundries, and Stephen Woo of Samsung Electronics—responded to questions after their presentations at a press lunch.
Patton summarized the key points, that 32 and 28 nm production is now on-line. They have integrated hi-K metal gate into the flow and are in production on a gate-first process. The 20 nm process is coming along, but they still need to resolve some issues related to lithography and source-mask optimization. The 20 nm process will change to a gate-last process to better address user requests for size and performance. All of the partners have realized extensions in interconnect and packaging.
Because the process changes from a gate-first to a gate-last flow, they will have to start a new manufacturing learning curve. Continued scaling requires innovation, especially in the middle of the line parts of the process. The main substrate will continue to be bulk silicon, but SOI will continue to be offered to those customers who need the higher performance. SOI also has integrated embedded DRAM, which is the on-chip memory of choice for memory arrays of up to 100 MB. 85-90 percent of their high performance users want the high performance and high memory bandwidth available from the SOI offerings.
One promising alternative is ETSOI (extra-thin silicon on insulator) which has the advantage that it is a planar process very similar to those currently in production. Packaging changes and the movement towards 3-D packaging leads to thermal issues, but new materials for the packages and interposers are improving the thermal conductivity
Hwee noted that some potential users are considering skipping the 32/28 nm node and jumping directly to the 20 nm process. Their biggest issue is the need for a process qualification. At the same time, the key driver for most users is time to market, so the 28 nm process, which is a shrink from the already qualified 32 nm process, requires no changes to the fab or manufacturing flow. All of the partners are trying to develop hedging strategies to balance capacity and demand for the processes.
Unlike Intel, who is 5-6 quarters ahead of the rest of the industry, the majority of users are more interested in total integration and lower power rather than the absolute highest performance like Intel with its CPUs. The more significant driver is time to market.
Woo opined that most customers will stay on the 32 nm node for a while. If the customers ask for it, they will be offered the 28 nm shrink.
In response to a query about the fab synchronization program, Hwee replied that the partners all use the same machinery, but cannot readily share masks because of differences in equipment setup. As a result, they only have GDS-II compatibility. Because they have matching electrical characteristics, no redesign is needed to transfer to one of the other fabs, just get a new mask set.
Woo added that the partners offer multiple sources and a track record of common and shared customers. Patton noted that they also have common test sites and alignment markers on the wafers so it is possible to run a partial wafer through a fab then complete it at another fab. All three agreed that the Common Platform and the Joint Development Agreement partners reduce the need for customer resources by having a single target process. The technology, process, and data sharing enables second sourcing and reduces risk for the users and the foundries.
At the same time, being competitors and partners requires a lot of work, especially when working with common customers. The whole business becomes more complicated and requires trust and a drive from the top of the corporation to succeed. At the same time, all of the parties involved have more to lose than to gain by these practices.
The companies forecast their production capacity and suggested tight capacity this year. Hwee stated that the 45/40 capacity is ramping up and is marginally keeping up with demand. 32/28 is limited for the next two years. Overall, ’12 depends upon demand. Woo is focused on low power designs and is seeing increased demand. He expects capacity to be tight in ’11 and ’12.
