Processor Whispers: About possible foundries and impossible processes
by Andreas Stiller
The start of December is when the International Electron Devices Meeting IEDM traditionally takes place. Along with the presentation of numerous new designs (with and without silicon), its agenda always includes an overview of the latest process technology.
This time at the IEDM in San Francisco, among other things, Intel released details about the low-power 22nm process for SoCs, which provides a choice of six different transistor types. Thanks to new tri-gate transistors, an SRAM test chip manufactured using them shows a leakage current of only 10pA/cell at a clock speed of 2.6GHz.
This new SoC process is probably primarily intended for the next Atom generation, Valley View or Avoton with Silvermont architecture, scheduled for the end of 2013. Meanwhile, however, Intel has started to also manufacture chips for external companies. Currently, three of the smaller companies, Achronix, Netronome and Tabula, are known to rely on Intel as contract manufacturer but, during an interview at the last IDF, Mark Bohr, head of manufacturing, hinted at there being more, "non-disclosed" companies already and the possibility of many more joining the list soon. Under discussion are Cisco, Xilinx and Altera, and now Intel is also being tipped as potential contract manufacturer for Apple's next or second-next generation of ARM processors – supposedly in exchange for Apple switching to Atom processors for future iPads.
At the Bernstein Technology Conference at the start of December, Otellini – for now still the CEO of Intel – announced that, as a contract manufacturer, his company does not intend to compete with TSMC, but that there is a lot of stuff in the pipeline. He declined to comment directly on the production of ARM processors, but emphasized that Intel is open to strike a deal as chip manufacturer for the right strategic relationship with the right client.
At the IEDM, there was no mention of such things, but of disappointing delays. For instance, Achronix boss John Holt conceded that the 6 million transistor FPGA HD1000, promised for the end of 2012, will enter sampling in the first quarter of 2013 at the earliest. Also Intel's Haswell processors will probably only arrive later in the second half of 2013. And the chips to follow in the 14nm process – Broadwell and Airmont – will only be released at a later time in the course of 2014. As a result, 600 Irish employees who were undergoing training in the US for the fab in Leixlip, where the chips should be produced, have been sent home. In spite of what was reported here and there, Intel's spokesperson Dan Snyder told the Maximum PC web site that Broadwell and later chips will continue to appear as a socket version for the "foreseeable future".
Haswell's desktop family will definitely be released as a socket version, presumably for the new socket LGA1150. This was learned from a roadmap leaked to the Chinese web site chinese.vr-zone.com that lists the 14 planned Haswell desktop versions for 2013: with two or four cores, with and without hyperthreading, with clock speeds between 2.7 and 3.5GHz, with 4 to 8MB of L3 cache – and all of them equipped with HD 4600 graphics. Probably this will be the moderately powerful GT2 graphics with only 20 execution units (EU), which would mean that the GT3 with 40 EUs would be reserved for the high-end notebooks.
Naturally, Globalfoundries was also present in numerous speeches at the IEDM. They hope to have the 14XM process with 14nm FinFet transistors up and running by 2014. The head of Globalfoundries' factory in Dresden, Rutger Wijburg, is meant to take care of that and is now in charge of the new fab in the state of New York as well. Together with its partner Toppan Photomask, Globalfoundries intends to revive the jointly owned photomask factory AMTC in Dresden, which has been suffering some neglect since the bankruptcy of Qimonda. They extended the joint venture by five years and plan larger investments as well as an increase in staff in the coming years. The Center Nanoelectronic Technologies (CNT), on the contrary, which is also located in Dresden and which is led in cooperation with Fraunhofer, has fallen by the wayside. Some weeks ago, the Fraunhofer-Gesellschaft decided to dissolve the CNT as an independent institution at the end of the year because of a lack of economic viability. Profitable areas are going to be merged with the department for photonic microsystems.
Globalfoundries' primary client, AMD, is currently struggling with falling demand and has therefore renegotiated the Wafer Supply Agreement. Already in the current fourth quarter of 2012, AMD is buying fewer wafers than initially ordered and will therefore be making a termination payment of $320 million. For 2013, the agreement foresees purchases of $1.15 billion. A closer look at the agreement also reveals a hint that AMD plans to switch to a standard 28nm CMOS process, leaving behind the current silicon on insulator (SOI) in favour of bulk. All microprocessors are going to be manufactured by Globalfoundries. All in all, AMD wants to focus on a small number of products with high sales volumes. As of now, AMD is having its chips produced by two foundries (Globalfoundries and TSMC) and in nine different processes. In the future, AMD CEO Rory Read plans to have only two manufacturing processes for all AMD products.
While AMD is apparently set on moving away from SOI, Globalfoundries' development partner IBM is sticking to SOI and also to "gate first" for the 22nm SoC process – the latter comes as a surprise as it was generally deemed certain that IBM would switch to the "gate last" procedure favoured by Intel. Another bit of news is the insertion of carbon as doping element. This way, the silicon-germanium composition can be put under even more stress in order to achieve a higher charge carrier mobility – about 25 per cent higher than in the current 32nm process. Until recently, carbon was considered to be too unstable, but apparently IBM got it under control.
However, in spite of all advances, Moore's Law is finally running out of steam, as IMEC's lithography specialist Kurt Rose explained in his keynote. Because of a lack of EUV lithography suitable for mass production, technology pioneer Intel has to use lots of tricks and magic to realise the 14nm process with 193nm laser technology, which, however, is less efficient in terms of performance – only 15 to 20 per cent in comparison to the previous generation, instead of 30 per cent as was usual until now. Additionally, this increases production costs because three exposures (triple patterning) are necessary instead of just one, as would be the case with EUV. According to head of manufacturing Mark Bohr, Intel has somehow even managed to find a way to go down to 10nm with 193nm light, but they hope the severe problems with EUV can be solved before then – if not, the curve of Moore's Law will probably suffer a rather substantial dent.