You are here

Facebook Acquires Sonics

This week, Facebook announced it had acquired Sonics, Inc., a 3rd Party Semiconductor Intellectual Property (SIP) vendor. Sonics is one of the mainstays of the SIP market for Interconnect SIP, which is used to tie together the multiple tens or hundreds of SIP blocks found on contemporary System-on-a-Chip (SoC) silicon solutions today. Sonics was founded in 1996 and was at the heart of the emerging SoC market in its early years along with ARC, ARM, MIPS, Analog Bits, Virage Logic and many others.
It is no surprise that a company like Facebook, who is designing their own AI-focused SoCs, would purchase a company like Sonics.  While no details of this acquisition have been released, some reasoned insights are possible. 
Implications for the Market
This acquisition removes a major player for Interconnect SIP from the market. Sonics’ customers must now determine how long Sonics will support them and at what level. It is reasonable to expect these customers will look to other suppliers for their interconnect requirements.
However, Facebook could take the point of view that Sonics will continue to license their existing SIP into the market just as Wave Computing is doing with MIPS CPU cores. Any new products developed for Facebook would not be available to outside companies. While this would relieve Sonics’ existing customer base from having to scramble for a replacement, it is unlikely Facebook would allow new licensing of Sonics products to occur.
There are other major players in the Interconnect SIP market – Arteris and ARM. NetSpeed, another established, but newer, Interconnect SIP vendor was acquired by Intel in September, 2018, for likely the same reason as the Facebook acquisition of Sonics. Provino is a newer start-up in the same space focusing on interconnect solutions for SoCs with an emphasis on AI-focused SoCs.
It is possible that other start-ups will try to enter the Interconnect SIP market. The barriers to entry in the Interconnect SIP space are high since it requires a very deep understanding of not just silicon design, but system-level design to ensure high degrees of efficiency, data security and performance necessary for Advanced Performance Multicore SoCs. Arteris and ARM have extensive patent portfolios that make the job of a start-up in this space very difficult to get around without infringement and a reasonably high amount of investment. It is also likely that Facebook took this into consideration when profiling Sonics for the acquisition. Sonics has an extensive patent portfolio which would be difficult to navigate around when trying to duplicate their current products.
Arteris will now be considered the pre-eminent supplier of Interconnect SIP in the market. We should expect their revenues to accelerate even beyond the good growth they are experiencing.
While it might seem likely that Arteris would now also be a target for acquisition, Qualcomm owns the patent portfolio for the FlexNoc products dating back to October 2013. Arteris has an unlimited use license for its FlexNoc patents, as well as access to the source code and the right to modify it as needed. If another company were interested in acquiring Arteris, the patent ownership question would be a difficult one to answer. It is unlikely that Arteris would be acquired unless the acquisition price was very high and would satisfy Qualcomm’s long-term goals and requirements for this technology.
Renewed scrutiny will be placed on ARM’s Interconnect SIP products. There will be pressure on ARM to place more emphasis on Interconnect SIP solutions directly aimed at AI-focused SoC solutions.
Future Directions
In general, given the emergence of AI SoCs for training in the cloud and SoCs for inference in end-point devices, greater emphasis will be placed on vendors of Interconnect SIP to create solutions directly aimed at this class of semiconductor products. This is going to be an extensive effort since the memory interface and type of memory figure prominently in the performance of the silicon solution. Both Arteris and ARM have the technical resources to accomplish this.
Semico believes that, given the continuing evolution in the AI market, there will be a third architecture that will emerge in the near future that is a hybrid between the SoCs for training in the cloud and SoCs for inference operations in end-point devices. This hybrid will feature some limited training capabilities and have a heavy emphasis on inference capabilities.
We believe this will happen because, while the need for local processing of inference data in end-point devices has been acknowledged, the ability to add new targets for these inference operations is not yet possible except by re-training the algorithm in the cloud and then redeploying that algorithm back to the end-point device. We believe this will be too time-consuming and expensive as a long-term solution. A much better solution would be to have some limited training capability where it is needed in end-point devices. No such solution exists – yet, but we feel it is only a matter of time before such a development happens.
Recent product introductions by Crossbar and CogitAI also seem to be signaling a trend in this direction, and it would make sense from an efficiency and productivity point of view. This also plays into the overall trend of the fast-paced evolution of the AI market in general. Things are moving very fast with improvements in both algorithmic concept and design and silicon architectures. Further improvements in Interconnect SIP to support such a hybrid architecture would seem to be warranted.
It is Semico’s view that some level of AI functionality is going to be incorporated in almost every application and SoC silicon solution over the next few years. It would make sense for the Interconnect SIP vendors to also trend in this direction since Interconnect SIP is at the heart of every SoC today and has been for several years. The near future is going to be a very exciting place to be – indeed!