CEO interview: Flex Logix’ Geoff Tate on licensing FPGA
Flex Logix, founded in 2014, provides licensable field-programmable gate array (FPGA) fabric and has produced fabric cores for multiple manufacturing processes including TSMC’s 40, 28 and 16nm processes. “We are part of the TSMC IP Alliance to help provide access but we have also conducted a port of our technology for Sandia National Laboratory to a 180nm CMOS platform for radiation hard applications,” said Tate.
“We’re happy to port to additional process nodes and styles but we don’t do ports without a customer. We will port to what they want and bear that cost,” Tate explained.
Clearly 10nm and 7nm are the next FinFET nodes on TSMC’s roadmap. So how soon will Flex Logix be there?
“The customers who are thinking about 7nm are in wireless, networking, data centers and so on and from what we can see very few people are doing 7nm design right now. They have plans but they’ve not yet started. Also 7nm keeps evolving. If you start to soon you could end up being on a version that’s not much used. TSMC technology symposium in Santa Clara on May 1 is when we expect more information. And we expect it [the process] to arrive late 2018 early 2019.”
Tate added: “And for us 7nm is the next node. We haven’t had any customers asking for 10nm. We do need to anticipate the needs of the customers but in FinFET it takes us about eight months to do a port. If we start at with the customer we can be proven in silicon before they tape out.”
With the digital logic developing rapidly at present we asked Tate whether Flex Logix sees an opportunity in diversifying its IP offering beyond FPGA fabric?
“Well we have moved from a first to a second generation fabric – from a 4-input look-up-table (LUT) to 6-input LUT (see FPGA fabric offered for TSMC 16nm FinFET). But that’s not really a diversification. All our software improvements are now focused on the 6-input LUT. It’s better for density and performance.”
But what about additional structures to better support neural networks?
Next: Machine learning?
“We are not neural network experts. It is certainly a hot topic but there are a lot of people in that space. There are probably multiple markets and multiple solutions. And FPGAs are being used to implement neural networks,” said Tate.
“Neural networks tend to want more multipliers and smaller data types. We could get involved with an AI-optimized FPGA fabric. It would be more of a tuning than a total change in architecture. We have a 22bit by 22bit MAC [multiply accumulator] that is also two 11bit by 11bit MACs. It is there to support complex arithmetic but we could support its independent use.
The idea of applying FPGA fabric within system chips has been around for 20 years but did not gain much traction early on resulting with a number of startup companies coming and going. So we asked Tate if FPGA fabric licensing is taking off now when it didn’t gain traction 10 to 20 years ago?
“The density on offer in the past was not as good as what we can offer. Customers do want high density. Also integrating embedded FPGA in a chip is a lot harder than it may appear because of process complexity. Different users of a node can choose a different metal stacks and can have varying metal thicknesses And you have to be compatible with them. But the complexity of processes and the increasing mask costs also mean that companies need to get a better return-on-investment on their chips and are going for FPGA fabric to achieve that.”
Tate continued: “We see two main categories of application for FPGA fabric. 1) Customizing but without changing the silicon in the field. So there is one chip in manufacture that can be multiple chips; for example in an MCU portfolio. 2) Chips that will be updated in the field. Such as networking chips where the SoC maker is looking for longevity. They do want to update the chips and protocols from time to time.”
The third category is much more dynamic where the RTL is being swapped in out, Tate said. “Here you can imagine a series of accelerators being swapped in and out in response to the software work load. We’re just seeing these applications emerging right now.”
So where is the FPGA market going more generally?
“FPGA vendors such as Xilinx and Altera – the latter is now Intel’s programmable solutions group – are going for the fastest and biggest chips and high volume applications, and have added processor cores so their silicon becomes a platform. It’s quite a different market to the one we are going after.”
Next: Tate uses the ARM analogy
Tate makes an analogy to ARM’s competition as an IP licensor with Intel as a chip vendor. “ARM didn’t try to compete head-to-head in the personal computer. ARM went after different embedded markets and new applications.” Clearly in this analogy Xilinx and Altera play the role of Intel and AMD defining the high-volume product and selling ‘platform’ silicon in the form CPUs to meet the need. Flex Logix and others play the role of ARM offering to license in IP while leaving the customer free to define the application and the system chip.
Flex Logix has five licensees that Tate can talk about: DARPA and Sandia, a consortium that includes Israeli semiconductor and systems companies including Mellanox, Satixfy, DSP Group and AutoTalks under a license with Bar Ilan University; Harvard University – which is working on an AI chip – and SiFive Inc.
To illustrate the progess Flex Logix is making Tate said: “We have multiple customers apart from these and our competitors haven’t announced any.”
Will Flex Logix need to raise more capital? It raised $5 million in a Series B round in May 2017. “We are very cash efficient. That means more process ports in less time from less people. A lot of the series B is still in the bank and we see rising sales so we may not need a Series C,” said Tate. However, rather than write off the possibility he added: “There may be Series C later in 2018 or early in 2019.”
Tate said Series C money might be useful to pick up the pace of engagement with customers but that it would not be used for an acquisition to get into other areas of IP such as neural networks. “If you are working in gold mine why would you go out and work in a silver mine.”
However, a partnership is a possibility he said. “We are considering some partnerships but they or may not happen, so I won’t say any more right now.”
Related links and articles: