Pourkeramati joined FMC as CEO in 2019 (see Ferroelectric Memory Co. puts Spansion veteran in charge) having held several senior positions at Spansion Inc. and Cypress Semiconductor from 2005 to 2015. He had founded flash memory company Azalea Microelectronics Corp. in 1995 and was its CEO up until 2004.
FMC’s latest backers include Merck, SK Hynix, IMEC, Robert Bosch and Tokyo Electron plus existing investor eCapital (see SK Hynix backs Germany’s ferroelectric memory startup). These companies wanted to provide more than $20 million, and there were others including leading foundries that wanted to invest, but FMC declined these additional offers for now, Pourkeramati said.
We will come to that later but Pourkeramati’s first point was to praise the properties of hafnium oxide as a ferroelectric material that also happens to be in widespread use in wafer fabs as an insulating layer in modern ICs. And he claims the ferroelectric field effect transistor (FeFET) is a natural replacement for embedded flash which struggles to scale beyond 28nm.
Unlike embedded flash, FeFET’s endurance is high at 10^11 cycles and is expected to go closer to that of MRAM at 10^15 cycles and with switching speeds of less than 1ns and switching energy of less than 1fJ per bit, Pourkeramati points out.
The FeFET works by using non-volatile ferroelectric property to shift the transistor threshold voltage. Source: FMC.
The fundamental switching speed is very high but in the array it does depend on the loading of the word and bit lines. We have a 32Mbit memory macro at about 10ns. If you want to read faster you go smaller and more distributed so there is a design aspect to it,” said Pourkeramati.
Next: Logic friendly
“This is a logic friendly technology that requires two or three non-critical layers. For the logic and foundry guys this is heaven,” he added.
Add to that high temperature stability, the scalability and familiarity of the materials and it might seem that FeFET faces an easy win.
But there is one figure of merit that FMC needs to drive higher. On his chart of measured and published characteristics of FE-HfO2 data retention stands at 1,000 hours at 125C. The expectation is that 10 years at 175C can be reached, but there is clearly work to be done.
As a refinement of high-k dielectric the ferroelectric material is equally applicable to gate-all-around and nanosheet or nanoribbon transistors. Source: FMC.
A second objection is that MRAM – and in STMicroelectronics case phase-change memory – has already been adopted by leading foundries for embedded memory at 28nm and below so the fight is not with embedded flash. “MRAM needs additional materials. Ferroelectric memory needs no special materials and scales with logic so we can go to FinFET,” riposted Pourkeramati. MRAM is also susceptible to magnetic fields.
Next Pourkeramati emphasized that ferroelectric memory can address embedded non-volatile memory for cache memories as well as stand-alone chips for storage-class memory. FMC has also done some preliminary work specific to a 3D stacked form of memory. “We are scalable to n layers,” said Pourkeramati. And it can be used for neuromorphic applications. “We have created IP [intellectual property] for all applications.”
Next: Business model
FMC’s business model is one of licensing intellectual property and taking royalties on a per chip basis. And to this end the company has divided its offering in to three strata. One is process-related IP such as the hafnium chemistry recipe. The second is transistor device level IP. The third strata of IP is design or circuit based. “We license these various pools of IP to the foundries, the fabless chip companies and to the IDMs [integrated device manufacturers,” he said.
Roadmap for the introduction of FeFET technology. Source: FMC
Pourkeramati said first silicon for consumer applications will come in 2023 and on 28nm planar CMOS. A move to 22nm FinFET and/or FDSOI would happen after that and with other sectors coming on in subsequent years.
Considering FMC was founded in 2016 and it is now 2020 FMC’s roadmap appears somewhat conservative. “These things take time; time in development and time to market. We think this is reasonable and acceptable. If you want to develop in FinFET then work with us to go through the nodes as quickly as possible,” said Pourkeramati.
It is noticeable that ferroelectric memory is becoming a hot topic in academic circles and numerous large semiconductor companies are talking about the technology (see IMEC, ferroelectrics prominent in virtual IEDM program and Sony, Kioxia pursue ferroelectric non-volatile memory). Is there a risk that with FMC having done much of the pioneering work with hafnium oxide and zirconium-doped hafnium oxide for memory others could reap the benefits?
Pourkeramati said not and that FMC’s patent position is comprehensive covering both hafnium and zirconium.
Next: Come back later
Which leads us back to round to why would Pourkeramati have declined or deferred some investment. There is a well-known aphorism in startup circles that it is best to take venture capital when you don’t need it. Because when you do need it, it won’t be available.
Pourkeramati has worked in venture capital having previously been a managing partner at Influence Capital Partners Inc. immediately prior to joining FMC in 2019 and puts the counter argument. He argues that FMC will be able to raise more money at a much higher valuation in times to come.
There were additional companies that wanted to invest in FMC. Similarly, there were foundries that wanted to invest in FMC but Pourkeramati said he was concerned that being closely aligned to any one foundry might inhibit others from embracing the technology, and so the company declined.
With FMC eager to attack multiple sectors simultaneously and in the game for the long-haul, those investors and others will no doubt have further opportunities to invest, but FMC needs to continue to make progress as an alternative to the already deployed MRAM.
Related links and articles: