FMC, a Dresden, Germany-based fabless Ferro-Electric Memory company, has raised €100 million ($116.2 million) in C-round funding to take its FERAM chip technology into AI datacenters and replace DRAM and SRAM, and also succeed as a storage-class memory supplier where Intel, with Optane, failed.
FERAM has similar speed to DRAM and SRAM but is non-volatile and uses less electricity. FMC was founded in 2016 as a spinout from the Nano- and Microelectronics Laboratory (NaMLab gGmbH) at the Technical University of Dresden (TU Dresden). The founders were the now departed CEO Dr Stefan Müller and VP Marketing Menno Mennenga. FMC raised Secured €600,000 ($697,000) seed funding from High-Tech Gründerfonds (HTGF) and €4M ($4,650,000) Series A from eCAPITAL/HTGF. A B-round in 2020 brought in €17.2 million($20 million). Now it has raised more cash taking the total raised to $141.6 million; serious money.
Of the €100 million ($116.2 million) raised, €77 million ($89.5 million) in equity comes from FMC’s oversubscribed Series C financing round, which is backed by prominent existing and new investors and ranks among the largest raise of its kind in the semiconductor industry. The other €23 million ($26.7 million) came from public funds, including contributions from the IPCEI ME/CT program and the European Innovation Council (EIC).
CEO Thomas Rückes said: “We are working on the next generation of memory chips and system solutions that are not only more sustainable and energy efficient, but also faster and less expensive than the current industry standard. While bandwidth has so far been the dominant metric of AI compute, energy efficiency is now becoming the key factor for the next generation of AI.”
We had thought that FERAM technology was stuck in the same trap as Optane; promising technology but with no viable route to the high production volumes necessary to provide low enough pricing for affordability and supplier profitability.
Perhaps FMC is finding a way through this technology-market obstacle course, with the key being the memory electricity needs of the GPU servers inside massive AI data centers.
Rückes asserts that: “Memory chips are the main bottleneck in the AI stack. FMC’s DRAM+ and 3D CACHE+ technology addresses precisely this issue: Faster and more energy efficient than established products. This lays the foundation for scaling up AI data centres and AI edge applications. Securing an equity financing of this magnitude emphasizes the significance of our technology, and we are grateful to have earned the trust of leading deep-tech investors for our vision.”
FMC has two products. The DRAM+ product technology is aimed at replacing DRAM and offers non-volatility (= less electricity) with memory also being storage, and more endurance. FMC’s CACHE+ tech is intended to replace SRAM, and offers 10X SRAM density, and reduces standby power10x as well, while being non-volatile again.
Both are fabricated with standard CMOS transistors and capacitors being transformed into FeFETs (ferroelectric field-effect transistors) and FeCAPs (ferroelectric field-effect capacitors) using existing semiconductor manufacturing equipment. That’s a plus but manufacturing feasibility is not the big problem as we understand it. The elephant in the room is upper supply chain acceptance.
We might imagine DRAM+ replacing DRAM conceptually but the cost feasibility of this will require an x86 server manufacturer to make a huge bet in building it into their servers. And a server operating system supplier to rewrite the memory-handling portion of their code, and – yes, there’s more – system application suppliers re-writing their code as DRAM+ will be both memory and storage, so persisting memory contents to storage for safety will no longer be necessary.
If we think about GPU servers and their HBM, then an HBM manufacturer would need to make a huge bet as well, and replace the DRAM stacks in their HBM with DRAM+ stacks. This is not totally off the wall as HBM supplier SK hynix is an investor in FMC. But that’s only half the story, as the GPU supplier would have to rewrite the memory handling part of their operating system on top of this.
If Nvidia decided this was the way to go then job done. Nvidia is so massive that it would move the market. If it doesn’t take that view, then FMC has to persuade AMD to adopt its FERAM and that would have a smaller effect on its market in terms of chip volume.
B&F thinks that, for FMS to replace x86 server DRAM, is more difficult, as the electricity supply problem doesn’t affect the manufacturers so much, and there being, we understand, fewer x86 CPUs in AI data centers than GPUs. Also, no one server manufacturer has as much market moving power as Nvidia.
FMC says its fresh funding will accelerate the commercialization of the company’s DRAM+ and 3D CACHE+ memory chips and system solutions, and expand its global presence. It claims its technology “builds on its superior energy efficiency and will facilitate the global ramp-up of AI data centers and AI edge applications, setting a new industry standard in the €100+ billion memory chip market.”
The investors and grant givers in FMC’s latest fund raise certainly think it’s a bet worth making. Are they right, and can FMS’s storage-class memory succeed where Optane did not? We should know the answer by 2030.
