//php echo do_shortcode(‘[responsivevoice_button voice=”US English Male” buttontext=”Listen to Post”]’) ?>
Fully homomorphic encryption (FHE) accelerator startup Niobium has taped out its first chip and raised $5.5 million in seed funding. Niobium’s initial chip will accelerate FHE by a factor of 1,000-5,000 versus standard CPUs, Niobium CEO Kevin Yoder told EE Times.
FHE is an encryption scheme that allows computation on encrypted data without decryption, with huge potential for the processing of sensitive data in the cloud. However, it is so computationally intensive that FHE is not currently practical to implement. Hardware accelerators for FHE are emerging from companies including Cornami, Optalysys, Niobium and Intel Labs.
Is a 1,000-5,000× performance boost sufficient to bring usable FHE within reach?
“We believe this is more than enough to get us going and we are working with some very large companies already,” Yoder said. “Our goal is to get as close as possible to real time with future generations.”
Niobium has spun out of cryptography and privacy R&D consultancy Galois to handle its development contract with DARPA as part of the Data Protection in Virtual Environments (DPRIVE) Project.
“The government contracts did a lot for us to be able to get here—we were able to tape out the chip and we will test it, package it and put it on a PCIe card using the money we got from those government projects,” Yoder said.
The company also recently raised $5.5 million in seed funding, which will be used to build Niobium’s software team, Yoder said, to begin to engage with the company’s first customers as part of an early access program. The program currently offers customers with FHE software access to a simulation of the Niobium hardware, but it will also offer hardware access when available.
“We’ve had five or six companies show interest in our early access program so far,” Yoder said.
The funding will also be used to develop commercial applications for FHE, including healthcare and pharmaceutical research, financial fraud detection, blockchain and digital advertising.
Room to grow
Niobium’s first chip is in a government-friendly GlobalFoundries 12-nm low-power process node.
“There is certainly headroom to go to more advanced process technologies,” Yoder said. “We haven’t had to make that decision yet. We think that what the current performance levels get us is very good insight from a number of applications out there in the financial community, in the insurance community, in some healthcare applications, and the feedback we’ll get through the proof of concept work we’ll do later this year will help us understand what we do with the architecture to get to the next level that’s required to open up more applications.”
Niobium’s accelerator has a highly parallel architecture resembling a DSP, with MAC units based on a proprietary instruction set architecture.
“Our design was conceptualized from the beginning for FHE, which makes a big difference,” Niobium chief product officer Jorge Myszne told EE Times. “We are at the beginning of this technology, so there is a lot of room to grow. 1,000 times is definitely a game changer. We see a lot of applications or proof-of-concepts that were done on the software, but when they try to move to production it’s just too slow, the functionality is there but the performance is not. With this accelerator, we will unlock all those use cases.”
There are other use cases Niobium may not have even thought about yet, according to Myszne.
“As always, there is a learning curve and a technology curve, and we have a couple more orders of magnitude to grow,” he said.
Software will be a major hurdle for FHE adoption. Niobium is already working on its software stack as part of the third phase of the DPRIVE project, which finishes at the end of 2024.
“We think we’ve got a pretty compelling software suite, but it’s not necessarily application related,” Yoder said. “We’ll continue to do compiler optimization, but over time, some customers will want to do more on the application side, so we’re already working on the ecosystem to have third-party software companies that can do specific application development that we think will be needed.”
Myszne said Niobium’s stack already connects with available FHE libraries.
“There are 4 FHE schemes at this point in time, and our accelerator is designed to work with all of them,” Myszne said, noting that the four schemes are essentially optimizations for different applications. One is designed for binary operation, and is used mainly for blockchain. Another is an approximate calculation that is mainly used for AI. Two others, which use exact math, are used for other applications. “We will be agnostic to the scheme and be able to work with anybody,” he said.
One of the most compelling use cases for FHE in AI is training models on encrypted sensitive data from multiple sources, like medical records. This is a valid use case, Myszne said, noting that inference on encrypted data is one of the benchmarks of the company’s existing DARPA contract. Both training on encrypted data and encrypting models are also possible, he said.
“Sometimes the model is more valuable than the data, and people are afraid of reverse engineering the model,” he said. “That’s something we’ll be able to do, too.”
Seed funding
Niobium retains rights to the IP developed under its government contracts, plus it has access to other IP Galois has worked on in the privacy space on concepts like zero knowledge proofs and multiparty computation. It can also use or hire key Galois engineers, if required, Yoder said.
While there are currently no specific restrictions in place regarding who Niobium could sell its future product to, Yoder said the company is “sensitive” to who they would give access to. Target customers for the early access program are mainly in the U.S. and Europe, he said. Feedback from the company’s early access program should start to come in later in 2024.
Customer pilots of Niobium’s technology are due to begin in Q4.
Source link
lol
