It is reported that the startup Niobium has raised $5.5 million in seed funding and launched its first FHE (fully homomorphic cryptographic accelerator) chip. The company says Niobium's first chip will make FHE 1,000-5,000 times faster than standard CPUs.
Fully Homomorphic Encryption (FHE) is an advanced encryption technology that allows various computations to be performed on ciphertext without decrypting the data, and the result of decryption is the same as that of corresponding computation directly on the plaintext. However, due to the very large amount of computation, FHE is not currently achievable. Companies including Cornami, Optalysys, Niobium, and Intel are all planning to launch FHE hardware accelerators.
As for whether the 1000-5000 times performance increase can achieve a usable FHE, the person in charge said that they believe that this performance increase is feasible and have already worked with some large companies.
"We have been greatly helped by the government's support, which allows us to tape out the chips and then test them, package them, and put them on PCIe cards," said the person in charge of the project. He added that the company has also raised $5.5 million in seed funding, which will be used to invest in Niobium's software design and development team. The funding will also be used to develop commercial applications of FHE, including healthcare and pharmaceutical research, financial fraud detection, blockchain, and digital advertising.
Niobium's first chip features the government-backed GlobalFoundries 12nm low-power process node. Its accelerator has a highly parallel architecture similar to a DSP, and its MAC unit is based on a proprietary instruction set architecture.
figure:Niobium raises $5.5 million to launch FHE chips
The head of Niobium said that its design was conceived for FHE from the beginning, and the project is in its infancy, so there is a lot of room for development. The person in charge said that the adoption of software will be a major obstacle for FHE, and Niobium has already started developing a software stack, and the project will be completed by the end of 2024.
"We think we already have a very compelling software suite, but it's not necessarily application-related. We're going to continue to do compiler optimizations, but over time, some customers will want to do more on the application side, so we've worked in the ecosystem to enable third-party software companies to do the specific application development that we think is needed. "Currently, Niobium's stack is connected to the available FHE libraries.
One of the most compelling use cases for FHE in AI is to train models with cryptographically sensitive data from multiple sources, such as medical records. The person in charge said this is a valid use case, noting that inference on encrypted data is one of the benchmarks for the company's existing DARPA contracts. Training of encrypted data and encryption models is also possible, he said.
