2024.05.14 Tue
Hidenobu Mori
In modern society, where huge amounts of data are exchanged, encrypting confidential data is one of the most important issues. Accelerating Fully Homomorphic Encryption (FHE), a type of encryption method, has been considered a means of achieving secure communication of encrypted data sets, but it requires intensive calculations and has not been commercialized. The problem so far has been that it is not suitable. Cornami (headquartered in Texas, USA) is a startup that has developed “TruStream” technology, a reconfigurable massively parallel computing fabric architecture that can solve this problem. We asked CEO Dr. Walden C. Rhines about the company’s unique technology.
The founder’s knowledge that made the “impossible” possible
– Please tell us about your background before joining Cornami.
I spent most of my time as a student through graduate school studying in the field of semiconductor engineering. He then spent 21 years at Texas Instruments and became CEO of Mentor Graphics, a provider of electronic design software. After 23 years there, the company was acquired by Siemens.
During that time, I had the opportunity to explore some things, and was commissioned by the US Defense Advanced Research Projects Agency (DARPA) to research FHE. DARPA was trying to apply FHE to cybersecurity. Therefore, I was asked to research companies that were working on FHE, but there were no such companies.
DARPA then launched a program to provide funding to companies developing the software and hardware for this technology. Later, I reconnected with an old friend, Gordon Campbell (Founder and Executive Chairman of Cornami), who runs a start-up company dealing with highly innovative software and hardware technology. They were mainly doing machine learning AI inference, but they also said they could do FHE, which is something that no one else can do.
I didn’t think they could actually do it, but after studying their data and simulations for about five months, I concluded that they could, in fact, make FHE a reality. Then they appointed me CEO in 2020.
-What is the history behind Cornami’s founding?
The company was founded by entrepreneur Gordon Campbell and engineers Paul Master and Fred Furtek. Master and Furtek developed a unique compiler technique that vectorizes computer applications, dividing them into completely independent streams of control and data. This technology allows us to create scalable parallel computers.
Making a computer faster is as simple as increasing the number of streams, chips, or cores, but traditional computers have many bottlenecks due to Amdahl’s law, especially shared memory and synchronization problems. To address these challenges, we spent seven years optimizing our software and ultimately designed a chip to implement this software.
Typically, we create innovative chips first and figure out how to program them, but Cornami is different. We first spent seven years optimizing the programming and then developed the chip. We built an emulator and tested the application for about three years, and the results were very positive. We plan to release this chip in production soon, and we also plan to introduce servers in 2024.
Encryption mechanism that can withstand quantum computers
-Please tell us about FHE’s technology.
With FHE, data owners encrypt their data using highly secure encryption methods. It is quantum computer proof, provides absolute protection and you don’t have to worry about data regulations like GDPR (EU General Data Protection Regulation). This high level of security allows you to transfer encrypted data to the cloud and perform calculations while the data remains encrypted.
FHE encryption does not require you to give your key to anyone, so no one else can see your data. The results are also sent back in encrypted form. That means only you can see your data and results. For this reason, the US Department of Defense calls it the “holy grail of encryption.” There is a non-zero possibility that someday all data centers and all operating systems will be hacked. The Department of Defense wants to ensure that no one can see the data when such an event occurs. FHE makes that possible.
Using this technology, machine learning models can be rented out while encrypted. Models can be provided without revealing model data to users.
-FHE processing seems to be a burden. That’s where a unique chip comes in handy.
Our processors need to run about a million times faster than NVIDIA and Intel processors.
Therefore, compared to NVIDIA and Intel, we have increased the number of cores per chip by about 50 times. Currently, we have 70 patents, and NVIDIA supports a type of parallel processing called “single instruction, multiple data.” Intel, on the other hand, supports a format called “multiple instructions, multiple data”. There are 55 other forms of parallelism, and we support them all.
Our chips can perform computations on encrypted data almost as fast as Intel and NVIDIA chips can on unencrypted data.
-You say that servers will be provided in 2024, but please tell us about the demonstrations that have been made so far in the pilot stage.
We provide our clients with remote access to our servers for processing. It is a SaaS-based server. We also sell servers and chips if the client wishes.
Also, the client runs the application on our emulator. This emulator is a computer that executes the chip’s logic. Several clients are currently completing evaluations with the emulator. We are particularly pleased with the positive results we have achieved in areas such as financial services, medical diagnostics, and improving manufacturing efficiency. Data security is important in these fields, but the advantage is that the data can be analyzed and shared without being made public.
Our clients are very excited and want the actual product, not an emulator. Our customers are very satisfied with the results as it guarantees security and speeds up their applications. Recently, we have customers who are leveraging generative AI using large-scale language models (LLMs). They are very security conscious and appreciate being able to protect their data.
We also have other customers using it for blockchain applications and other purposes. The people at Softbank, who have invested in our company (Softbank Vision Fund 2 led our Series C funding), taught us the importance of this technology. They explained that this technology has the potential to turn data and information into a multi-trillion dollar business. Even if you have valuable data and want to provide it to others, you are worried that it will be stolen. However, if you encrypt it, you don’t have to worry about that.
For example, Tesla developed an Autopilot feature for its cars, but it was not encrypted and was copied by competitors. American companies Meta and Google received large fines from the European Union (EU) for violating the GDPR because they kept data in their data centers. However, if the data is encrypted, even if the data center is hacked, there will be no problem. Furthermore, it is predicted that quantum computers will break traditional cryptography over the next three to four years, but FHE’s method cannot be broken by quantum computers.
FHE’s technology was mathematically complex and very slow, so many people considered it impractical. Cornami has proven its high performance and is now suddenly attracting attention.
-Do you mean that there is no other company that can bring FHE’s technology to a practical level?
There will be some competitors in the future. At least three companies are currently developing FHE-compatible chips, one of which uses optical processing technology. However, these companies develop chips specifically for specific algorithms. FHE’s algorithms are constantly changing, so a programmable chip is required.
What’s unique about Cornami is that our products are configurable. As technology changes, we can reconfigure the processor in software and optimize it to the algorithms we need.
There is no need to make any changes to your software during installation. If you’ve written an application that follows one of the standard interfaces and you want to do machine learning, if your application supports PyTorch or TensorFlow, etc., we can run that software directly without modification. .
Installing a server in a data center is the same as installing any other computer. Our servers and racks come with standard configurations, and if you have a rack with Intel or NVIDIA servers, you can add our cards to it.
Series C funding led by SVF2, with logistical support as well
– Please tell us about your goals for the next 12 to 24 months and your business in the Japanese market.
We plan to offer chip and server products, which is expected to increase revenue. We also plan to start designing products using more advanced technology. Although the technology used to date is relatively mature, a rapid transition to advanced semiconductor technology promises to improve performance, reduce costs, and provide many other benefits.
In the Japanese market, in March 2022, Softbank Vision Fund 2 led a $68 million funding round, which was used to develop chips and servers. Softbank not only provides funding, but also introduces suitable personnel and provides other advice. When the product goes on the market, we plan to provide marketing support, especially in Japan.
Additionally, we are seeking partnerships with companies in Japan and abroad who have customers who require data security. The targets are companies that provide servers and data centers. We are also in discussions with several Japanese car companies about embedded chips. However, since the first generation chips are for servers, embedded chips will not be available for some time.
– Security is important now that data utilization has progressed. Please tell us about your company’s long-term vision for achieving robust security.
We expect rapid growth, especially in areas where security is critical. This includes many of the categories I talked about, such as financial services, medical diagnostics, and government agencies. In the future, as the threat of quantum computers increases, everyone will want to move towards secure computing. Therefore, we expect our technology to become a part of most data processing around the world.
As people begin to realize the value of their data and data models, and as generative AI becomes capable of producing highly complex models, we believe that security when buying, selling, sharing, or using these models will become critical. I am. We believe we are at the center of this business and technology evolution, and we think this is going to be huge.
Read the full story here. (Original Japanese article)