IT companies have spent billions of dollars creating ways to move data more efficiently and cheaply in an increasingly distributed world of on-premises datacenters, multiple clouds, and the edge. This has compelled the development of advanced protocols, interconnects, and fiber optics to grease the network skids and compression algorithms to reduce the size of the data, enabling more of it to be sent at one time.
There is a lot of incremental evolution of these technologies, but Charles Yeomans, chief executive officer of Atombeam Technologies, says his seven-year-old company has a unique way of crunching down and transmitting data.
Last July, Atombeam unveiled its first product – Neurpac – which is central to the company’s data-as-codewords strategy of shrinking the size of data being transferred by an average of 75 percent in near real-time, resulting in an average four-times increase in effective bandwidth.
Finding more efficient ways for managing, storing, and moving data will become increasingly important as the cloud, the Internet of Things (IoT), and edge computing continue to rapidly expand, Yeomans says. Statista analysts expect the amount of data created globally will jump from 149 zettabytes in 2024 to 394 ZB in 2028. Neurpac supports the MMQT machine-to-machine protocol that underpins IoT.
“That’s a problem that’s getting to be more and more acute because everybody’s always thought about fixing this by saying, ‘We need bigger pipes, we need more satellites, we need faster processors, we need more processing on the edge,’ all those things,” Yeomans tells The Next Platform. “But nobody’s ever sat down and said, ‘What about the data itself? Can you do something to make the data more efficient?’ That’s where we come in.”
Data Compaction, Not Compression
There is data compression, but that’s primarily for storing data, he says. Atombeam’s technology, which uses AI, machine learning, and other technologies, reduces bandwidth, compute power, and energy during transmission. Neurpac, which comes with a SDK, also makes data randomly addressable and searchable, which Yeoman notes typical compression can’t. In addition, it addresses the issue of security by offering built-in deep data obfuscation, a key point when talking about IoT data, 98 percent of which doesn’t have security, Yeomans says.
Yeomans and Asghar Riahi founded Atombeam in 2017, with Yeomans bringing years of executive experience with firms such as insurance brokers Portal Group Holdings and Frenkel and Co, and he spent nine years leading biotech company Trigemina before launching Atombeam. Riahi came with deep systems experience with a range of well-known IT companies, including Red Hat, Siemens, Seagate, and thirteen years with pre-split Hewlett-Packard as master technologist. He left Atombeam in August 2023.
Codes And Codebooks
Neurpac is a cloud-based platform that includes an encoder, decoder, and trainer that can integrate into an organization’s cloud infrastructure. According to Atombeam, the trainer, using AI and machine learning, creates a set of small codewords – a codebook – with each codeword about three to ten bits in length. The codewords correspond to larger patterns found in a data sample – usually 64, 128, or 200 bits long. The codebook is installed on both ends of the communication link – at the sending and the receiving ends, connected by a satellite.
When a message is sent, the encoder – running as an edge module – uses the codebook to quickly replace the content with codewords. It’s the same information, but smaller, and the decoder, a server module, at the receiving point restores the original data using the codebook. The system’s gateway supports the MQTT protocol and can use others.
Yeomans says it is unlike data compression, which re-encodes information using fewer bits than the original. Because it’s on an “individual-only, one-at-a-time basis you can’t really use the compressed files for anything other than storing. If you said, ‘I’ve got a datalake full of this data and I want to do some research on it – I want to find out why this part is burning out on this tractor – just finding the data from those particular line of tractors is really hard, so 78 percent of the time the analysts spend is in finding and cleaning up the data; very little is actually analyzing it,” he says.
“Instead of in each individual file finding these patterns, we use machine learning to look at thousands of these little files and we look at it all at once. What we’re doing is in some ways the same thing compression is doing. We are looking for patterns that repeat, but we’re not looking in one file. We’re looking in thousands all at once as a sample.”
Atombeam’s technology is protected by 27 issued patents and 16 others that are pending, according to the company’s fundraising site.
The security comes via the obfuscation, which Yeomans says is not encryption, “but it obfuscates the data to an extent that it would take a national effort to hack into a codebook. That’s because, even though it’s a substitution cipher, it’s thousands of randomness, and the randomness is based on patterns in data, not on letters or numbers or so forth.”
The company is pulling in high-profile customers, including Saab, the U.S Air Force, and CrowdPoint, a blockchain company. In addition, partners include Intel, data engine provider Cribl, and satellite company Inmarsat.
Lots Of Investors, Lots Of Money
The company is taking the path less traveled to fund its work. Rather collecting money from handfuls of venture capitalists and other such firms, Yeomans and Riahi turned to StartEngine, an equity crowdfunding platform, with an initial goal of $20 million. The company last month said it has raised more than $15 million from more than 5,000 investors. In July 2024, Atombeam launched its third campaign on StartEngine and because of the response – more than $3.1 million raised since – it extended the deadline from December to January 30.
Yeomans said the crowdfunding decision comes in part from a natural inclination toward democratizing capital and because of the experience of Atombeam’s leadership, there wasn’t the need for the advice and direction that tends to come with taking venture funds. That said, he says it’s that as the company’s ambitions grow, “we could be looking at taking in a bigger slug. And if we have to do that, then most likely will be institutional.”
Private Clouds And Encryption Up Next
Those ambitions include the next generation of Neurpac, which will be released by mid-2025. The next generation of the product will allow companies to use Neurpac in private clouds. Right now the data goes into Atombeam’s cloud, but companies like Chevron, ExxonMobile, and BMW want what Neurpac offers, but don’t want to send their data to Atombeam’s cloud, Yeomans says.
The company also has contracted with the US Air Force to create an enhanced encryption technology – Neurpac+ – that can automatically be turned on once the data is processed by the trainer and will be stronger than traditional encryption but use much less compute power. Atombeam already has developed prototypes and the capability will be integrated into the next-gen Neurpac, the CEO says.
It also will play a role as Atombeam works to puts its technology into chips and other components. The company last month announced a partnership with Ericsson to integrate Neurpac and Neurpac+ into the networking giant’s Cradlepoint routers and NetCloud. The company is working on a reference design that could be used for any chip, Yeomans says, noting that it’s also gotten interest from military branches, including the Air Force, Navy, and Space Force.
“The cool thing about Neurpac is that it will work on a 10 cent processor,” he says. “It’s nothing. It is just doing a lookup and so it will work. Our plan is to put it into chips and have the chips cheap, so for anybody building anything – like a phone – it will work.”
Neurcom On The Way
By the end of this year, Atombeam also plans to release Neurcom, a complementary technology to Neurpac that will focus on video and audio codecs that will accelerate data throughput and improve the quality of the image or audio. It’s being developed with the Air Force Research Laboratory and according to Atombeam can double the number of high-density images that are transmitted over a network.
Where Neurpac is lightweight, Neurcom will be different.
“We take a codec and we plug a neural net AI into it to quantize the data,” the CEO says. “This is going to be fairly heavy duty because it’s got this neural net and it’s a codec and a video codec, but the net-net is that it marries well to Nerpac because it’s doing the things that Nerpac doesn’t do.”
Right now, the focus of Neurcom is synthetic aperture radar (SAR) images, but the Air Force is looking to apply it to other areas, including hyperspectral images and thermal environments, he says.
An interesting and much needed tech imho, but also a bit of a head scratcher. Could this be viewed as a homomorphic variant of LZ78 for example, where codebooks replace dictionaries, and are derived through ML (or is it something completely different and non-analogous I wonder)?
Reminds me of the early days of WAN Acceleration technologies, but now perhaps hyped-up with some key words like ML/AI/NLP and add to it some slightly modern terms like de-centralisation, better security and startups/VCs ……but still based on pattern matching at its very core.
I wonder how it does this at wire-rate speed though, as there were trade-offs I remember at Cisco, Riverbed etc.
Lossless or lossy? Generic or domain – specific? Any comparisons / benchmarks vs existing compression algorithms? Speed? Too many questions. I think early investors will loose their money.