Having already teamed up with Intel and Nvidia, AI-RAN pioneer DeepSig is integral to the OCUDU open-source initiative.
A pilot signal acts as a scout in the mobile network, surveying conditions and clearing traffic to gallop through the wireless channel. Using AI, DeepSig was apparently able to remove the pilot like a redundant employee and eliminate overhead without endangering performance. Founded in 2016, the US startup is at the frontier of inserting AI into a radio access network (RAN) as an efficient substitute for the conventional algorithms written and refined by people. As AI shoots to the top of the telecom agenda, DeepSig has been lifted out of relative obscurity, propelled by tie-ups with some big hitters.
DeepSig most recently showed up as a significant factor in OCUDU, the initiative unveiled by the Linux Foundation and the US Department of Defense (DoD) before this year’s MWC Barcelona. Standing for “open central unit distributed unit,” it seeks to inject open-source code into the baseband or computational side of the RAN. Today’s US baseband market is dominated by Ericsson, Nokia and Samsung with their proprietary technologies. The shift to an open-source foundation would theoretically help smaller players like DeepSig slot in next to those giants.
DeepSig, though, looks integral to OCUDU, named alongside Ireland’s Software Radio Systems (SRS) as one of the two startups chosen by US government authorities to build the initial software stack. “The National Spectrum Consortium had an RFQ for doing an open-source stack, and so we partnered with SRS,” said Jim Shea, DeepSig’s CEO. “They had their basic stack, which is very capable, but it needed bringing to carrier class, which involved adding a lot of new features and really hardening it.”
Ericsson and Nokia were immediately identified as “premier members” of a group called the OCUDU Ecosystem Foundation when the project was announced. In principle, they could adopt OCUDU and use it to integrate third-party algorithms into their own products. Yet hardly anyone in the industry expects either of the incumbents to wholeheartedly embrace open source and abandon proprietary software.
In Nokia’s case, that much seemed to be confirmed by CEO Justin Hotard’s remarks about OCUDU at the Finnish vendor’s MWC press conference this year. “We see this as a natural evolution of the ecosystem,” he said. “It is necessary to accelerate innovation cycles and what it means is you are no longer dependent on a vertically integrated stack to put the pieces in. But it doesn’t necessarily mean that everything is open source.”
That said, OCUDU is being championed by US government agencies including the powerful DoD, keen to open 5G and 6G to US applications with military uses. For Ericsson and Nokia, defense may currently seem like a more lucrative growth opportunity than mainstream 5G. Both Nordic vendors also count the US as one of their biggest and most profitable markets.
The Shea and O’Shea double act
DeepSig’s fortunes do not hinge entirely on OCUDU, either. The company was partly the brainchild of Tim O’Shea, its chief technology officer, while he was a research scientist who studied deep learning, a subset of AI, at Virginia Tech. “He realized you could apply this in a wireless area,” the similarly named Jim Shea told Light Reading at MWC, with O’Shea seated nearby. “You can apply the technology to improving the way communication systems operate by replacing many of the traditional algorithms with deep learning,” Shea said.
Still up for debate is what precise impact all this will have. For decades, Shannon’s law has seemed to put a hard limit on how much data can be shoved into a given amount of bandwidth. The experts involved in refining conventional algorithms say today’s networks are already brushing up against it. Shea acknowledges that DeepSig is not breaking any fundamental theoretical limits. And yet, if the economics make sense, even marginal improvements will look attractive to a resource-constrained sector.
DeepSig accordingly developed two products. The first, called OmniSIG, essentially uses AI to do radiofrequency-based sensing. That soon led to some revenue-generating work for DeepSig with defense contractors. OmniPHY, DeepSig’s other product, perhaps has more relevance for telecom as an AI-based contribution to the physical layer, or Layer 1, of the RAN – the baseband software responsible for the most demanding network functions.
Shea and O’Shea started out working with Intel through FlexRAN, the chipmaker’s own, sometimes maligned, baseband reference stack. “We went in and we replaced a lot of the upstream processing, signal processing, with AI and got improved performance,” said Shea.
By working with Intel and FlexRAN, DeepSig could address what once seemed to be the world’s largest virtual RAN platform, running off general-purpose central processing units (CPUs) and standard servers. Unfortunately, having not advanced as quickly as Intel would have hoped, virtual RAN still accounts for a relatively small percentage of the total RAN market. What’s more, smaller RAN vendors using FlexRAN have failed to capture market share. And the big suppliers that offer Intel-rooted virtual RAN products – Ericsson and Samsung – claim to have replaced FlexRAN with their own stacks.
OCUDU or CUDA?
While it has not ditched Intel, DeepSig’s priority these days appears to be a more recent partnership with Nvidia. About a year before OCUDU hit the headlines, the giant maker of graphics processing units (GPUs) said it had opened Aerial, its own baseband reference stack, to other developers. They would, naturally, have to work within the boundaries of CUDA, the software platform (often described as its “moat”) that Nvidia has built for its GPUs, just as FlexRAN developers were restricted to Intel’s CPUs. But Nvidia’s ecosystem was blossoming, while Intel’s looked gloomier.
Shea’s company has also been able to run some of its software on Nvidia’s GPUs without the need for Aerial. “We actually have a demonstration running here on a DGX Spark where we did not take Aerial,” he said, referring to one of Nvidia’s GPU products. “We just took the basic SRS stack and accelerated portions of the RAN with the GPU.”
The intention is to be as hardware-agnostic as possible. For that reason, Shea takes a dimmer view of the application-specific integrated circuits (ASICs) that still account for most of the baseband silicon in commercial use. The interdependency of hardware and software in those ASICs is a clear obstacle for companies such as DeepSig. But Shea is evidently in the camp that doubts custom silicon will have a long and lively existence.
“I think the problem with the ASIC approach is the cost of doing the production of these three-nanometer, two-nanometer, one-nanometer chips,” Shea said. “I just don’t know if the RAN market is big enough to support a specific ASIC for that. Over time, I think the Intel’s and the Nvidia’s that are doing this for other markets can really ride this technology curve.” While that view has opponents, with Ericsson among them, it is shared by some executives at Samsung. Nokia’s decision last year to develop RAN software for Nvidia’s GPUs, and the recent talk by Hotard of shifting from proprietary to general-purpose hardware, would suggest the Finnish company agrees.
The challenge is moving software from an Intel CPU to an Nvidia GPU, or vice versa. Their fundamentally different architectures would demand a rewrite of code, even if the core algorithms do not have to be changed. That could be awkward for smaller companies with limited resources. It is partly why Shea seems drawn to OCUDU as a platform that is independent of a big hardware sponsor and could simplify systems integration. “We think OCUDU is going to help address some of that,” he said. “We’re going to come up with basically one way of doing everything, and it’s going to be an open-source framework.”
Of course, Nvidia’s growing ubiquity makes people less worried about relying on its technology. Aerial will look extremely inviting if GPUs become widely adopted in the RAN. Nokia, in which Nvidia has invested $1 billion, even sounds willing to integrate third-party software designed for Aerial into its “telco-grade” implementation. Whether it comes through Aerial, OCUDU or something else, a partnership with one of the world’s biggest RAN vendors would put DeepSig firmly on the map.
https://www.lightreading.com/5g/small-deepsig-is-at-heart-of-ai-ran-challenge-to-ericsson-nokia
