G REIGNS, a division of HTC that develops Private 5G vRAN and core solutions, has agreed to integrate software from DeepSig, a company that develops AI-native PHY software that can be used in RAN basebands.
HTC’s G REIGNS is producing a compact private 5G network-in-a-box solution, called REIGN CORE 2. It first saw the light of day at the 2022 Mobile World Congress, where your writer noticed it on the exhibition floor.
The companies said this week that they will work on updating this solution to include DeepSig’s OmniPHY 5G software in HTC’s solution, and also work on further “AI-driven vRAN collaborations”.
DeepSig was first covered by TMN in 2021. Its OmniPHY 5G is software that applies AI/ML in the upper-L1 baseband processing in a DU. Intel has integrated the software within its FlexRAN Layer 1 reference software. The software operates fully within the 5G Open Distributed Unit (O-DU) and replaces multiple 5G NR signal processing algorithms with a Deep Neural Network (DNN).
DeepSig said in a May 2022 white paper it produced with Intel, “This approach with DNN potentially requires less computation while significantly improving network capacity and resilience to interference by learning the real-world characteristics of the local wireless environment where the Radio Unit (RU) operates.”
The DeepSig PHY provides replacements to the PUSCH (Physical Uplink Shared Channel) channel estimation SDK routines for standard MIMO, and to SRS (Sounding Reference Signal) channel estimation and pre-coding routines for massive MIMO. Inclusion in Intel’s FlexRAN software means DU vendors can access the algorithms without making additional changes in their hardware or software stack.
Intel and DeepSig said in a Febuary 2023 press release it had delivered its first operational release of OmniPHY 5G compiled into Intel FlexRAN L1 reference software to select Open RAN ecosystem partners.
Adrian Tung, General Manager at HTC G REIGNS said, “We are committed to bringing advanced, O-RAN-compliant wireless technology to market, and look forward to partnering with DeepSig on optimising their AI software into our 5G private network offerings for low-latency use cases, IOT and smart city technologies.”
David Oberholzer – Vice President, Business Development, told TMN in February that as well as a Taiwanese customer (which seems likely to be HTC) DeepSig also has a US customer that will make an announcement in early 2023.
DeepSig says that its advantage is that it can apply machine learning techniques to PHY processing, making Layer 1 processing more efficient. “The software drops right into Layer One in the DU. So it’s not off board. It’s not on a card. It is actually in the DU. And it replaces static algorithms for the uplink processing – and uplink obviously is the most complex to compute in the link budget of the base station.”
In demos with Intel, Oberholzer said that the software had been able to improve both speed and accuracy in calculations like determining the best channel for a UE. The 2022 white paper said that operating on 3rd Gen Xeon scaleable hardware, OmniPHY could deliver up to 20% increases in PUSCH processing speed. Gains would be even better in 4th Gen (SaphhireRapids) based products.
“A lot of traditional methods will go really fast, but they won’t be as accurate,” Oberhozer said. “So you won’t have that maximum throughput available. So what our approach does, through machine learning, is improve both accuracy and speed at the same time in massive MIMO. So we’re taking a very complex process, running it faster with fewer cycles needed on the CPU, meaning we’re doing it with less compute load, which translates actually to power savings.”
The industry has recently seen Cohere Technologies make a pitch for more accurate and longer lasting MIMO channel predictions by running calculations based on Delay Doppler-based measurements. As a means of getting to market Cohere is exploring the deployment of its technology as an xApp in a near real time RIC, as well as deployment directly in the base station.
DeepSig sees, instead, the benefit of deploying its PHY software right in the DU, avoiding the latency budget of taking data to the RIC for processing.
“So I’m not saying what we do is exactly the same as them. But what we do doesn’t require that offload processing. We’re reducing the compute load, that’s one of the main benefits to reducing the processor load in the DU, and increasing capacity in the DU.”
“Especially in open ran, one of the earliest issues that arose from operators was addressing power consumption, or server efficiency. It’s a great expense in the R&D to go through this. Now they have a licensable software solution to apply to that.”
As vendors continue to discuss the benefits of lookaside and inline acceleration for Layer 1 PHY, DeepSig will continue to push for the benefits of its approach of an AI PHY to improve performance.
“AI-native is going to be the path that will lead us forward into 6G, and this [Intel] implementation is proof of it,” Oberholzer said.