At MWC, the DU was everywhere, cropping up on operator, server manufacturer, chip company, software vendor and network equipment provider stands.
The DU (Distributed Unit) is the bit of a disaggregated Cloud RAN baseband that provides time-sensitive processing of the lower layers of the RAN “stack”. Although the move to Cloud RAN, and its sort-of cousin Open RAN, is some way off, there is a rush on at the moment to provide the most efficient way of doing lower layer processing at the DU. The issue is, if you take this L1 software away from tightly integrated and optimised RAN compute platforms and stick it on a server, you need to design it so that the most intensive processes are “accelerated” in some way – otherwise you’d need huge amounts of processors and power, which isn’t really an option. Hence the DU gets tied to decisions about how best to provide this acceleration.
Dell’s accelerator card options for DU integration to its PoweredgeXR8000 server. Dell’s own card with Marvell, Nokia’s SmartNIC (also with Marvell), the Nokia SmartNIC, and Nvidia’s A100X processing engine.
How to provide lower layer acceleration in a DU may seem like an “angels dancing on pinheads” debate, especially as Cloud RAN adoption is in nearly all cases still some years off. However, major vendors say that operators are now fully engaged with Cloud RAN discussions, and tying it into their own future strategies, as they work out what such platforms may enable them to do from a business point of view. But these discussions are also exposing some strategic differences, and choices, in the industry.
Nokia, as TMN first reported on the Sunday before MWC kicked off, is enabling “Cloud RAN” by launching a product it is calling the SmartNIC. This is a card that can slot into a server to support DU functionality in a Cloud RAN deployment, with acceleration for L1 processing provided by Nokia’s Reef Shark SoC, which is actually based on Marvell’s own chip.
Nokia’s SmartNIC under some perspex at MWC23
Nokia said that the card can slot into pretty much any commercial server, including those in public cloud data centres. It was showing integrations with Dell and HPE, as well as AWS at the event. It can also be run within any cloud environment – eg Red Hat or AWS and so on.
But the openness stops there. If you want Nokia cloud RAN, you need its DU (SmartNIC). You can’t take Nokia’s vRAN software and run it, for example, on Qualcomm’s DU accelerator platform, rather than on the Nokia SmartNIC.
Why? Adrian Hazon, VP Global Product Sales, Nokia, says that the key reason is that Nokia wants to provide feature parity between its Cloud and its integrated solutions, and that it wants to be able to support the same software release cycle across any environment.
How do we think as an industry that you can suddenly switch and put that on bare metal and it will just run the same? You know, it’s just not physically possible at this stage.”
Hazon said: “Having our own accelerator with our own system-on-chip means we can absolutely guarantee that performance, and that feature parity. We wouldn’t discourage anybody else for doing their own accelerator card. But that [feature parity] is certainly a consideration for today, which is why we couldn’t hang around and wait for AN Other to come up with an accelerator card – we have to do our own. I think that’s absolutely the right decision and, as I say, we are not ruling anything out at this stage, but that’s why we are having our own accelerator card.”
So, just to be clear. Nokia says, you can have Cloud RAN on any server or cloud environment, but it’s tying it into its own accelerators, in effect offering another layer of integration?
“It is,” says Hazon, “and there’s a big debate about that, of course, as there should quite rightly be in the industry. But if you actually analyse what it takes… it has taken 30 odd years to get to the level of performance, power consumption, features, parity, everything we have with purpose built… how do we think as an industry that you can suddenly switch and put that on bare metal and it will just run the same? Eventually, yes, that’s the right sort of aim and strategy. But at this stage, no, it doesn’t work that way. Because you’ll end up putting too many servers out there. You wouldn’t be able to fit them all in the room. You know, it’s just not physically possible at this stage.”
Our solution is completely disaggregated, while some of the other companies are integrating their L1 acceleration with their L1 software
However, there is opposition to this view, and it comes from what may be a surprising source – Ericsson. Mårten Lerner, Head of Cloud RAN at Ericsson, has this to say.
“I think the difference is that we want to become fully portable between different hardware platforms. If we look at the alternatives out on the market, and I think SmartNIC is one way of expressing it, there is also an acceleration component, which means that you connect again the software to the hardware in a way that we have initially taken a step to go away.
“So our solution is completely disaggregated, while some of the other companies are integrating their L1 acceleration with their L1 software. And I think this has not been fully understood in the industry about the portability of the solution, because if you then go with smartNIC or their acceleration, you’re still becoming proprietary to that L1, versus us being agnostic across the different hardware platforms that we support.”
In fact Ericsson laid out this vision at its Capital Markets Day in December 2022, explaining how it wants to create parallel paths between deploying its RAN software on its own RAN compute and on commercial hardware, using the same interfaces.
Ericsson’s Cloud RAN strategy – providing “optionality”. Open RAN fans will note the reference to “next-generation LLS” (Lower Layer Split).
Lerner also stated that Ericsson is not tied to the migration path of any one partner that might provide acceleration within the DU.
“That is also a bit of a misconception in the industry,” he said, “that we have connected L1 directly to Intel. And I think that is coming partly because Intel has its FlexRAN reference stack. [Intel’s FlexRAN is not considered easy to port away from, once a vRAN vendor is committed]. But when we started to build with Cloud RAN, we had this portability vision in mind. So we took the decision of not using the FlexRAN as that is connected to their accelerators, but instead building our own agnostic L1 stack. So we can go in between both the generations (RAN Compute and general purpose hardware) without making any kind of changes to the software.”
Ericsson + Intel on the HPE ProLiant DL110 Gen10 Plus Telco Server. Ericsson says its software is agnostic to any hardware or chip provider.
Other accelerators are available
So why retain flexibility between the hardware and lower layer software, and upper software layers? Does it make any difference? One company that argues that it does, is Qualcomm, which has been marketing its X100, an accelerator card for the DU, for a number of months and is now shipping to partners.
Gerardo Giaretta,VP & GM of 5G RAN Infrastructure, Qualcomm, says that how acceleration of L1 workloads is provided is a key issue. For example, it can relate directly to power efficiency and performance. At MWC it was showing a demonstration of the X100 operating at output power of just 16-18W – Giaretta admitted that “apple to apple comparisons are difficult” but notwithstanding, claimed that this is better not just than any other Open RAN DU, but better than integrated versions as well.
The Qualcomm X100 power consumption demo, showing 16-18W. (Max power at full load is 35-40W)
Qualcomm’s partner announcements at MWC included Dell, Mavenir, NEC and Viettel, and its QDU was visible on a number of those partner booths during MWC.
Clearly, operators too are making their assessments. Dr. Femi Adeyemi, Head of Wireless, Fujitsu said that NTT DoCoMo has been deploying its CU-DU in Japan since 2019 for 5G, and will now move to a virtualised version.
“Now we have containerised the software, and Docomo will start deploying that image, so it will be vRAN ORAN. The same solution will be introduced into the North America market as well,” he said. Fujitsu announced that is vCU-vDU solutions will be based on NVIDIA’s GPU technology, applying NVIDIA’s GPU processing engine “NVIDIA A100X” to the physical layer processing at the base station. It was developed as part of the “5G Open RAN Ecosystem” (OREC) project promoted by NTT DOCOMO.
At the Vodafone booth you could see its current set-up, with a Samsung DU on a current generation Dell server in a Wind River cloud environment. But you could also see a desk presenting the work it is doing in its Mallaga chip research centre, which showcased Nokia’s SmartNIC, Marvell, Nvidia. EdgeQ and Intel solutions.
An array of server and chip options at Vodafone’s booth at MWC.
That last one is interesting, because Vodafone’s Open RAN lead, Santiago Tenorio, has made some pretty public comments doubting that Intel’s path is the correct one, leading the operator to investigate Marvell-based options with Samsung and with Nokia. But Intel is attempting to protect its position in the DU accelerator space by introducing vRAN Boost – an architecture that provides dedicated on-die acceleration for L1. You can read much more about that here, but in essence, its 4th Generation XEON Scaleable processor is designed to support on-die acceleration in the CPU. Intel says that will double performance and markedly improve power efficiency. Like others, it has racked up some partners for this approach, including Samsung.
Not everyone is impressed, however. Qualcomm’s Giaretta said that even if the approach is more efficient than Intel’s previous lookaside architecture, it doesn’t get around the issue that Intel is sending CPUs to do a DSP’s job.
“What Intel has announced, in my understanding,” he said, “is that they took FEC (Forward Error Correction) IP that they were doing for the lookaside accelerator and then integrated it into the CPU. So in that sense for sure the solution is more efficient than their previous solution because you don’t need a separate card. But when you look at it from a computing architecture perspective, and where the layer one workloads are implemented, they are still implemented on the CPU. And FEC is only a very small part of the L1 of 5G, there are a lot of other things – modulation, demodulation, equalisation – that you do on the DU that are very low latency.
“In 5G, you really have to do processing on a tenths of a microsecond perspective. CPU is not good on that and in order to make it good enough for that you have to add a lot of cores. You parallelise a lot.”
Qualcomm’s architecture that has inline acceleration uses DSP for 5G processing and so “those are naturally very low latency”.
Giaretta: “I can see vRAN Boost is better than their [Intel’s] previous generation. We still see that with our solution we can pretty much half the number of cores that are needed on a CPU, because the L1 is not needed on the CPU any more, and that is not magic, it is computing architecture.”
