Yesterday saw news that greenfield US operator Dish is going to use Amazon’s cloud environment and infrastructure to run its network software, including parts of the radio network. Although Dish has also intimated there will still be a role for the Intel FlexRAN platform – (Intel’s vRAN reference platform that runs on COTS hardware based on Xeon Scaleable or Xeon D chips) – it has also said that parts of the DU would be sited on Amazon’s own Graviton servers.
The move bears out a strategy line from well-backed new chip developer EdgeQ, whose CEO Vinay Ravuri told The Mobile Network in a recent briefing that he thinks hyperscalers (Google Cloud, AWS, Microsoft Azure) will end up making their own boxes to support 5G radio networks that provide enterprises with very low latency cloud service connectivity.
EdgeQ, with a new chip architecture for the hardware platforms that will host vRAN software and edge compute use cases, wants to be in on that.
Ravuri said that such a move would mirror what happened with SDN and datacentres switches, where the GAA gang moved on from working with the likes of Dell and Cisco on servers and switches, to making their own. He likens the O-RAN infrastructure, with the RIC defined as an intelligent controller of “dumb, stateless” DU-RU base stations, to SDN’s Controller/forwarding path architecture.
That creates a need for dumb but high-performance DUs, with the RIC hosting the network intelligence. EdgeQ’s aim is to use its RISC-V based CPUs to provide a flexible, programmable platform for the software that is required for the specific application. Its website markets it as a “base station on a chip”, talking about enabling the whole space from the access point to the RIC, with a fully programmable L1 PHY, and support for O-RAN compliant RU-DU-CUs.
Talking of the Open RAN opportunity, Ravuri said, “Where it becomes interesting is the centralisation and cloudification of the RIC. If you go that route it’s what SDN was for switching. SDN was a centralised controller with open interfaces to the hardware, which was super brawny, but stateless and dumb. The silicon and lowest layer of the switching elements were very flexible and programmable but all the intelligence is removed and put in the SDN Controller. The RU-DU-CU architecture brings almost the same concept to cellular. The RU-DU are stateless and dumb but very brawny – they pack a lot of punch – but are also programmable.”
“Now if you think about what happened in SDN, who were the winners and losers? OpenFlow as a protocol didn’t go anywhere. Cisco didn’t get out of business but they did get hit. And Broadcom was the big winner for commodity silicon. But the biggest inventors and innovators of SDN were the hyperscalers. They showed how SDN could work and then over time other vendors fit themselves into that world with YANG information models, and NetConf programming interfaces and so on. And that’s how Facebook and Amazon did their switches.”
“I can see much of this happening here [in wireless networks] as well, where hyperscalers will go down this path and do their own. Then the whole landscape changes.”
Ravuri said that Nokia’s three-way partnership announcement with AWS, Google and Microsoft showed the way things are headed, just as Microsoft’s acquisition of core network companies had.
“One of the things these companies want to do, if you look at today where the big push is, it’s enterprise. That has started, but the new thing is going to be the data produced by enterprise itself, a factory or a campus network. And a lot of that intelligence is not inside any cloud. As things get more automated, there’s data to be monetised. That data has to get into their cloud – and how does it get there? They are seeing operators as a bottleneck here. Sometimes they work together, but sometimes they’re the bottleneck they don’t move fast enough.
“And then latency is another piece. These applications are extremely latency sensitive. You can’t have real time control unless it’s very fast. So we are starting to see hyperscalers buy assets and get more into the cellular space, because they want to be at the edge themselves. This is how I see these people getting into the edge computing space, really via 5G.
“Then, given that the networks are getting upgraded into 5G they ask themselves ‘Why not be one of those people. Why can’t I have a DU-as-a-Service and offer that up to the end enterprises?’
If you look at Microsoft they got into core network by acquisitions, and the next step is to partner with operators and then say to customers, “Go into my core instead.” That’s happening as we speak today. But really they are looking at Private Networks that are yet to be built. Those are faster to evolve and more likely to be open to hooking up to a hyperscaler because they would want that data to go to him instead of an operator-based cloud. And then the next target is to either partner with box makers, or chip makers, and do this similar thing that you saw in SDN, where you build out your own network. They have the wherewithal and the funds to do it. And that is my guess on where this world is heading.”
Ravuri said that EdgeQ – which stuffed with ex-Qualcomm talent (plus ex-intel and ex-BroadCom) – is basing its architecture on a flexible, programmable chip platform that can be purposed via software.
“We are making both hardware and software, one is the chip and then the software in the cellular stack – whether that runs on or outside of our chip. We’ve chosen to go with RISC-V in the hardware layer. The RISC-V piece is CPU technology that we have modified so that the cellular baseband is more power efficient and area-efficient.”
The idea is that the same hardware that does all wireless things for you based on software.
Ravuri said that the company went with RISC-V, “and not Arm, for example,” because cellular comes with a lot of custom instructions – “think of things like matrix manipulations.”
“You’ve heard that things like machine learning don’t work well on CPUs, but instead they have custom chips or GPUs to do that. This is no different. If you take wireless algorithms/software and try to run it on a CPU it will be over x10 worse in terms of performance. Now RISC-V fundamentally in itself is no different, it’s a CPU technology, but I get to modify, gut and add anything I want in it, loyalty and license-free. That’s not possible with ARM or with other processors. That’s why we chose this. And we went this way because it has a benefit of the open source ecosystem that I can leverage.”
“Because they are RISC-V CPUs, this whole thing is super-soft, now. So just the firmware makes things one or the other. So it could be a DU or a RIC, the same chip, same programming model, but different product and firmware. Today this world is so new that you make chips for each of these things separately. The idea is that the same hardware does all wireless things for you based on software.”
“And interestingly, if you take Microsoft, Google, Amazon announcing that partnership with Nokia, I look as that as step one, as in the older days when they bought servers from Dell. And then they made their own servers. They bought switches from Cisco and then decided to make their own boxes. I can see much of this happening here as well, where they will go down this path and do their own.”
Of course, there’s one potential flaw in this vision. You still need the spectrum to get the data from your robots, automated vehicles and all the rest into your edge cloud. That might well require an operator holding the requisite spectrum license to operate. But Ravuri is confident that either hyperscalers can partner with mobile operators on this low margin piece of the business, or they can take advantage of license free and shared spectrum, or of privately licensed low-power spectrum, such as the 100MHz set aside in Germany’s midband for industrial and enteprise use.