There is a lot of discussion about the need to increase chipset support for vRAN workloads, with one critical element being L1 acceleration for the DU. This is also a critical element of speeding Open RAN adoption, as Open RAN systems will need to achieve at least feature and performance parity with the integrated solutions, and the radio processing capabilities of the chip platform will be critical for that.
Just yesterday Vodafone announced that it is working with Marvell in combination with Samsung and Nokia to create Open RAN units with Marvell’s in-line L1 acceleration.
Intel’s lookaside acceleration, which supports its FlexRAN platform by acting as an external accelerator to its CPU, has been criticised for being less efficient than other solutions. Qualcomm is working on a range of SoC solutions that integrate inline accleration for Du-RU use cases. Vodafone, in its releases, said it wanted to work with Marvell, and vendors Samsung and Nokia, to “significantly improve” the performance and power efficiency of vRAN networks that run on standard COTS servers. (It sees such a solution as working in alliance with x86 platforms carrying out higher layer processing.)
But Rakuten Mobile’s CEO Tareq Amin is not for turning. Rakuten will announce in November that it is committed to continuing with Intel’s chips for L1 processing in its next phase of DUs.
But first, Amin acknowledges the limitations of Rakuten’s existing DU, which bolts together hardware acceleration around Intel’s CPU.
“Honestly it [the DU] is a big accomplishment but I’m going to be truthful with you – this is not my dream. It is a beginning of something bigger. What I do not like about this design is that there are too many discrete components, too many auxiliary hardware or chips around the CPU. I am of the opinion that additive acceleration cards are a narrow patch.”
Rakuten Mobile’s DU: “A big accomplishment but not my dream.”
But he sees an upcoming solution from Intel as providing a more elegant solution. Intel has announced that it is working on a next gen chip that it regards as akin to inline acceleration, in that it integrates acceleration in the CPU.
In February 2022, Intel said its upcoming next-generation Intel Xeon Scalable processor, Sapphire Rapids, includes support for RAN-specific signal processing and will deliver up to “two times capacity gains” for vRAN and support advanced capabilities like high-cell density for 64T64R Massive MIMO.
In addition, Intel also highlighted the fact that the Sapphire Rapids family will be built out with new chips with integrated acceleration that are optimised for vRAN workloads.
Amin and Rakuten Mobile are on board.
“We and Intel have agreed that we’re going to do an analyst day, on the 17th of November, we’re going to invite the entire analyst to share the future roadmap of a CPU accelerated product that is on the roadmap,” he said.
So this is the Sapphire Rapids product? “Exactly so. Except this time, on the 17th of November, I am forcing Intel to publicly tell the world about this innovation.”
What Amin wants to see is a solution that builds acceleration as building blocks into the CPU.
”So what I envision is the CPUs must get reduced, not increased.
“The right thing is to get the likes of Intel, AMD, maybe Nvidia to start thinking about a totally new architecture in which the CPU and the acceleration blocks are totally integrated into the same die.”
“Yes today there are a lot more CPU cores needed to support a complex multiband environment. But the fact of the matter is you need like that many CPU cores because there’s still not enough hardware acceleration blocks. As the CPU architecture changes, the likes of Intel are super committed to our roadmap of increasing the amount of acceleration that they’re going to keep adding to the CPU block.”
One question, as Rakuten Mobile evolves its network, is whether it could move away from Intel and its FlexRan platform if it wanted to. Did it look at Marvell, for instance, the company that Vodafone thinks can solve its issues for higher end vRAN?
“So we looked at we looked at Marvell, Qualcomm, we looked nVidia. And the reason that we stuck the course so far with Intel is primarily driven by the software ecosystem that they have built and their commitment to the future. For us, we like their roadmap, we like where they’re heading, and we like some of the process manufacturing improvement to improve power efficiency on the future generation of CPU products.
“At the same time, I hope others are going to come in, you know, I hope Marvell will be successful., I hope that even the likes of AMD [post acquisition of Xilinx] will have an opportunity to really get into the space as well. But importantly for us, we don’t want to create custom silicon. It’s a nightmare for me, it just doesn’t work. If I’m going to chase custom silicon development, it will ruin the business model for symphony.”
To take Marvell, subject of yesterday’s announcements. “Someone in the industry needs to the be the champion to port L1-L3 software on top of Marvell. This is not trivial. So, for me, that’s an investment that needs to happen and someone needs to do that.” Again, you could ask if Samsung and Nokia going to be those champions, given yesterday’s news with Vodafone, and previous announcements from those companies.
“Any time that I see someone talking about Open Ran, and really endorsing it, I think it’s a good thing for our ecosystem. The question is, will someone step up to give a L1, as Intel has done for FlexRan? See I have a L1 software, but it also contains a lot of intellectual properties and IP, so it is very difficult for me to tell the world hey, I’m going to open source my L1. I think there needs to be a reference design that gets created for everyone. By the way, Qualcomm is good at that.
“That becomes a catalyst for the industry to start building their own IPs on top of the reference design, that is maybe not necessarily hardened but at least you have a starting point.
“So the question becomes, for Marvell, would they be willing to invest in software, the way that Intel is investing in software? Those are the things that that needs to be seen. But I have no doubt they’re going to build credible hardware. The question becomes about what is the software roadmap is going to look like?”
In terms of pragmatic industry adoption, Marvell would no doubt point to integrations with Dell for its Open RAN Accelerator Card, a vRAN Distributed Unit (DU) solution with Marvell’s ARM-based OCTEON Fusion platform that incorporates inline 5G Layer 1 acceleration.
As for its continuation with Intel, Amin acknowledges that the company is committed to the FlexRAN platform partly because of the cost and effort of moving away. Not that he sees that is something particular to Intel.
“Once you are on a platform, whether it’s Nokia or Ericsson or Samsung, or ZTE or Huawei, the process of port out from one platform to the other is actually not trivial. To stabilise your software code base to run on any new chip architecture is just a time consuming thing. It doesn’t mean it cannot be done, but every platform or vendor has to determine the ROI of doing such a thing. For Symphony at this stage we don’t see clear ROI by us moving away from our partner that we have selected to work with. They’ve been really good for us so I am not necessarily diverging path at this stage. It doesn’t mean that we’re not open to look in the future if a good reference design comes from Marvell or others.”
