The timing is perhaps the most notable thing. As confidence in Open RAN falters in some areas, and operators increasingly announced single vendor “Open RAN-ready” contracts, Japanese company Kyocera is making a big bet on the development of an AI-powered open vRAN platform. The bet includes developing an AI plus vRAN platform based on an Nividia GPU platform, and the formation of an O-RU integration alliance.
As we reported last year, Open RAN can look different from outside the European T1 operator landscape, where advances are cautious – and where the imminent outcome of Vodafone’s Spring 6 tender is still unknown. In Japan NTT DoCoMo, which has its own multi-vendor RAN, is pushing ahead with an ecosystem integration play called OREX SAI, and claims to have deals lining up in East and South East Asia. Rakuten Mobile, of course, is slowly adding subscribers to mobile services delivered over its own Open RAN network. NEC may have scaled back its direct efforts to make headway in the Open RAN space, but it is coming again with a new near real time RIC platform announced this week.
And there’s a new kid on the block too – AI RAN. Here the likes of SoftBank and KDDI, as well as T-Mobile USA, are investigating the use of compute platforms that host not just Open RAN as cloud or virtualised workloads, but use the compute power that accelerates process-heavy vRAN workloads to also carry out AI tasks. These tasks could be AI for the network (ie optimising coverage and capacity algorithms) or they could be external AI workloads carried out on “spare” compute capacity that is rented out by the telco.
Into this environment this week comes Kyocera, which said it will be developing a new 5G AI-RAN platform built on the Nvidia Grace-Hopper GPU-CPU platform – in this case the GH200 superchip. This is a CU-DU platform that will be able to provide dual mode (sub 6 GHx and mmWave) connectivity, as well as connect to RUs shared by operators in the MORAN acrchitecure. It will also be able to run AI on the base station platform.
The company is also developing O-RUs for both FR1 (sub 6 GHz) and FR2 (mmWave). There is evidence online of Kyocera having previously explored the Open RAN and O-RU market, but the company said that the AI-RAN product is a “new departure.”
“We are at the finished stage of developing the AI-RAN prototype, and we are now in the phase of developing the commercial product,” Nishikido Masamitsu, General Manager, Radio Network System Development Division, said. However the company did not want to say when a commercial product would be ready.
Yoshihisa Nakagawa, Department Manager, Strategy Planning Department, Kyocera Corporation said, “We are refining the product to get to the stage where operators can bring it in to their networks. Where and how and how much business will be the next step – but we are at a very early stage at the moment. These are internal discussions but are premature to announce.”
The company was also unwilling to be drawn on which operators it is in discussions with. “The answer is yes we are discussing with several operators but nothing is confirmed yet.”
So why is the company judging is now the time to throw resources at and Open RAN and AI-RAN approach? Really it is that it is attracted by the way that Cloud RAN economics will change, based on advanced processors. Describing a landscape from 2029/30 onwards it envisions the potential for those platforms to reduce eventual TCO for the Cloud RAN. Again, geogrpahy may well play a part. It’s easier to imaging this future in countries where fronthaul capacity to connect data centre platforms to multiple remote sites is more abundant.
As for those economics, one key discussion on AI-RAN has been that the compute platform is likely to b e much more expensive than competitor platforms, given that it involves buying the Nvidia GPU. So Kyocera, as a company that has developed an Nvidia-powered AI-RAN prototype should have some insight into the cost profile of designing an Nvidia-based solution.
Masamitsu-san outlined the broad arguments in favour of the AI RAN economic model, “We think the use of the superchip pays off. Looking at one server it is expensive, however due to strong processing ability of superchip this platform can cover a plurality of BTS. Eventually it is cost effective.
“There is another reason. A shared CU-DU server enables the server to work both for the RAN and the AI, so the server can be used for AI in the night time, which can be a revenue source for operators.”
The platform itself is designed to support network sharing via MORAN, in order to make TCO for the platform further attractive. And although the superchip approach is not necessary to enable MORAN, which is an architectural specification that sees operators radiate their own frequency through shared antennas, Masamitsu-san said a high performance GPU can have some advantages “because the processor ability is strong and that can contribute to the MORAN aspect.”
O-RU Alliance
As it announced its AI-RAN platform, the company also said it would be forming an Alliance with a group of O-RU manufacturers. The main aim here is to have a stable of pre-validated O-RU to DU-CU integrations, so give operators a disaggregated but pre-integrated solution. That Alliance will be rbber stamped with a signing ceremony during MWC.
Kyocera’s O-RU partners are mostly from Taiwan and South Korea, with one in India, and are:
- Alpha Networks Inc. (Taiwan)
- HFR, Inc. (Korea)
- Microelectronics Technology Inc. (Taiwan)
- SOLiD Inc. (Korea)
- VVDN Technologies Pvt. Ltd (India)
- WNC (Wistron NeWeb Corporation) (Taiwan)
Amit
Proud to be witness of this evolution.