Want to run 5G virtualized RAN on commercial off-the-shelf (COTS) servers? You might find that adding additional accelerator cards will better the performance and lower the power consumption of vRAN or Open RAN deployments.
Using silicon to accelerate the physical layer (Layer 1) of a cellular network is something that chip designers have alighted on as a way to improve the performance of commercial off-the-shelf (COTS) servers in the cloud-native scenario. That's because general purpose processors have proved to offer worse performance and power consumption than specialist silicon.
Intel has taken the early COTS lead with its “lookaside” x86 acceleration, which has initially been based on field programmable gate array (FPGA) cards and its FlexRAN software. The company is now starting to deliver its 4th generation Xeon server chips that include an integrated vRAN boost inside.
Principal and founder of analyst firm AvidThink Roy Chua said of vRAN deployments with silicon enhancements that – surprise! – “there are likely more FPGA-assisted, Intel CPU deployments with FlexRAN out there today" than any other option.
“For example, Dish and Rakuten Mobile both support Intel's FlexRAN architecture,” he said. Intel has also said that Verizon has deployed more than 10,000 cell sites using its 3rd generation Xeon processor.
Lookaside vs. inline
Lookaside acceleration offloads some but not all of compute intensive 5G processing from the x86 processor. Intel is also fine-tuning the power consumption of its server chips with its latest Xeon release.
“We believe that Sapphire Rapids will match or be better than any L1 [system on a chip] SOC solution that exists or could be entering the market,” Cristina Rodriguez, VP of Intel’s network and edge group, told Silverlinings recently.
There are plenty of “inline” acceleration cards, however, entering the 5G acceleration fray. These cards can offload the entire layer 1 to the accelerator, which, proponents claim can lead to a less data-heavy interface between the CPU and the accelerator. The end result is better data processing performance and power consumption results than lookaside options.
Marvell, Qualcomm and Nokia are among the major backers of inline acceleration.
Chua noted that “inline L1 cards with custom silicon (Marvell/Dell, HPE/Qualcomm, Nokia's own) in an open/disaggregated RAN deployment...It's only a handful running in pilots or limited production. Some of the cards aren't generally available yet.”
“The inline accelerators are superior performance-wise to the lookaside solution,” Chua noted. “But the CPU-based vRAN boost will help Intel hold its own against them.”
So who will win the acceleration game? It’s still too early to say, Chua said.
“In speaking to the server manufacturers, and even some of the SmartNIC vendors, it's still unclear who the advantage will go to,” Chua said. “Some of the telcos are planning on using L1 accelerator cards from Marvell (incl with Dell tie-up) and Qualcomm (with HPE tie-up) in their open RAN deployments, while others are waiting to evaluate performance comparisons in their own labs (vs. Intel's touted results) between Xeon 4th Gen, before committing,” he concluded.
It seems, then, the acceleration battle has barely even begun for many mobile operators.