Announced this week at Mobile World Congress (MWC) by the Linux Foundation was the establishing of the OCUDU Ecosystem Foundation for advancing open-source AI-RAN (Radio Access Network) innovations. OCUDU is building a reference platform and innovations around 5G and early 6G network solutions. With OCUDU being benchmark-friendly, I have been putting the early code through some performance tests on current AMD EPYC and Intel Xeon server platforms.
The OCUDU software is an evolution of the srsRAN Project by Software Radio Systems (SRS) that was focused on open-source 5G solutions. OCUDU is focused on an open-source stack for CU (Centralized Unit) and DU (Distributed Unit) as part of Open RAN. Development on the Open Centralized Unit Distributed Unit (OCUDU) radio access network software began last year with funding from the US Department of War National Spectrum Consortium (DoW NSC) by DeepSig and SRS. OCUDU premier members also include SoftBank, NVIDIA, AMD, Nokia, Ericsson, Nokia, AT&T, and Verizon. With this level of industry backing, OCUDU will likely play a big role in the 5G/6G space moving forward. At the general member level are also other backers like Cisco, Marvell, T-Mobile, and Red Hat.
Those wanting to learn more about OCUDU and its new foundation as a Linux Foundation project can do so via OCUDU.org. For my purposes I’m interested in it as a benchmark given that the srsRAN predecessor has worked out well for CPU benchmarking too.
For my testing I benchmarked the current state of the OCUDU codebase on a dual 128-core Intel Xeon 6980P server for that flagship Granite Rapids server up against dual 128-core AMD EPYC 9755 “Turin” Zen 5 server flagship plus the dual 192-core AMD EPYC 9965 “Turin” Zen 5C too. Due to only having the Xeon 6980P processors of Granite Rapids currently with no other Xeon 6P review samples from Intel, it’s just a top-of-stack comparison for this initial OCUDU benchmarking.
The AMD EPYC Turin and Intel Granite Rapids servers were running in their maxed out memory configuration, including MRDIMM-8800 modules for the Xeon 6980P processors. Ubuntu 25.10 with the Linux 6.18 kernel was used for testing and the same PCIe Gen 5 NVMe SSD storage used by both the AMD and Intel server platforms. From there it was running the CPU benchmarks of OCUDU recommended by its developers. Looked at for this initial benchmarking was the per-thread performance of OCUDU. There is a total throughput benchmark as well with a hard-coded limit of 256 threads. When manually adjusting that limit, some other issues were encountered that are yet to be resolved and thus focusing on the per-thread performance for this first round of OCUDU benchmarking.
With OCUDU’s PDSCH (Physical Data Shared Channel) benchmark, the per-thread performance of the Intel Xeon 6980P was matching that of the dense Zen 5C cores in the EPYC 9965 2P processor. Though on a total throughput basis, the 192-core EPYC 9965 with 1.5x the cores/threads of the 128-core Xeon 6980P would easily win. With the EPYC 9755 processor using the “full fat” Zen 5 cores, its performance for the OCUDU PDSCH benchmark was 11% faster than the Intel Granite Rapids single thread performance — even with the latter being backed by the MRDIMM-8800 memory too.
With OCUDU’s PUSCH (Physical Uplink Shared Channel) benchmark, the results were even more compelling for AMD’s current generation processors. The dense Zen 5C cores in the EPYC 9965 were 4% faster on a per-thread basis while the full Zen 5 cores in the EPYC 9755 was around 20% faster than that top-end Intel Xeon 6 Granite Rapids server processor.
Benchmarks on other (random collection) assortment of CPUs with the new OCUDU can also be found via this test page, including the throughput benchmarks on sub-256 thread CPUs. More performance tests to come with OCUDU on Phoronix.
OCUDU is working toward its first official release in April and these initial numbers are very strong for AMD’s current EPYC 9005 line-up. Last week AMD also announced their EPYC 8005 “Sorano” processors that will be catering to Telco/RAN and thus a great fit for OCUDU / Open RAN software. Given these per-thread results, AMD’s Sorano should be very well positioned for when these processors begin shipping. AMD EPYC Venice based on Zen 6 is also not too far out for pushing AMD’s server performance capabilities even further.
If you enjoyed this article consider joining Phoronix Premium to view this site ad-free, multi-page articles on a single page, and other benefits. PayPal or Stripe tips are also graciously accepted. Thanks for your support.
