Premday #3 - Cloudflareʼs OpenBMC Journey

Cloudflare details how servers evolved since 2019 and how OpenBMC was used on them..

Transcript

1. Cloudflareʼs OpenBMC Journey PremDay 2026
2. Cloudflareʼs OpenBMC Journey - PremDay 2026 Generation 9 Edge Servers 2 x Intel Cascade Lake 24 Core In the beginning 2019 2019 100% Cloudflare servers on AMI BMC • Closed source, binary only • Beholden to server vendor for: ◦ Bug and vulnerability fixes ◦ Feature add / removal • Requests go "round trip": ◦ CF > Vendor > AMI > Vendor > CF ◦ Can take weeks or even months
3. Cloudflareʼs OpenBMC Journey - PremDay 2026 The Journey Starts 2020 2020 First OpenBMC development engineer hired • Engineers with OpenBMC experience are rare ◦ A smart embedded systems engineer can learn it ◦ TIP: search for resumes with the word “yoctoˮ • Pitfall: Proprietary Binary Libraries ◦ Used server vendor's binary library to speed development ◦ No access to source for that portion of the image ◦ "Don't control your own destiny" Same as AMI firmware) Generation 10 Servers AMD Rome 48-core
4. Cloudflareʼs OpenBMC Journey - PremDay 2026 Phase Two 2023 OpenBMC factory installed on all Gen 11 purchases •Limited resources meant limited coverage of legacy SKUs •Dev team expanded to three ◦ Ideal minimum team size to maintain tribal knowledge •Separate OpenBMC images for all SKUs ◦ 2 x edge servers & 4 x core server SKUs ◦ Specialty SKUs: ARM & GPU ◦ Added support for passively cooled GPUs Generation 11 Servers AMD Milan 64-core
5. Phase Three 2023 BMC chip abstracted to DCSCM • Implementation open sourced as Project Argus • Allows reuse of code between images • Enables use of lower cost non PCIe Gen5 PCB Generation 12 Servers AMD GenoaX 96-core Cloudflareʼs OpenBMC Journey - PremDay 2026 * Datacenter Secure Control Module
6. Phase Four 2026 Shared chassis supports multiple hardware use cases • Edge web • Edge storage R2 ◦ OpenBMC support added for additional U.2 drive bays • Edge GPU ◦ OpenBMC customized for additional cooling for high wattage GPUs • Core Kubernetes Generation 13 Servers AMD Turin 192-core Cloudflareʼs OpenBMC Journey - PremDay 2026
7. Phase Five Today and the future) OpenBMC team has fully integrated AI coding tools • Automated unit testing • Support for new SKUs accelerated ◦ ARM image took days vs. weeks • Accelerate porting older images to newest OpenBMC version • AI tools are a force multiplier - they donʼt replace people • Unified firmware image between Gen 12 and Gen 13 AIDriven Development Efficiency & Reliability Cloudflareʼs OpenBMC Journey - PremDay 2026
8. Thank you!