I’m Kevin, come from RustFS team. We are building an AI-native storage infrastructure (https://github.com/rustfs/rustfs) that recently surpassed 26,000 GitHub stars. Our goal is to provide a high-performance, Rust-based alternative to MinIO, optimized for the “silicon-based life” era.
We have recently joined the NVIDIA Inception program and are eager to leverage BlueField DPU’s hardware acceleration. Specifically, we want to offload Erasure Coding (EC) and RDMA operations via DOCA to achieve maximum throughput.
I have two specific questions for the community:
Rust Integration: As our project is entirely written in Rust, are there any official or community-recommended Rust SDKs or FFI bindings for DOCA (specifically for DOCA Storage and EC)? What is the recommended path for Rust-based DPU development?
Development Environment: To support BlueField-4 (BF4) adaptation, should we focus on procuring physical hardware, or does NVIDIA provide a cloud-based lab environment (such as NVIDIA LaunchPad) where we can test our code on BF4/BF3 DPUs remotely?
There’s currently no official Rust SDK or native Rust bindings for DOCA (including DOCA Erasure Coding or DOCA Storage). The supported path is to call the C DOCA APIs from Rust via FFI (e.g., bindgen or a small C shim), following the DOCA C docs and samples for EC/RDMA.
2. BF3/BF4 development environment
For early experimentation, you can use NVIDIA LaunchPad, which provides free remote access to preconfigured systems with BlueField and DOCA. For serious performance work and BF4-specific tuning, you should still plan to get physical BlueField hardware; since you’re in NVIDIA Inception, coordinating hardware/BF4 access with your Inception contact is recommended.
Thanks for the quick reply—this is super helpful. Got it on the C FFI approach; we’ll start our early experimentation via LaunchPad with bindgen and coordinate with our Inception contact for the physical hardware.
Just out of curiosity, is a native Rust DOCA SDK on the NVIDIA roadmap at all? We’d love to see official support eventually. Since we are pushing for strict memory safety and zero-overhead performance, a native SDK would be a huge win for cross-platform compatibility and smoother toolchain integration.
Also, looking a bit further down the road: once we nail the RDMA and EC offloading on the current BlueField DPUs, how should we be thinking about the new Vera architecture? We want to make sure the architectural decisions we make today will cleanly align with where NVIDIA is heading next.
If there are any early-access roadmaps, architectural guidelines, or best practices you can share regarding Vera, that would be amazing.