Nvidia's two-track Rust support puts the safety payoff in the host code

A Rust CUDA build has exactly one handoff, and it is PTX text. The device module compiles for the nvptx64-nvidia-cuda target, build.rs writes kernel.ptx into OUT_DIR, and the host binary reads that file back with include_str! before passing it to Module::from_ptx and launching on a non-blocking stream ^[7][13]. The driver loads PTX, and by then the source language is gone. "Rust doesn't make your kernels magically faster. The GPU doesn't care what language emitted the PTX," the guide says ^[11].

The sample kernel shows what that leaves. The file is declared #![no_std] with #![feature(abi_ptx)], and the entry point is a pub unsafe extern "ptx-kernel" fn taking three raw pointers and an i32 ^[8]. The body derives its index from _thread_idx_x, _block_idx_x and _block_dim_x, checks idx < n by hand, then writes *c.add(i) = *a.add(i) + *b.add(i) ^[9]. That single statement dereferences three raw pointers inside an unsafe function ^[10]. The bounds check is written out, the same way it is in the C++ version printed beside it ^[19]. On where the win sits, the guide is straight: host-side code managing allocation, launches and transfers, where it puts use-after-free on device pointers and forgotten cudaFree calls ^[12].

The walkthrough itself runs on the community path. rust-cuda, reached through ptx-builder and the nvptx64-nvidia-cuda target, appears on the guide's list of unofficial tooling that first-party backing supersedes, alongside wgpu and cudarc ^[4]. The setup steps then add nvptx64-nvidia-cuda, and the host program is written against cust ^[7][13][20]. Nowhere does it cite Nvidia's announcement or name a crate, a release or a version ^[16].

For the official-support framing to change a build-versus-bind decision, the toolchain would have to show crates published by Nvidia, device codegen on stable instead of an unstable feature gate, and cuda-gdb and nsight resolving Rust frames. That last item is precisely what the guide says community crates lacked, along with guaranteed compatibility across CUDA toolkit releases ^[5]. Today the adoption cost is a pinned nightly in CI and nvcc on the build host ^[7][14].

In my view the two-track split is honest about who it serves. Track two is an iterator-style API compared to rayon, for people who do not want to size grids and blocks for every operation ^[3]. Track one gives direct control over thread blocks, shared memory, warps and memory coalescing, and the guide says it is aimed at engineers porting existing CUDA code ^[2][15]. For a team running an inference server, the part that changes is the host: swapping FFI glue for a first-party host API removes the bug class the guide names, and the kernel still needs the same review it needed in C++. The guide says Nvidia is targeting both the performance-obsessed kernel author and the application developer who wants GPU acceleration without becoming a CUDA architecture expert ^[17].