VectorWare brings Rust portable SIMD to GPU hardware

VectorWare, which describes itself as the first GPU-native software company, announced that Rust's portable SIMD library (core::simd) can now execute on GPU hardware. The implementation maps Rust's generic Simd<T,N> type directly onto GPU warp lanes, exploiting the architectural similarity between CPU SIMD and NVIDIA's SIMT (Single Instruction, Multiple Thread) execution model. In SIMT, a warp issues one instruction across 32 lanes simultaneously — structurally identical to SIMD data parallelism. For example, a Simd<i16, 32> vector assigns one i16 element to each of the warp's 32 lanes; an addition compiles to a single warp instruction (add.s16 in PTX), mirroring the vpaddw instruction generated on x86-64. This builds on VectorWare's earlier work mapping Rust's std::thread to GPU warps. Because core::simd lives in core rather than std, it required no additional std-layer porting. The result completes a full parallelism hierarchy: on both CPU and GPU, threads (or warps) contain SIMD/warp lanes, and core::simd drives those lanes uniformly. VectorWare states the same source code that compiles to x86-64 SIMD on a laptop now lowers to GPU warp instructions without modification.