DSpark
Distributed execution & kernel compilation framework for DGX Spark clusters
Tools
In plain English
The coordination software that lets two or four DGX Sparks share a single giant AI model over high-speed cables as if they were one machine.
Technically
An optimized distributed pipeline and tensor parallelism execution runtime tailored for Grace Blackwell (GB10) clusters. Orchestrates memory-mapped model weights across RoCEv2 ConnectX links with fused cross-node GEMM kernels to minimize communication bubbles.