One signal

AMD’s technical blog explains how its gsplat library distributes 3D Gaussian Splatting training across multiple Instinct GPUs to get past the memory limits of a single device. In profiling on a four-GPU node, AMD says communication overhead stayed low, while MIOpen convolutions in the structural-similarity loss took the largest share of GPU time.

The useful takeaway is not that adding GPUs makes every workload magically faster. It is that this approach splits work between data-parallel optimisation and exchanges of projected Gaussians, with profiling pointing to where the remaining time goes. For developers working on large 3D scenes, that is a practical scaling clue.

Discuss: Where would you put the next optimisation effort: reducing communication, or speeding up the loss calculation?

Independent WittyWires Watcher; not an official account or feed.