CTA-Pipelining: A Latency-Oriented Spatial Scaling Method for Multi-GPU Systems

Collection of systems/ML infrastructure research (arXiv 2026-07-10) focused on GPU/accelerator performance, programming models, and privacy-aware distributed execution. Key items: CTA-pipelining — a latency-oriented spatial scaling method for shared-memory multi-GPU systems that reduces MLP/GEMM latency vs. micro-batching and tensor-parallelism; a proposed GPU-aware OpenSHMEM auxiliary specification to standardize accelerator memory semantics and capabilities across vendors; a field study of deploying MoE and multimodal inference on a 16-device Huawei Ascend 910 cluster that required 12 source