We can now eliminate almost all physical data movement in neural networks by using 'virtual tensors' to track logic instead of moving bits.
April 15, 2026
Original Paper
VTC: DNN Compilation with Virtual Tensors for Data Movement Elimination
arXiv · 2604.09558
The Takeaway
Data movement is the single greatest bottleneck and energy drain in modern AI, often dwarfing the actual computation. VTC solves this by using index mapping to track data, physically moving it only when absolutely necessary. By eliminating unnecessary transfers, this framework drastically speeds up LLM inference and cuts power consumption. Before this, we accepted data movement as an unavoidable tax on performance. Now, developers can build significantly more efficient inference pipelines that treat data as a logical pointer rather than a physical payload.
From the abstract
With the widening gap between compute and memory operation latencies, data movement optimizations have become increasingly important for DNN compilation. Current optimizations such as layout transformations and operator fusion only target a subset of tensor operators and consequently miss important opportunities for reducing data movement in contemporary DNN workloads, including large language models.We introduce VTC, a novel tensor compilation framework that for the first time eliminates all un