We report the expected speedup (combined forward and backward pass) and memory savings achieved by PulsarAttention compared to PyTorch’s standard attention implementation. By improving memory access efficiency and compute utilization, PulsarAttention enables faster model execution with significantly reduced VRAM usage.