Blog
Simulator Field Guide
A living index of architecture simulators — NoC, memory, accelerators, CIM, and mapping frameworks, with notes from my own exploration.
An Optimal Mapper Meets a Real GPU: Hand-Lowering AccelForge's Attention Mappings to Triton
AccelForge's FFM mapper provably finds optimal fused mappings — optimal, that is, inside its own cost model. We hand-lowered its attention mappings to audited Triton kernels on an RTX 4090 to find out what that optimality is worth on hardware: top-10 rankings invert, pairwise orderings survive, and the model's fusion incentive turns out to be the same gap as its ranking failure.