A Thread-Register Decoupled GPU Execution Model for Efficient Tensor Computation
Summary
The paper introduces FIBER, a thread-register decoupled GPU execution model aimed at enhancing tensor computation efficiency. It identifies fixed parallelism and coarse-grained scheduling as bottlenecks when interleaving diverse non-GEMM and GEMM operations, and demonstrates substantial speedups across GPUs (Ampere, Hopper, Blackwell) with end-to-end gains up to 2.25x and kernel-level gains up to 2.49x. The approach includes ISA, microarchitecture, and compiler extensions to support shared-register addressing and dynamic dataflow scheduling.