INFRA Signal 142
FIBER architecture decouples GPU threads from private registers to achieve 2.25x speedup on Ampere
↗
Illustration only Photo by Vishnu Mohanan on Unsplash
The FIBER architecture extends the GPU SIMT model by decoupling execution instances from private register ownership, yielding significant end-to-end speedups for mixed-precision LLM serving across Ampere, Hopper, and Blackwell GPUs.
Current GPU architectures struggle with fixed parallelism and coarse-grained scheduling when running modern AI workloads that mix GEMM and non-GEMM operations. By allowing execution instances to access shared registers dynamically, FIBER removes operand supply redundancies and better utilizes Tensor Cores for large language models.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
FIBER introduces a "fiber" execution instance that carries minimal control state and accesses an SM's registers through a shared view instead of private ownership.
The architecture addresses bottlenecks in modern AI workloads caused by fixed parallelism and coarse-grained scheduling interleaving diverse non-GEMM operations with GEMM.
In a mixed-precision LLM serving scenario, FIBER achieves a 2.25x end-to-end speedup on Ampere, 1.8x on Hopper, and 2.09x on Blackwell.
THE CLUSTER