Roofline analysis of BitNet b1.58 2B4T inference across H100, MI300X, Groq LPU, Cerebras WSE-3, and two hypothetical ternary chips. Five phases covering gate counts, precision sweeps, memory bandwidth, hybrid activation Pareto, and prefill regime, plus empirical weight-distribution validation against the published microsoft/bitnet weights.
quantization gpu-architecture bitnet roofline-model inference-optimization llm-inference ternary-computing hardware-modeling
-
Updated
May 10, 2026 - Python