Fp4

Match each GPU generation to the quantization format that lights up its native tensor cores.

Pick the serving changes TensorRT-LLM 0.18 (early 2026) actually shipped

Match Hopper FP8 and Blackwell FP4 tensor cores to their throughput and bandwidth effects on decode

Match each 4-bit format to its niche in 2026 inference and fine-tuning.

FP4 vs NF4, pick the answer that captures the actual structural difference