FP8 weights at BF16 memory size, KV cache remains the VRAM bottleneck

AlpinDale · x · 2026-08-28

Discusses the performance trade-off when FP8 checkpoints occupy as much memory as BF16 on hardware supporting FP8 activations (e.g., sm89+). The view is that unless the model is very large, the KV cache remains the biggest VRAM consumer.

Original post →

More from Infra

Infra channel →