Skip to content
BLOKZ.dev

The KV Wall

KV cache size (GiB) vs context length for Llama 3.1 8B, 70B, and 405B on log-log axes, against A10G / H100 VRAM ceilings. Toggle BF16, INT4, or DeltaKV compression and drag the cursor — the panel shows whether the live KV state fits on a single GPU.

chart SVG Machine LearningInfrastructureLLMs Jun 21, 2026
⬢ loading artifact…
The KV Wall — drag or click to move the context cursor · arrow keys ← → to step through context lengths · compression tabs toggle BF16 / INT4 / DeltaKV curves · data as of · Meta Llama 3.1 (arXiv:2407.21783); DeltaKV (arXiv:2602.08005); ARKV (arXiv:2603.08727) ↗ open artifact ↗

View artifact source on GitHub ↗

Appears in

Type to search the archive.