Memoriant/dgx-spark-kv-cache-benchmark

KV cache quantization benchmarks on NVIDIA DGX Spark GB10 — three novel findings including the dequantization cliff and unified memory paradox

3 best points · 0 comments · 1 Hacker News post · 13 GitHub stars · View on GitHub

Hacker News discussions

• Q4_0 KV cache collapses 92.5% at 64K on DGX Spark GB10 (unified memory paradox) (3 points, 0 comments, by memoriant, 2026-03-31)