Recovering 4.1 GiB of hidden RAM on NVIDIA DGX Spark, 3.6x bigger KV pool
marian_nmt · x · 2026-10-02
The author found Linux exposes less than the DGX Spark's 128 GB unified memory and recovered 4.1 GiB per unit by stacking two tricks: switching to a 64 KiB kernel (2.1 GiB back from reduced vmemmap metadata) and reclaiming the 2046 MiB headless display carveout. On a 2x setup, this grew the KV pool for GLM-5.3-Flash from 262k to 938k tokens. Full measurements and a runbook are in the linked blog post.
Related event: Two tricks recover 4.1 GiB of hidden memory on DGX Spark(2 posts)→
More from Infra
- US data center construction hits record $85B/yr, up 513% since ChatGPT launch — HaydnBelfield · 2026-10-02
- HBM shortage to persist for years, analyst pushes back on Wall Street's 2029 memory crash bet — JOBhakdi · 2026-10-02
- Hugging Face adds Train Models docs: fine-tune on one A10G for ~$0.10 in minutes — vanstriendaniel · 2026-10-02
- Gemini 4 may roll out to Ultra users next week as Google's TPU capacity reportedly runs tight — haider1 · 2026-10-02
- Banks tighten data center lending as Beff bets thermodynamic computing can fix AI power crunch — beffjezos · 2026-10-02
- Local Qwen 27B 4-bit proved its worth in a hospital with no signal or Wi-Fi — AaronBergman18 · 2026-10-02