# Insights: Measured Answers to AI Inference Storage Questions

> Articles on KV cache tiering, TTFT, LMCache, NVMe-oF, accelerator enablement and acceptance testing. Each gives the answer first, then the measured data and its source.

URL: https://vuke.top/en/insights · Updated November 12, 2026

The questions buyers actually ask, one per article. Each gives the answer first, then the data and the sources.

- [Low GPU Utilization From Frequent Model Switching](https://vuke.top/en/insights/gpu-utilization-model-switching) (2026-11-12): Each model switch leaves GPUs waiting for weights to load. At 20 switches per hour, shorter load times raise effective compute utilization from 46.7% to 62.8%.
- [How to Verify a Vendor's LLM Inference Benchmarks](https://vuke.top/en/insights/how-to-verify-vendor-benchmarks) (2026-11-12): Five checks for vendor inference benchmarks: baseline, platform and topology, run-to-run deviation, third-party report, reproduction material. Plus common traps.
- [How to Accept an Inference Storage Project: 4 Gates](https://vuke.top/en/insights/inference-storage-acceptance-gates) (2026-11-12): Split acceptance of an inference storage project into four gates on a weekly schedule: delivery, single-node baseline, main joint-test gate and stability sign-off.
- [Slow Model Loading From NFS: Ascend 910B Results](https://vuke.top/en/insights/model-loading-nfs-bottleneck) (2026-11-12): On Huawei Ascend 910B × 8, moving model storage from NFS to an FX100 all-flash array: measured serving load times for DeepSeek-32B and 70B, and where they apply.
- [Local NVMe or an NVMe-oF Array for KV Cache?](https://vuke.top/en/insights/nvme-of-vs-local-nvme-for-kv-cache) (2026-11-12): Both avoid full recompute. Across three concurrency levels on a 480B model, an external FX100 array gives more throughput and lower TTFT than one local NVMe drive.
- [Recompute or Read Back? KV Cache on a 480B Model](https://vuke.top/en/insights/recompute-vs-read-back-480b) (2026-11-12): Full recompute versus reading KV cache back from FX100 on a 480B model at concurrency 16: time to first token, throughput, test conditions and how to reproduce.
- [How to Cut LLM Time to First Token Without More GPUs](https://vuke.top/en/insights/reduce-ttft-without-adding-gpus) (2026-11-12): With long contexts, high time to first token usually means evicted sessions are being recomputed. Three fixes that need no new GPUs, with results from a 480B model.
- [What Is KV Cache Tiering and When Does It Help?](https://vuke.top/en/insights/what-is-kv-cache-tiering) (2026-11-12): KV cache tiering keeps the KV cache that overflows GPU memory in host RAM and external flash, then reads it back instead of recomputing. Structure, gains and limits.
- [When Does KV Cache Tiering Not Help?](https://vuke.top/en/insights/when-kv-cache-tiering-does-not-help) (2026-11-12): The gain from KV cache tiering comes from cache hits. Short chats, low concurrency, never-repeated contexts and output-bound workloads benefit little. Here is why.