# Apple NVMe small random read latency under F_NOCACHE at queue depth 16 > Existing benchmarks that skip F_NOCACHE can serve reads from the unified buffer cache, so results vary with installed RAM (author's claim; a forum reply disputes it on the basis that SATA and HDD speeds looked like device speeds). Parent: [Mac local LLMs: MoE streaming and offload](https://llms-explorer.com/tree/mac-local-llms-moe-streaming-and-offload/) · 1 facets · 8 facts · page: https://llms-explorer.com/tree/apple-nvme-small-random-read-latency-under-f-noc/ ## Facts - Existing benchmarks that skip F_NOCACHE can serve reads from the unified buffer cache, so results vary with installed RAM (author's claim; a forum reply disputes it on the basis that SATA and HDD speeds looked like device speeds). — [source](https://forums.macrumors.com/threads/beta-disktruth-f_nocache-mac-disk-benchmark-slc-cliff-detection-negotiated-interface-speed-testflight-now-open.2481094/) - DiskTruth author: BlackMagic and AmorphousDiskMark do not use F_NOCACHE and are inflated by RAM caching on large-memory Macs. A MacRumors reply (bzgnyc2/!!!) says those tools gave device-consistent speeds on SATA SSD (<500 MB/s) and HDD (~200 MB/s), so caching is not evident. Shown side by side, not resolved. — [source](https://forums.macrumors.com/threads/beta-disktruth-f_nocache-mac-disk-benchmark-slc-cliff-detection-negotiated-interface-speed-testflight-now-open.2481094/) - DiskTruth (beta announced 2026-04-16) sets fcntl(fd, F_NOCACHE, 1) on every file descriptor before any I/O. — [source](https://forums.macrumors.com/threads/beta-disktruth-f_nocache-mac-disk-benchmark-slc-cliff-detection-negotiated-interface-speed-testflight-now-open.2481094/) - DiskTruth measures random 4K IOPS and latency at QD1 with 10,000 operations per pass and reports average and p99. — [source](https://forums.macrumors.com/threads/beta-disktruth-f_nocache-mac-disk-benchmark-slc-cliff-detection-negotiated-interface-speed-testflight-now-open.2481094/) - DiskTruth measures queue-depth scaling at QD1, QD4, QD16 and QD32 with 3,000 ops per level. — [source](https://forums.macrumors.com/threads/beta-disktruth-f_nocache-mac-disk-benchmark-slc-cliff-detection-negotiated-interface-speed-testflight-now-open.2481094/) - Its author states Apple Silicon internal NVMe has "extraordinary QD1 latency" and less dramatic scaling at higher depths than desktop NVMe, with no numbers given in the post. — [source](https://forums.macrumors.com/threads/beta-disktruth-f_nocache-mac-disk-benchmark-slc-cliff-detection-negotiated-interface-speed-testflight-now-open.2481094/) - DiskTruth uses a 16 GB test file, 3 iterations and the median for sequential tests. — [source](https://forums.macrumors.com/threads/beta-disktruth-f_nocache-mac-disk-benchmark-slc-cliff-detection-negotiated-interface-speed-testflight-now-open.2481094/) - The ds4 QD16 stall estimate remains incomplete: no measured latency number exists in the sources read. — source: `asserted`