DirectX Shader Cache Sizing and NVMe SSD Queue Depth Saturation: Competitive Benchmark Analysis and Hårdvara Interactions

[Carter Peterson]

I’ve been chasing frametime spikes during competitive matches and I’m increasingly suspecting my shader cache configuration and NVMe read behavior are colliding. When a scene swaps to rarely-used shaders, the decompression pipeline stalls the render thread and I see 16ms hitches that wreck my flick shots. Before I start resizing the DirectX shader cache across different drive capacities, I want real-world data: what cache size range have you found actually reduces stutter without eating into usable storage, and does a larger cache help in fast-shifting titles like Valorant or CS2, or does it just shift the stutter later?

Mer importantly, I’m digging into whether my NVMe is saturating during those low-queue-depth random 4K reads. The datasheets talk about queue depth 1 latency, but real-world shader streaming might behave differently. At what point do you start seeing queue depth saturation choke the decompression throughput, and are you tuning the driver’s NVMe settings or leaving them stock? Has anyone benchmarked frametime consistency with different queue depths and seen a measurable scaling pattern?

ITX Battlestation: Ryzen 9 7950X3D | RTX 4070 Super | SF750 Platinum

[Avery Howard]

Good topic to dig into. The shader cache and NVMe queue depth interaction is something most people overlook because they’re measuring the wrong thing. Here’s the deal: the DX11/DX12 shader cache in Windows stores precompiled shaders so games don’t recompile them on every launch, and the default cache size (often capped around a few GB) forces the OS to evict older entries under heavy use. For competitive titles like Valorant, CS2, or Apex, shader compilation stutters are exactly what ruins your frame times in the middle of a fight. I bump the shader cache up to 15-30GB via the registry key HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Control\GraphicsDrivers\Config\ShaderCacheSizeInMB (or use a tool like the “ShaderCacheSize” tweak) and clear it out periodically. A bigger cache means fewer on-the-fly compilations, which is a much bigger win than chasing raw sequential read speeds.

Now the NVMe side, which is where the real interaction lives. Shader cache reads are tiny, highly random I/O operations, and what actually matters here is queue depth at low depths (QD1-QD4), not the advertised sequential numbers or the 64K max

"⌨️ Custom Tofu60 Redux + Lubed Gateron Oil Kings + PBT Keycaps (Tape & Force Break Mod)"

[Blake Howard]

On the shader cache side, the practical lever most people overlook is that you want it large enough to stay resident but not so large that it pollutes your drive with stale entries. On Windows 11, the MaxShaderCacheSize registry keys (set under HKLM\SOFTWARE\Microsoft\Windows\Dwm) let you cap this—setting it to something like 8-16GB for a dedicated game drive is the sweet spot

"~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~"

"Veni, vidi, vici." • Esport & PC gaming enthusiast"

[Author]

Posts