DirectX Shader Cache Sizing and NVMe SSD Queue Depth Saturation: Advanced Implementation and Real-World Architecture

[Grace Morgan]

So here’s where I’m hitting a wall. On our competitive rig, we’re seeing stutter spikes that line up almost perfectly with shader cache thrashing during map transitions, and I’m not convinced it’s purely the cache size itself. We’ve been tuning the DirectX shader cache between 4 and 8 GB, but when decompression throughput stalls, the CPU ends up idle waiting on 4K reads at queue depth 1 off our NVMe. Has anyone actually benchmarked the latency cliff on their flagship drive at QD1 versus QD4 for these small random reads, or is the controller just going to choke regardless?

Mais importantly, on the engine side, how are you handling shader compilation on the main thread versus offloading to a worker pool? We tried streaming compiled shaders ahead of known transition points, but we keep getting pipeline bubbles when the decompression can’t keep up with the renderer’s draw calls. Do you pre-warm the cache during loading screens, or does that just push the stalls into a different part of the frame? And has anyone measured whether forcing the cache to an NVMe versus a SATA drive actually eliminates the low-queue-depth stalls, or is DRAM caching the only real fix?

"i9-14900K | RX 7900 XTX | Custom Water Loop (EKWB)"

[Author]

Posts