DirectX Shader Cache Sizing and NVMe SSD Queue Depth Saturation: Historical Paradigm Evolution and Next-Gen Architectural Projections

[Dean Cooper]

I’ve been benchmarking shader cache behavior on my current build and hitting some weird wall around 128MB of allocated cache space—beyond that point, I’m not seeing meaningful latency drops on the 4K random reads that actually matter during shader compilation stutters in competitive shooters. For anyone who’s dug into how shader decompression throughput interacts with queue depth saturation on NVMe drives, what’s your observed sweet spot? Are you tuning cache size to stay under a specific queue depth threshold, or letting the driver manage it? And does anyone have real-world data showing where low-queue-depth latency actually becomes the bottleneck versus just being noise in the measurements?

I’m also curious how people think this evolves with the newer hardware—specifically whether hardware-accelerated shader decompression is going to shift the whole conversation about cache sizing, or if we’re still going to be fighting the same fundamental limits. For the folks running multi-queue setups, have you noticed the legacy single-queue methods just falling apart under modern workloads, or is there still a viable tuning path? What’s your experience telling you is going obsolete in the next couple of generations, and where do you think the architecture’s actually heading?

"@dean_trigger • Gamer24h"

[Jake Walker]

Great topic. On the shader cache side, the practical sweet spot for most competitive setups sits around 10–15GB in the registry (LocalSystem\MSSCachingV2), and honestly past ~24GB you’re leaving money on the table with zero real-world benefit—modern engines like Unreal 5 and Frostbite pre-bake so much that the cache just becomes a big LRU that thrashes. What actually matters for 1% lows is consistency, not raw size: a 12GB cache that stays warm across your session beats a 24GB one that constantly evicts because you’re juggling 6+ games. One thing people overlook is that the cache lives on your system drive, so if it’s on a nearly-full or DRAM-cache-depleted QLC drive, you’ll see the opposite effect—more shader compilation stalls, not fewer. If you’re chasing frame-time consistency, I’d pin the cache to a Gen4 with good

"[Audio: Sennheiser HD 560S + FiiO K5 Pro] • [RX 7900 XTX]"

[Author]

Posts