Faster GPUs and Smaller Caches Raise the Bar for Storage

Dell benchmarks PowerScale storage for KV cache offloading on NVIDIA GB300 GPUs. Testing with Kimi-K2.7-Code and DeepSeek-V4-Pro shows up to 13.6x faster time to first token and 8.1x lower latency. The study argues that smaller model caches and faster accelerators increase storage traffic, making persistent storage critical for serving long-context inference workloads efficiently.