IT之家 · 智能时代T2·72 pts
- A single SSD packs an NPU and 64GB memory with 130GB/s bandwidth, running vLLM and llama.cpp locally
- Pushing compute to storage can cut data movement 10-100x on certain workloads and offload host CPUs
- The mixed-media design replaces dedicated cache drives, losing only 4% cache capacity per failed drive