ARCHITECTURE
Every memory tier has a role.
Strata keeps the experts a token is most likely to need in VRAM, the rest in system RAM, and large lookup tables on NVMe. The agent tunes the split for each PC: how much RAM to pin, how much expert work to send over PCIe versus compute on the CPU, and how many CPU workers to use.
| Tier | Role |
|---|---|
| VRAM | Dense weights, hot experts, KV cache window |
| RAM | Most experts, pinned for fast transfer |
| NVMe | Lookup table read during long prompts |
Capacity is not speed: a GPU with less VRAM leans more on the CPU and PCIe, which is why every node is benchmarked rather than trusted from its specs.