A neural accelerator is a memory system with arithmetic attached. NP-1 is organized around keeping weights and activations on-die so the array never waits on a bus.
Each PE is self-contained: it holds its tile of weights, executes against them, and reduces locally before ever touching global SRAM.
Three tiers, three orders of magnitude of latency and bandwidth. The compiler's job is to keep traffic in the top two tiers.
| Tier | Capacity | Aggregate bandwidth | Latency | Purpose |
|---|---|---|---|---|
| Tile SRAM | 16 MB · 1 MB × 16 PE | ~12 TB/s | ~2 ns | Weights and activations local to one PE |
| Global SRAM | 80 MB · 16 banks | ~2 TB/s | ~8 ns | Cross-PE reduction, KV cache, staging |
| LPDDR5X | up to 64 GB | 256 GB/s | ~80 ns | Long context, multi-tenant working set |
HBM buys bandwidth at the cost of latency, thermals, and a co-packaged stack that fights the edge power budget. Edge inference is latency-bound, not bandwidth-bound at the margin. A large on-chip SRAM pool plus commodity LPDDR5X covers the workload envelope without the package cost.
The sixteen PEs are wired as a two-dimensional mesh. Neighbor-to-neighbor transfers are single-hop; global reductions are folded across rows then columns. The scheduler places dependent operators on adjacent PEs to keep hops minimal.
One MAC array, five precisions: INT4, INT8, FP8, FP16, and FP4. Conversion to and from higher precision happens in the quant/dequant block, not in software. Sparsity is 2:4 structured — two of every four weights are skipped with metadata, so the speedup is deterministic rather than data-dependent.