2.4.4. Neutron v7 Architecture#

The Neutron v7 architecture expands into an NPU system in which multiple NPU Cores work together. Unlike previous versions, which focus on vNPU partitioning within a single Core, Neutron v7 emphasizes multi-Core parallelism, shared on-chip memory, task-dependency synchronization, and coordinated scheduling of computation and data movement.

The compiler divides a model into executable tasks that are mapped to one or more NPU Cores. Each NPU Core contains compute EUs, data-movement EUs, Local OCM, and a SyncManager. Multiple Cores can exchange data and coordinate tasks through Shared OCM and SyncManager.

  • Multi-Core parallelism: Model subgraphs or tasks can be scheduled across multiple NPU Cores to improve parallelism and throughput.

  • On-chip memory reuse: Local OCM enables data reuse within a Core, while Shared OCM exchanges data across Cores and reduces DDR access pressure.

On the AX8860 platform, the --npu_mode option of pulsar2 build specifies the number of NPU Cores used to compile the model. NPU1, NPU2, and NPU4 correspond to 1, 2, and 4 NPU Cores respectively; they do not identify particular Core numbers.

2.4.4.1. Runtime modes and npu_mode#

2.4.4.1.1. AXEngine runtime modes#

Chip platform

Compute resources

AX8860

To be updated.

2.4.4.1.2. Offline npu_mode and hardware resources#

Chip platform

npu_mode

Hardware compute resources

AX8860

NPU1

One NPU Core. Each Core contains 4 CONV, 4 TENG, 2 SDMA, and 4 DAU.

AX8860

NPU2

Two NPU Cores. Each Core contains 4 CONV, 4 TENG, 2 SDMA, and 4 DAU, with cross-Core Shared OCM support.

AX8860

NPU4

Four NPU Cores. Each Core contains 4 CONV, 4 TENG, 2 SDMA, and 4 DAU, with cross-Core Shared OCM support.