128 TOPS, sparse acceleration.
High-efficiency NPU with 128 TOPS for edge and cloud AI inference.
Tile-based renderer, Vulkan 1.3, ray tracing.