Space-Mate : SMoE-based NeRF-SLAM Processor
본문
Overview
Space-Mate is the first SLAM chip that produces a complete 3D map and accurate position in real-time, which is achieved with 3 key features: 1) Out-of-Order (OoO) SMoE router to alleviate data transactions for low latency, 2) single-skip (SS) and dual-skip (DS) heterogeneous core architecture to exploit coarse-grained sparsity caused by similar zero patterns in the same expert for high throughput and energy-efficiency, and 3) loss-prediction based 2D sampling unit for low energy per frame. Space-Mate shows 32.85 FPS with 303.5mW power consumption.
Implementation results
Performance Comparison
Architecture
Features
- Out-of-Order Sparse-Mixture-of-Experts (OoO-SMoE)
- Single-skip core (SSCore) and dual-skip core (DSCore)
- Loss-prediction based 2D sampling unit
Related Papers
- ISSCC 2024