01 Network Optimization
network_optGraph-level optimization: operator fusion and equivalent rewrites that fit the NPU’s constraints.
Graph levelMobilint designs the NPU architecture, compiler, quantizer and runtime together, so one architecture runs CNNs, LLMs and multimodal models efficiently and scales from a single core to a full server.
CNNs, LLMs and multimodal models use different compute patterns. One design rarely runs them all efficiently.
GELU and SiLU are expensive to build directly in hardware.
Lower precision saves memory but costs accuracy. Higher precision costs power and area.
Hardware and compiler are designed together, so one chip runs CNNs, LLMs and multimodal models efficiently.
GELU and SiLU run on lookup tables at minimal cost, and complex functions stay accurate.
Analyzes data distributions and sets the right precision for each layer automatically, cutting memory traffic to up to 75% without losing accuracy.
Throughput scales almost linearly from one to eight cores, and one programming model covers every level.
Each cluster pairs three specialized engines with a shared scratchpad memory (SPM). Data moves between engines on-chip, so off-chip memory access stays to a minimum.
Shared data within the cluster → minimal off-chip access
Matrix multiplication for CNN and LLM workloads
DMA, data movement and buffer management
LUT-based activations and vector operations
Performance rises almost linearly as cores are added, with higher throughput than competing designs in the same silicon area.
The same code runs at full performance from compact devices to servers, with no recompiling.
Mobilint’s mixed-precision engine sets precision layer by layer automatically, preserving model quality.
The qb Compiler sets precision in three levels.
Supported weights: 1, 2, 3, 4 and 8-bit. Activations: 8 and 16-bit. Values are restored to full precision inside the chip as they’re read, so smaller weights mean less memory traffic.
Each layer gets its own precision.
Precision is set per group of values within a layer.
Outliers stay in high precision while the rest compress. Unevenly spread values use lookup-table quantization.
Up to 75% less memory traffic.
Over 99% of FP16 accuracy retained, with no retraining.
80 TOPS (Dense, INT8)
For edge servers, robots and industrial systems. Available as PCIe cards, MXM and M.2 modules, edge AI boxes and AI servers.
10 TOPS (Dense, INT8)
For cameras, drones and compact devices. Available as a chip, SoM, evaluation kit, single board computers, USB dongle and mini AI box.
