FP8 is the format modern quantised LLM inference increasingly targets, and INT2 suits aggressively compressed always-on models. Since Qualcomm publishes no throughput figure, supported precision is the one axis on which its generations can honestly be compared.
Snapdragon 8 Elite Gen 5
The real change is the precision set: INT2 and FP8 arrive, which matters more for modern quantised inference than the headline percentage.
Silicon from Qualcomm
9 chips on record, newest first.
-
2026Relative only
-
2026Relative only
-
2025
Snapdragon X2 Elite Extreme
Not stated by Qualcomm Hexagon NPU
The Qualcomm Hexagon NPU features 80 TOPS of AI processing
Absolute -
2025
Snapdragon 8 Elite Gen 5
3nm Hexagon NPU with Fused AI Accelerator
37% faster Hexagon NPU, 16% better performance per watt
Relative only -
2024Relative only
-
2023
Qualcomm Cloud AI 100 Ultra
Not stated by Qualcomm PCIe card, 64 AI cores, 576 MB SRAM, 150 W
ML capacity (INT8): 870 TOPs
Absolute -
2023
Snapdragon 8 Gen 3
4nm Hexagon NPU
Up to 98% faster Hexagon NPU; on-device models up to 10B parameters at up to 20 tokens/sec (Llama 2 7B)
Relative only -
2022Relative only
-
2021
Snapdragon 8 Gen 1
4nm 7th Gen AI Engine with Hexagon processor
AI Engine up to 4x faster than predecessor
Relative only