Rockchip RK3588 NPU (Orange Pi / SBC)
other RK3588 (ARM SoC + 6 TOPS NPU) 2022
Fields
id
accelerators/rockchip-rk3588-npuVendor
other
Architecture
RK3588 (ARM SoC + 6 TOPS NPU)
Release year
2022
Process (nm)
8
Form factors
sbcintegrated
Memory type
LPDDR4x/LPDDR5 (shared system memory)
Bandwidth basis
Unified system-memory SBC; the 6 TOPS NPU shares DRAM with the CPU. No discrete VRAM or GPU-class bandwidth figure applies. The binding constraint is the NPU compute ceiling, not bandwidth.
Peak FLOPS
precisionint8tflops6denseyesvendor_claimyessource_idairockchip-rknn-llm-benchmark-tablebasisvendor-spec-sheetbasis_detailRK3588 NPU 6 TOPS (INT8), vendor-published.unit—
Unified memory
yes
Consumer part
yes
notes
TRAP for 7B+; edge toy for small models. The $80-200 RK3588 SBC runs Rockchip's RKLLM (airockchip/rknn-llm, v1.3.1), which is real and actively maintained, but the measured benchmark table (RK3588, w8a8, 128-in/64-out) makes the wall obvious: Qwen2.5-1.5B 16.7 tok/s, Phi3-3.8B 7.45, ChatGLM3-6B 4.98 - the largest they benchmark is already ~5 tok/s. A real 7B lands ~3-4 tok/s decode and wants ~5-6 GB for weights+KV at w8a8. Fine for a sub-4B on-device assistant or VLM demo; a trap if the goal is 7B+.
Notes
TRAP for 7B+; edge toy for small models. The $80-200 RK3588 SBC runs Rockchip's RKLLM (airockchip/rknn-llm, v1.3.1), which is real and actively maintained, but the measured benchmark table (RK3588, w8a8, 128-in/64-out) makes the wall obvious: Qwen2.5-1.5B 16.7 tok/s, Phi3-3.8B 7.45, ChatGLM3-6B 4.98 - the largest they benchmark is already ~5 tok/s. A real 7B lands ~3-4 tok/s decode and wants ~5-6 GB for weights+KV at w8a8. Fine for a sub-4B on-device assistant or VLM demo; a trap if the goal is 7B+.
Cross-references
This record references no other records — it is a leaf. It is cited by 7.