Skip to main content

Sophgo BM1688

Product Overview​

The BM1688 is the highly integrated edge TPU processor released by SOPHGO in 2023, positioned as the "edge intelligence heart" for deep learning and machine vision. Compared with the BM1684X, it further strengthens integration and professional imaging capability, integrating an 8-core ARM Cortex-A53, a proprietary NPU, a professional security ISP, and a video codec unit on a single chip.

The BM1688 delivers 16 TOPS INT8 / 32 TOPS INT4 / 4 TFLOPS FP16-BF16 / 0.5 TFLOPS FP32 of compute, supports intelligent analysis of 16 HD video streams, 16-channel hardware decoding, and 10-channel hardware encoding, and integrates a 4Kp60 professional security ISP (with hardware acceleration for binocular depth, image stitching, fisheye expansion, etc.). It can run edge large models such as ChatGLM3-6B, Qwen-7B, and Qwen2.5-VL-3B locally, making it ideal for smart cameras, edge boxes, and micro-servers.

Core Specifications​

ParameterValue
ArchitectureSophgo highly integrated edge TPU, integrating 8-core ARM Cortex-A53 @ 1.6GHz + proprietary NPU
Process NodeNot disclosed (estimated 12nm-class, same generation as the BM1684X)
FP16 / BF16 Compute4 TFLOPS
INT8 Compute16 TOPS (up to 32 TOPS at INT4)
FP32 Compute0.5 TFLOPS
Memory Capacity8 / 16GB (LPDDR4/LPDDR4X 64-bit @ 4266 Mbps, also supports DDR4 2×32-bit)
Memory TypeLPDDR4 / LPDDR4X
Memory BandwidthNot disclosed
TDPNot disclosed (paired whole devices typically draw about 7.2 W; lower for the chip alone)
InterconnectDual Gigabit Ethernet, PCIe, USB, HDMI, MIPI-CSI, CAN FD
InterfaceSoC on-board (micro-server SE9 / module)
Launch2023
Mass Production/AvailabilityReleased and mass-produced in 2023

Key Features​

  • Edge large models: supports local inference of ChatGLM3-6B, Qwen-7B, Qwen2.5-VL-3B, and more.
  • Professional security ISP: 4Kp60 image processing with hardware acceleration for binocular depth, stitching, and fisheye expansion.
  • Multi-stream video: 16-channel HD intelligent analysis, 16-channel decoding, 10-channel encoding.
  • Full precision support: INT4 / INT8 / FP16-BF16 / FP32 mixed precision.
  • Rich peripherals: PCIe / USB / HDMI / MIPI-CSI / CAN FD, convenient secondary development.
  • Full-stack frameworks: mainstream frameworks such as PyTorch and TensorFlow, with the SophonSDK one-stop toolchain.

Vendor Information​

ParameterValue
CompanySOPHGO
HeadquartersBeijing, China
Founded2019

Use Cases​

  • ✅ Smart cameras, edge boxes, micro-servers, smart cities, intelligent transportation, private large-model deployment
  • ❌ Large-model training, ultra-high-density (>16-channel) concurrent analytics (weaker than the BM1684X)

References​