Specifications
Processor Class
Octa-core 64-bit RISC-V RV64GCV application processor
Performance Cluster
4x RISC-V C920 cores up to 2.3 GHz with 64KB I-cache and 64KB D-cache each, 1MB shared L2
Efficiency Cluster
4x RISC-V C908 cores up to 1.9 GHz with 32KB I-cache and 32KB D-cache each, 512KB shared L2
NPU Throughput
12 TOPS INT8 with support for INT4, INT8, INT16, FP8, FP16 and BF16 datatypes
LLM Capability
DeepSeek-7B runs at 8 tokens/s; a 4-die multi-chip cascade reaches 25 tokens/s
AI Frameworks
TensorFlow, Caffe, HuggingFace and ONNX model support
GPU
Unnamed 3D engine rated 50.34 GFLOPS and 3.14 GPixel at 786 MHz
Graphics APIs
Vulkan 1.1/1.2, OpenCL 1.1/1.2/2.0 and OpenGL ES 1.1 through 3.2
Video Decode
Up to 4Kp120 H.265/HEVC, H.264/AVC and AV1, plus 1080p60 VP6/7/8, AVS/AVS+, VC1 and MPEG4
Video Encode
4Kp60 H.265/HEVC and H.264/AVC; JPEG encode up to 1080p at 312 FPS
Memory
Up to 16GB via 2x 32-bit LPDDR4/LPDDR4X channels at up to 4266 MT/s
Storage Interface
PCIe 3.1 / SATA 3.0 combo interface for NVMe or SATA media
Display Outputs
HDMI 1.4/2.0 up to 4Kp60 with HDCP 1.4, plus DisplayPort over a USB3.1/DP 1.4/eDP 1.5 combo
Camera Interface
4-lane MIPI CSI-2 D-PHY interface with flexible lane mapping
Module Form Factor
A210 SODIMM V2 system-on-module with a companion carrier board
Target Workloads
Vision inference, transformer and LLM inference, robotics perception and industrial analytics
Overview
The Zhihe A210 is one of the first high-performance RISC-V application processors aimed squarely at edge AI rather than microcontroller-class inference. It pairs a big.LITTLE-style arrangement of four C920 performance cores and four C908 efficiency cores with a 12 TOPS INT8 NPU that accepts INT4 through BF16 datatypes, which allows the same silicon to run quantised vision models and transformer workloads.
The headline capability is generative inference at the edge: the vendor documents DeepSeek-7B at 8 tokens/s on a single die and 25 tokens/s when four dies are cascaded in a multi-chip configuration. Storage is handled over a PCIe 3.1 or SATA 3.0 combo interface, so local NVMe model storage does not require a bridge chip.
Documentation has surfaced for both the SoC and an A210 SODIMM V2 development platform built from a carrier board plus a SODIMM-format module, which is the practical route to evaluation. Typical deployments are vision-heavy industrial analytics, robotics perception stacks and on-premise LLM endpoints where an x86 or GPU solution would be over-specified on power.
Key Benefits
12 TOPS INT8 NPU with INT4 to BF16 datatype coverage, 8 tokens/s DeepSeek-7B edge LLM inference, 4Kp120 AV1/H.265 decode, and an open RISC-V RV64GCV architecture that avoids proprietary accelerator lock-in.
Applications
RISC-V edge AI gateways, robotics perception controllers, on-premise LLM inference endpoints, industrial vision inspection, smart retail analytics, and multi-die cascade inference clusters.
Request a Quote — ZHIHE A210 — 12 TOPS RISC-V EDGE AI SOC AND A210 SODIMM V2 DEVELOPMENT PLATFORM
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →