Intellifusion DeepEye1000 (IPU)
Product Overview
Intellifusion is a Shenzhen-based AI company known for its full-stack "algorithms + chips + big data" capabilities, focusing on digital cities, smart security, and edge vision. DeepEye1000 (also known as "Yuntian Chuxin DeepEye1000") is the company's first deep learning neural network processor chip for computer vision, released on November 13, 2019 at the 21st China Hi-Tech Fair, and serves as the carrier of its second-generation neural network processor, the NNP200.
The chip is positioned for edge and endpoint vision inference, targeting smart security, intelligent transportation, industrial inspection, robotics, drones, and similar scenarios. Its biggest innovation is the compute-storage fusion architecture + reconfigurable compute array: compute efficiency exceeds 99%, DDR memory access bandwidth drops by 77% and power consumption drops by 60% compared with traditional solutions — a typical domestic "compute-storage/reconfigurable" innovative architecture rather than a general-purpose GPU.
Around DeepEye1000, Intellifusion has launched edge inference product forms such as the IPU X1000 module and IPU X1000 / X2000 / X5000 accelerator cards, commercializing this self-developed vision NPU as modules and PCIe cards.
Core Specifications
| Parameter | Value |
|---|---|
| Architecture | Heterogeneous multi-core visual AI coprocessor (compute-storage fusion + reconfigurable compute array, self-developed NNP200 quad-core neural network processor) |
| Process Node | 22nm FD-SOI |
| FP16 / BF16 Compute | Not disclosed (chip level; the IPU X1000 card is listed as supporting FP16 + INT16/12/8) |
| INT8 Compute | 2.0 TOPS (peak, INT16/INT12/INT8 mixed precision; IPO disclosure lists 1.5 TOPS at INT12) |
| FP32 Compute | Not disclosed |
| Memory Capacity | Not disclosed (chip level); the IPU X1000 module/card comes with 16GB DDR4 |
| Memory Type | DDR4 (module/card form) |
| Memory Bandwidth | Not disclosed (officially stated that compute-storage fusion reduces DDR access bandwidth by 77%) |
| TDP | Not disclosed (chip level); the IPU X1000 card has typical power of about 20W (passive cooling) |
| Interconnect | Not disclosed |
| Interface | SoC (chip); the IPU X1000 card uses PCIe 2.0 ×4 |
| Launch | 2019-11 |
| Mass Production/Availability | 2019-11 (in mass production, commercially deployed at scale) |
Note: Chip-level peak compute is about 2.0 TOPS (consistent across Baidu Baike and the SSE IPO inquiry response); "10 TOPS" appears in IPU X1000 accelerator card materials (with the NNP clock raised to 1GHz) and is a board/module-level figure, not the bare-chip peak — this table uses the bare-chip 2.0 TOPS. FP16/FP32 bare-chip compute has not been officially broken out and is therefore marked "Not disclosed".
Key Features
- Compute-storage fusion + reconfigurable compute array: compute efficiency above 99%, significantly reducing DDR access and power consumption — the core innovation distinguishing it from general-purpose GPUs.
- Self-developed instruction set and quad-core NNP200: ships with a custom instruction set and programming framework, supporting porting of mainstream CNN algorithms (facial recognition, etc.).
- Heterogeneous multi-core parallelism: the CPU is an Alibaba T-Head XuanTie 810 (RISC-V) embedded processor (1.2GHz) + dual-core vision DSP (550MHz), supporting 4K@30fps and parallel real-time analysis of 4 HD video streams.
- High energy efficiency: about 2 TOPS/W, better than contemporaneous competitors such as HiSilicon's Hi3559A (about 1.5 TOPS/W).
- Industrial grade and multiple form factors: operating temperature of -40℃ ~ 85℃ (industrial grade); deliverable as a module (mini-PCIe) or PCIe card.
Vendor Information
| Parameter | Value |
|---|---|
| Company | Shenzhen Intellifusion Technologies Co., Ltd. (Intellifusion) |
| Headquarters | Shenzhen, China |
| Founded | 2014 (listed on the STAR Market in 2023) |
Use Cases
- ✅ Edge/endpoint vision inference: smart security, intelligent transportation, industrial inspection
- ✅ Lightweight endpoint devices such as AI cameras, robots, and drones
- ✅ Multi-stream video structuring: a single chip supports real-time analysis of 4 channels of 1080P video
- ❌ Large-model training / high-compute data center training (inference positioning, limited compute)
- ❌ General-purpose GPU compute ecosystem (requires porting via the self-developed compilation framework, not CUDA)
Related Cards
- TsingMicro TX81 — Also on a domestic reconfigurable compute architecture innovation route
- Vastai VA10 — Domestic cloud vision inference accelerator card
- Lightmatter Envise — International compute-storage (silicon photonics) route comparison