LQ7000 - LQ Silicon

LQ7000

LQ7000 NPU focuses on cutting-edge AIPC scenarios, achieving an ultra-low power consumption ratio and industry-leading computational density at the edge through an innovative architecture. It deeply integrates with an independent software ecosystem, fully compatible with mainstream large language model frameworks and supporting dynamic expansion, enabling highly optimized deployment of large models. By redefining the AIPC experience with robust technology, it paves the way for a secure and controllable future of intelligent terminals.

Features

High-Performance Architecture

Utilizes the independently innovative EE-DNA architecture, supporting mainstream large models like LLAMA and DeepSeek, while also being compatible with traditional neural network models such as CNN and RNN, significantly improving inference performance and energy efficiency.
Powerful Computing Capability

A single chip delivers robust INT8 computing power and supports multi-chip collaboration for flexible scalability, meeting the demands of large-scale AI inference tasks.
Efficient Memory Support

Integrates a high-speed LPDDR5x@6400Mbps memory interface, supporting ample memory capacity to ensure efficient data access during large model inference.
Flexible Product Form Factors

Supports various form factors including chip-scale packaging and PCIe cards, facilitating integration into different hardware platforms to suit diverse application scenarios.
Low-Power Design

Achieves ultra-low power operation through advanced power management technology and dynamic frequency scaling, making it particularly suitable for mobile and edge computing devices.
Independent Compiler

The LLVM compiler and graph compiler work in synergy to optimize data movement between DDR, L2M, and LRM, reducing off-chip data transfer, thereby improving efficiency and lowering energy consumption.
Programmability

Provides an independent Software Development Kit (SDK) and programming interfaces, supporting user-defined model optimization and algorithm porting, ensuring the chip can adapt to the evolving demands of future AI models and algorithms.

Solution

Chipset Solutions

Heterogeneous chip packaging solution, building a high-efficiency, low-cost AIPC core engine.
Homogeneous chip packaging solution, seamlessly aggregating the computing power of multiple NPUs to achieve linear growth in performance.

Solution

System Solutions

Based on the LQ7000 chip, we create PCIe standard interface accelerator cards. Single or dual (and above) chips form a high-density AI inference solution for a single machine. Supports mixed precision, compatible with multiple platforms and mainstream operating systems, suitable for AI workstations, all-in-one large model machines, etc.
M.2 card, based on the standard M.2 interface specification, integrates the powerful computing capability of a single LQ7000 chip. Supports local operation of large models with tens of billions of parameters, features low-power design, and is compatible with multiple systems. It can be flexibly deployed in AI PCs and other devices, providing local large model inference capabilities for consumers, offices, and industrial applications, ensuring data privacy and low-latency response.