Beijing (Gasgoo)- On December 30, Black Sesame Technologies introduced its Huashan A2000 family, a high-performance chip platform designed for next-generation AI models.
Huashan A2000 family
The A2000 family includes three products: A2000 Lite, A2000, and A2000 Pro, catering to varying levels of autonomous driving needs. A2000 Lite focuses on urban intelligent driving, A2000 supports full-scenario general-purpose intelligent driving, and A2000 Pro is designed for advanced full-scenario applications.
The Huashan A2000 family represents a significant leap in autonomous driving technology. These chips integrate industry-leading CPU, DSP, GPU, NPU, MCU, ISP, and CV units, achieving high integration and multi-tasking capabilities on a single chip. Advanced ISP technology with 4-frame exposure and 150dB HDR improves performance in tunnels and nighttime scenarios, enhancing image processing capabilities.
The chips support seamless data processing, including desensitization, compression, encoding, and storage, ensuring smooth operation while advancing algorithm innovation. Offering up to four times the computing power of the company's current flagship chips, the A2000 family natively supports Transformer models and provides flexible scalability for multi-chip configurations, covering applications from NOA (Navigation on Autopilot) to Robotaxi.
Beyond smart vehicles, the A2000 chips also excel in robotics and general computing. Notably, they meet the "small brain" and "large brain" demands of robots, facilitating the transition from prototype development to mass production.
Jiushao NPU
Black Sesame also introduced its proprietary NPU architecture, "Jiushao" (name in Chinese pinyin), designed as the computational core for high-performance AI chips. Supported by innovations such as the BaRT universal AI toolchain and BLink dual-die interconnect technology, Jiushao enables enhanced performance and scalability for the A2000 family.
The Jiushao NPU features a cutting-edge large-core architecture, supporting real-time inference for AI large models, reducing latency, and prioritizing complex computational tasks. Black Sesame Technologies claims that it achieves the highest safety levels in the industry, preventing errors during inference and ensuring consistency in training and deployment for robust autonomous driving systems.
Key features include support for INT8/FP8/FP16 mixed precision, hardware acceleration for fine quantization and Transformer models, and a three-layer memory architecture for low latency and high throughput. This architecture balances performance, bandwidth, and cost by reducing dependence on external storage bandwidth while optimizing efficiency and scalability.
