China's leading mass-produced autonomous driving SoC platform — from L2+ highway pilot to L4 robotaxi
Black Sesame Technologies' Huashan (华山) series represents one of the most widely deployed domestic autonomous driving SoC platforms in China. The A1000 is built on 16nm automotive process, featuring a proprietary DynamAI NN engine delivering 58 TOPS (INT8) with support for 4-bit/8-bit mixed-precision inference, achieving 80% average MAC array utilization on convolution layers. It integrates a NeuralIQ ISP pipeline supporting 16-channel 8MP camera input with 3-exposure HDR at 140dB dynamic range.
The next-generation A2000 series advances to 7nm automotive process with a proprietary JIUSHAO (九韶) NPU architecture — the largest NPU core specification in the industry — delivering up to 1,000 TOPS per chip. Multi-chip BLink interconnect scales to 4,000 TOPS. ResNet-50 inference reaches 328 FPS, surpassing NVIDIA Thor's 300 FPS benchmark. The A2000 integrates Transformer hard-acceleration modules for native large-model acceleration. Mass production began in 2025 with tier-1 OEM design wins. Data source: Black Sesame official product specifications (2025).
Proprietary neural network processor with 58 TOPS (A1000) to 1000 TOPS (A2000) INT8 throughput, supporting INT4/INT8/FP8/FP16 mixed-precision inference.
16-channel (A1000) to 20-channel (A1000 Pro) HD camera input pipeline with 3-exposure HDR, 140dB dynamic range, low-light denoising, and LED flicker mitigation.
ISO 26262 ASIL-B (A1000) / ASIL-D (A2000) with dual-core lockstep safety processor, ECC/Parity protection, secure boot, and OTP key storage.
A2000 BLink interconnect enables scaling from single-chip 1000 TOPS to 4000 TOPS across 4 chips for L4/Robotaxi compute requirements.
Integrated Transformer hard-acceleration module in A2000 for native BEV/Occupancy Network/large-model inference without software fallback.
PCIe Gen3, USB 3.0, Gigabit Ethernet, CAN-FD, eMMC/SD, multi-lane MIPI-CSI, supporting full sensor fusion architecture.
A1000 deployed in Geely Galaxy E8, Lynk & Co 07/08, Dongfeng eπ007/008. 500K+ units shipped across multiple OEM programs.
A2000 achieves 328 FPS on ResNet-50, exceeding NVIDIA Thor's 300 FPS. 7nm process delivers efficiency comparable to 4nm competitors.
| Parameter | Specification |
|---|---|
| A1000 — Process | 16nm FFC Automotive |
| A1000 — CPU | 8× ARM Cortex-A55 @ 1.5GHz, 64-bit |
| A1000 — NPU | DynamAI NN, 58 TOPS (INT8), INT4/INT8 mixed-precision |
| A1000 — ISP | NeuralIQ, 16ch × 8MP, 3-exp HDR, 140dB |
| A1000 — CV | 5× Vision DSP + 4K H.264/H.265 codec |
| A1000 — Memory | 64-bit LPDDR4 automotive-grade |
| A1000 — TDP | 18W typical |
| A1000 — Package | FCBGA 25×25 mm, 0.8mm pitch |
| A1000 — Interfaces | PCIe Gen3 ×4, USB 3.0 DRD + USB 2.0, GbE, eMMC/SD, CAN-FD, UART/SPI/I2C/I2S/GPIO |
| A1000 — Safety | ISO 26262 ASIL-B + AEC-Q100 Grade 2 |
| A1000 — Op. Temp | -40°C to +105°C |
| A1000 Pro — NPU | JIUSHAO, 106 TOPS (INT8) |
| A1000 Pro — CPU | 16× Cortex-A55 @ 1.5GHz, 60K DMIPS |
| A1000 Pro — ISP | 20ch HDR, 140dB |
| A1000 Pro — Safety | ASIL-B + ASIL-D Safety Island |
| A2000 — Process | 7nm Automotive |
| A2000 — NPU | JIUSHAO, up to 1000 TOPS (INT8), INT8/FP8/FP16 |
| A2000 — Transformer | Native hard-acceleration module |
| A2000 — Multi-Chip | BLink interconnect, up to 4000 TOPS |
| A2000 — Benchmark | ResNet-50: 328 FPS (vs. NVIDIA Thor 300 FPS) |
| A2000 — Safety | ASIL-D certified |
| A2000 — Applications | L2+ to L3 AD, Robotaxi, AI BOX, Robotics |
| Mass Production | A1000: 500K+ units (2025); A2000: MP start 2025 |
| OEM Customers | Geely, BYD, Dongfeng, FAW, Chery |
| Data Source | Black Sesame official product specifications (2025) |
Contact our sales team for pricing, lead times, and customization options.