DeepSeek V4 Just Launched on Huawei Chips First - No Nvidia Required.

On April 18, 2023, DeepSeek launched its latest AI model, V4, exclusively on Huawei's Ascend processor, marking a significant shift in the global AI chip market dynamics. This move challenges Nvidia’s dominance and highlights the rapid development of China's domestic AI ecosystem.

Key points from the launch include:

1. Model Efficiency: V4 is highly optimized for inference compute, reducing per-token computation by 73% compared to its predecessor (V3.2). It leverages dual-mode compressed attention mechanisms (CSA + HCA) and FP4/FP8 mixed precision, aligning perfectly with Ascend’s hardware capabilities.

2. Training Data: V4 was pretrained on over 33 trillion tokens, nearly doubling the dataset size of its predecessor. This extensive training enhances model performance but also underscores the need for high-performance computing resources during training phases.

3. Ecosystem Support: The launch marks a critical milestone in the development of CANN (Compute Architecture Neural Network), Huawei’s AI framework. With 4 million developers and full open-source status, CANN is rapidly catching up to CUDA's extensive developer base and ecosystem depth.

4. Geopolitical Implications: U

Read the full article at Towards AI - Medium

---

Want to create content about this topic? Use Nemati AI tools to generate articles, social posts, and more.