回复@A大师-懒惰的投资人: 我接受黄说的Deepseek的创新-20250302022122

回复 @A大师-懒惰的投资人 : 我接受黄说的Deepseek的创新不会减少对算力的需求（其实就是算力需求会继续增长）的观点，也大致理解他认为NVDA在算力上的地位不会受到威胁的意思。
查看图片 // @A大师-懒惰的投资人 :回复 @babygirl :这是DS自己的回答 (Due to technical issues, the search service is temporarily unavailable.)
DeepSeek, like many AI and deep learning companies, likely relies on **Nvidia's CUDA** for GPU acceleration, given its widespread adoption in the industry. Here's a breakdown of why:
1. **GPU Dependency**: Training large AI models typically requires powerful GPUs. Nvidia's GPUs (e.g., A100, H100) dominate this space, and CUDA is the primary platform for optimizing computations on these GPUs.
2. **Frameworks and Libraries**: Popular deep learning frameworks (e.g., PyTorch, TensorFlow) are built on CUDA for GPU acceleration. Unless DeepSeek uses custom alternatives, they likely depend on these CUDA-optimized tools.
3. **Cloud Infrastructure**: If DeepSeek uses cloud providers (AWS, Azure, GCP), their GPU instances are predominantly Nvidia-based, requiring CUDA for efficient computation.
4. **Ecosystem Lock-In**: The deep learning ecosystem heavily favors CUDA due to compatibility with libraries (cuDNN, NCCL) and pre-trained models. Switching to alternatives like AMD ROCm would involve significant effort and potential performance trade-offs.
**Exceptions**: If DeepSeek uses custom hardware (e.g., TPU-like accelerators) or prioritizes non-Nvidia solutions, they might reduce CUDA dependency. However, there's no public evidence of this.
**In short**, while specific internal infrastructure details aren't public, it's highly probable that DeepSeek depends on CUDA for GPU-accelerated AI workloads, as do most companies in the field.