亞馬遜晶片年化營收超200億美元

🔑 增強重點摘要

• •Amazon's custom silicon strategy centers on the Graviton (CPU) and Trainium/Inferentia (AI accelerator) product lines, which allow AWS to decouple infrastructure costs from third-party chip vendor pricing.
• •The $20 billion run-rate is largely driven by internal consumption within AWS, where Amazon replaces expensive NVIDIA-based instances with its own cost-optimized silicon for high-scale cloud workloads.
• •Amazon is increasingly positioning its custom chips as a key differentiator in the 'AI sovereignty' market, enabling enterprise customers to run large-scale models with better price-performance ratios than standard GPU-as-a-service offerings.

🛠️ 技術深入

• •Trainium2 chips are designed specifically for large language model (LLM) training, featuring high-bandwidth memory (HBM) and optimized interconnects for multi-node scaling.
• •Inferentia2 utilizes a custom architecture optimized for low-latency, high-throughput inference, supporting transformer-based models with native hardware acceleration for common operations like LayerNorm and Softmax.
• •The AWS Neuron SDK serves as the software abstraction layer, allowing developers to compile models from frameworks like PyTorch and TensorFlow to run on custom silicon without extensive code refactoring.
• •Graviton4 processors utilize a 64-bit Neoverse V2 core architecture, providing significant improvements in performance-per-watt over previous generations for general-purpose compute.

🔮 前景展望基於引用來源的 AI 分析

⏳ 時間線