AI Compute Landscape: GPUs, Network, and Storage as Core Engines

ChatGPT's release is likened to the birth of Windows, positioning large language models as the new entry point for information systems and potentially reshaping the software ecosystem.

The report emphasizes that computation is the core engine of AI compute, with GPUs serving as the primary pillar for training and inference, outpacing Moore's Law due to strong AI and high‑performance market demand, and driving rapid growth in AI server markets.

Networking has become the main bottleneck for AI compute scaling; solutions such as NVIDIA's NVLink and NVSwitch under the InfiniBand architecture, along with high‑capacity optical modules (800 G, 1.6 T), are raising GPU inter‑communication capabilities to new heights.

Storage faces the "memory wall" challenge; conventional NAND and DRAM are approaching process limits, prompting exploration of multi‑dimensional solutions like 3D stacking, while High‑Bandwidth Memory (HBM) is emerging as a key component for high‑performance GPUs.

The analysis is extracted from the 2023 AI Series Research Report: "AI Compute Research Framework," which provides a comprehensive view of compute chips, technologies, networking, and storage dimensions.