<p>➀ Nvidia has transitioned from HGX B300 NVL16 to NVL8 to accommodate faster GPU-to-GPU interconnects for LLMs with sparse MoE architectures like DeepSeek and LLAMA4.</p><p>➁ NVL8 and NVL16 both have 16 GPU dies, but NVL8 integrates two GPU dies per package, connected via NV-HBI, offering up to 10 TB/s bandwidth.</p><p>➂ The shift to NVL8 will enable Nvidia to predominantly use CoWoS-L packaging, optimizing supply chain management.</p><p>➃ DGX Spark uses the GB10 chip, expected to be in mass production around June-July, with end product shipments in August-September.</p><p>➄ The N1X is Nvidia’s Windows-on-ARM processor, requiring significant software optimization for laptop use and Windows support, potentially delaying laptops into 2026.</p>