Ramine Roane, AMD | theCUBE + NYSE Wired: Media Week - Cyber & AI Innovators Summit
summary
AMD claims a pivot into AI with MI300X GPUs optimized for large models, reporting about $5 billion in sales this year and citing Frontier/MI250X as prior exascale proof points. Architecture uses chiplets and 3D stacking to pack 2.4X HBM memory advantage, hosting Llama 3-405B in a single 8‑GPU server (vs two for 80GB cards), supports HIP/CUDA compatibility, uses Infinity Fabric and Ethernet/Ultra Ethernet, and shows near‑linear scaling across hundreds to thousands of GPUs for training while targeting high‑volume inference.
AI-written summary · about
chapters (15)
- 0:00Introduction and welcome
- 0:23AMD’s AI leadership position
- 1:29Supercomputers and MI300X GPU
- 3:59Market growth and TAM
- 4:43Chiplet architecture advantage
- 6:29Memory density and SRAM limitations
- 8:15NVIDIA’s Achilles heel power
- 9:19CUDA moat and compatibility
- 11:07Systems design and ZT Systems
- 12:38Networking and Infinity Fabric
- 14:30Training vs inference markets
- 17:02Scaling laws and AI economics
- 19:33Data sources and AGI potential
- 22:03AGI timeline and robotics
- 23:50Closing remarks
