Marty Poniatowski, AMD | HPE Discover 2024
summary
AMD positions the Instinct MI300X with MI300X 192 GB memory as a high-end training alternative and notes Microsoft Azure exposes MI300X optimized with Hugging Face; software portability uses ROCm plugin for PyTorch/JAX so models run on multiple accelerators. EPYC CPUs drive efficiency and free budget for AI, with Ryzen at the client and Xilinx FPGA used for semi‑custom edge inference (automotive, medical imaging). Expect heavy power and cooling needs and rack design tradeoffs.
AI-written summary · about
chapters (15)
- 0:00Introduction and welcome
- 0:46AMD's GPU alternatives in AI
- 2:11Customer concerns about lock-in
- 3:56AMD's business strategy and history
- 4:40Software importance in AI
- 6:02Software development and AI workloads
- 8:10FPGA and GPU use cases
- 9:59Modernizing data centers
- 11:05Data privacy and AI challenges
- 12:29Data integrity in AI training
- 13:34Architectures and FPGA roles
- 15:00Custom cloud services and hardware
- 16:05Direct liquid cooling in AI
- 17:30Clustered systems and enterprise challenges
- 18:51Future of AI and closing remarks
