AMD Talks Up Open-Source Software For AI, Introduces Instinct MI300X

Beyond launching Bergamo and Genoa-X, AMD's AI Day also showcased the new AMD-Pensando DPU offerings and also previewed more of their next-gen Instinct MI300 accelerator APU.

AMD's Victor Peng talked up AMD's open-source software offerings around AI and their software ecosystem at large under Linux. Victor talked up AMD's open-source software offerings and the advancements made with the ROCm stack. Though for frequent Phoronix readers keeping up with my ROCm update articles, it's not particularly new. It's always great seeing AMD and hardware vendors at large promoting their open-source software efforts.

Hugging Face CEO Clément Delangue was also at the AMD AI Day event to talk-up the importance of open science and open-source in the world of AI. Hugging Face and AMD have also formed a partnership to ensure full AMD support going forward for their software and hosted models, CI/CD on AMD hardware, etc.

The AMD Instinct MI300A with 24 Zen 4 cores, 128GB of HBM3 memory, and CDNA3 is now sampling to customers. AMD also announced the Instinct MI300X as a GPU-only version of CDNA3 with 192GB of HBM3 memory and focused on handling large language models.

The AMD Instinct MI300X was demoed with the Falcon-40B large language model on a single MI300X running entirely in-memory. A single MI300X can handle large language models up to approximately 80 billion parameters.

Lisa also introduced the AMD Instinct Platform with eight MI300X within an OCP infrastructure platform and provides 1.5TB of HBM3 memory for handling very large AI LLMs.

The Instinct MI300A has been sampling since earlier this quarter to HPC/AI customers. The MI300X and the AMD Instinct Platform to begin sampling in Q3. Full production ramp is expected in Q4.