Oct 8, 2025, 7:51 AM ETConference Presentation, Keynote

Ramine Roane, CVP @ AMD: Unlocking the Next Wave of AI Open Foundations for Intelligent Systems

Scaling Law Dynamics: The 2020 scaling law established a proportional relationship between model intelligence and the log of compute/parameters for four years; however, public data exhaustion has shifted the growth driver.
New Compute Distribution: Intelligence is now driven by compute distributed across pre-training, post-training, and reasoning inference, rather than pre-training alone.
Inference Evolution: Models have transitioned from one-shot word prediction to verbose reasoning models that generate and assess multiple "chains of thought" to determine accuracy.
Post-Training Reinforcement: Training methodologies now utilize verification loops (e.g., re-injecting math results or compiling/running code) to reinforce correct behaviors via reinforcement learning.

Product Timeline: AMD is deploying the MI350 series currently and plans the MI400 series for 2026.
Memory Advantage: The MI350/MI400 racks feature HPM (High Performance Memory) delivering 2.8x the memory capacity of the NVIDIA Grace Blackwell NVL72.
Performance Metrics: In benchmarks against NVIDIA Blackwell, AMD MI355x achieved 1.3x higher performance, attributed to NVIDIA's reliance on closed-source TensorRT-LLM (limited to FP4) versus AMD's open-source optimization.
Edge Deployment: AMD chips are currently utilized across diverse endpoints including laptops, autonomous vehicles, satellites (Mars rovers), and medical equipment.

Open Standards: AMD's software stack is entirely open-source, rejecting closed-source proprietary models; their CUDA equivalent, ROCm, is integrated directly into PyTorch source code.
Framework Integration: Full compatibility and optimization exist for PyTorch, JAX, ONNX, Hugging Face (hosting 1.8 million models), and OpenAI Triton.
Inference Serving Optimization: AMD supports and optimizes VLLM, SGLang, and LMD (LLMD) for high-utilization inference serving.
Disaggregated Architecture: AMD enables LMD to split inference into "prefill" (prompt analysis) and "decode" (generation) tasks, allowing specialized GPU allocation to reduce cost per token and latency.
Developer Incentives: AMD is offering developers free GPU compute hours and accounts to test the stack, which is fully compatible with existing Hugging Face and PyTorch workflows.

Enterprise Stack: A new enterprise AI stack is scheduled for release in the second half of the year, utilizing Kubernetes, telemetry for GPU utilization monitoring, and distributed job management.
Strategic Rationale for Sovereignty: The presentation argues that nations must adopt sovereign compute to avoid historical economic setbacks similar to those experienced by China and India during the Industrial Revolution.
Key HPC Partnerships:
- United States: Oak Ridge National Laboratory and the National Energy Technology Laboratory (top two global supercomputers) utilize AMD GPUs for FP64 HPC and AI training.
- Europe: Lumi (Finland), the top European supercomputer, is based on AMD hardware; Silo AI (acquired by AMD) was a major driver of this work.
- France: Partnerships established with CEA and Gen-C for GPU deployment.
- United Arab Emirates: G42 is constructing a ~1 gigawatt data center in Grenoble, France, based on AMD GPUs.
- Middle East: A major deal has been announced with Saudi Arabia and other nations.
Competitive Differentiation: AMD advocates for a multi-vendor strategy for nations to prevent dependency on a single supplier controlling pricing and supply schedules.

Open Foundation: AMD commits to open software, open hardware, and open platform standards including OCP racks, Ultra Ethernet, and UCIe for scaling.
Market Impact: The company asserts that AI is the first non-human intelligence to compete with human intelligence, necessitating a global industrial-scale response similar to the Industrial Revolution.
Continuous Development: AMD partners with OpenAI to ensure full support for Triton kernels, enabling optimized model execution on AMD hardware.