2025-12-02 Technical Intelligence Report

🤖 ROCm Updates & Software

[2025-12-02] Announcing AMD Schola v2: Next-generation reinforcement learning for Unreal Engine

Source: AMD GPUOpen

Key takeaway relevant to AMD: > AMD Schola v2 significantly matures the AMD-led ecosystem for AI in gaming and robotics by providing a high-performance, modular bridge between Unreal Engine 5 and Python-based RL frameworks, now featuring native ONNX inference support.

Summary: > AMD has released Schola v2, a major update to its open-source reinforcement learning (RL) plugin for Unreal Engine 5. The update focuses on a modular architecture that decouples the inference process, introduces native support for standardized datasets (Minari), and allows for dynamic agent lifecycles within simulations.

Details:

  • Modular Architecture: Introduces a decoupled design featuring:
    • Agent Interface: Standardized definitions for actions and observations (AInferencePawn, AInferenceController).
    • Policy Interface: Supports UNNEPolicy (native ONNX inference via Unreal’s Neural Network Engine) and UBlueprintPolicy.
    • Stepper Objects: Includes PipelinedStepper, which allows inference to overlap with simulation for higher throughput.
  • Dynamic Populations: Now supports agents being spawned or deleted mid-episode, a critical feature for realistic NPC behavior and dynamic environment simulations that was missing in v1.
  • Dataset Standards: Native support for the Minari dataset format, facilitating offline RL and imitation learning data sharing.
  • Version Compatibility:
    • Unreal Engine: 5.5 - 5.6.
    • Python: 3.9 - 3.12.
    • RL Frameworks: Gymnasium 0.29+, Stable Baselines 3 v2.3+, and Ray RLlib 2.10+.
  • CLI Enhancements: New command-line interface built with cyclopts for better error handling and auto-completion during training.

🔲 AMD Hardware & Products

[2025-12-02] AMD and HPE Expand Collaboration to Advance Open Rack-Scale AI Infrastructure

Source: AMD Press Releases

Key takeaway relevant to AMD: > AMD is challenging NVIDIA’s “Blackwell” rack-scale dominance by partnering with HPE to launch the “Helios” architecture, leveraging next-gen “Venice” CPUs and MI455X GPUs to deliver massive FP4 performance using open standards like UALoE.

Summary: > AMD and HPE have announced the “Helios” rack-scale AI architecture, an open-standard platform for large-scale AI workloads. The collaboration integrates AMD’s latest compute (Instinct MI455X, Venice EPYC) with specialized HPE/Juniper networking. Simultaneously, AMD announced the “Herder” supercomputer for HLRS in Germany, utilizing MI430X GPUs.

Details:

  • Helios Performance: Delivers up to 2.9 exaFLOPS of FP4 performance per rack.
  • Hardware Stack:
    • GPU: AMD Instinct MI455X.
    • CPU: Next-generation AMD EPYC “Venice” processors.
    • Networking: AMD Pensando Vulcano NICs and HPE Juniper Networking switches.
  • Networking Standards: Employs Ultra Accelerator Link over Ethernet (UALoE), developed in collaboration with Broadcom, to provide high-bandwidth, low-latency connectivity as an alternative to proprietary interconnects.
  • Herder Supercomputer:
    • Hardware: AMD Instinct MI430X GPUs + EPYC “Venice” CPUs.
    • Platform: HPE Cray Supercomputing GX5000.
    • Timeline: Scheduled for delivery in 2H 2027 to replace the “Hunter” flagship system at HLRS.
  • Deployment: Helios AI Rack-Scale Architecture is expected to be available globally from HPE in 2026.
  • Software Integration: Fully unified via the ROCm open software ecosystem, focusing on sovereign AI and industrial innovation for European and enterprise researchers.