Performance
AI Hydra is designed for high-throughput training on consumer hardware, without relying on a GPU.
Its performance comes from two complementary design pillars:
1. Aligned Training Pipeline
The training path is tightly aligned across memory, batching, and model execution.
- Replay memory stores full episodes and exposes variable-length sequences via metadata
- Sequences are constructed without expensive transformation or reshaping
- Batches are fed directly into the model’s sequence-forward path
This creates a continuous pipeline:
memory → sequence assembly → batched training → model forward
Because each stage is designed to match the next, the system avoids unnecessary copying, slicing, or recomputation.
2. Lean Simulation Loop
The core run loop is intentionally minimal and stays focused on simulation and training only.
- No rendering or plotting occurs in the training process
- Telemetry is published asynchronously via a ZeroMQ PUB socket
- Visualization, plotting, and event correlation are handled by a separate client (TUI)
This keeps the hot path free of UI and analysis overhead, allowing the simulation to run at full speed.
Result
- High episode throughput on CPU-only systems
- Stable real-time telemetry without blocking the simulation
- Fast iteration cycles with immediate observability
AI Hydra achieves speed not through hardware, but through alignment and separation of concerns. The system is optimized so that the cost of learning is dominated by model computation, not data movement or orchestration.