diff --git a/tools/gsf/green-code-skill/AGENTS.md b/tools/gsf/green-code-skill/AGENTS.md index 774b0bc..a4c72de 100644 --- a/tools/gsf/green-code-skill/AGENTS.md +++ b/tools/gsf/green-code-skill/AGENTS.md @@ -179,6 +179,10 @@ For full detail on any pattern, the agent can fetch the URL. - Use energy-efficient AI/ML models — match model size to task, routing pattern. → /architecture/technology-selection/energy-efficent-models - Use serverless cloud services — FaaS, serverless containers, serverless databases. → /architecture/technology-selection/use-serverless +### Operations (1 pattern) + +- Optimize power-performance for edge AI inference — power-aware scheduling, thermal-aware throttling, battery-level policies, DVFS; named modes (Performance/Balanced/Power Saver) calibrated from energy-latency-accuracy triplet. → /operations/optimize-edge-ai-power-performance + ### Agentic AI patterns Applied to LLM-driven orchestration, agentic loops, and tool-calling workflows. These patterns target the energy and emissions profile of agentic AI systems, where each loop iteration is a model call and call counts can compound rapidly across planning, retries, and tool invocations. diff --git a/tools/gsf/green-code-skill/GETTING_STARTED.md b/tools/gsf/green-code-skill/GETTING_STARTED.md index 523535c..d411a67 100644 --- a/tools/gsf/green-code-skill/GETTING_STARTED.md +++ b/tools/gsf/green-code-skill/GETTING_STARTED.md @@ -221,5 +221,5 @@ Inside this skill package: - `references/sci-formula.md` — SCI grounding and what each SCI term represents - `references/patterns-development.md` — 23 Development patterns with detection signals and fixes - `references/patterns-architecture.md` — 17 Architecture patterns -- `references/patterns-ai-ml.md` — 9 AI/ML model and training patterns + 10 agentic AI workload patterns +- `references/patterns-ai-ml.md` — 10 AI/ML model and training patterns + 10 agentic AI workload patterns - `references/trade-offs.md` — contextual considerations for applying patterns well diff --git a/tools/gsf/green-code-skill/README.md b/tools/gsf/green-code-skill/README.md index fe85b68..63b69be 100644 --- a/tools/gsf/green-code-skill/README.md +++ b/tools/gsf/green-code-skill/README.md @@ -22,7 +22,7 @@ The skill covers: - **23 patterns from the GSF Development category** — data handling, media efficiency, web performance, cloud deployment - **17 patterns from the GSF Architecture category** — system topology and technology selection -- **9 AI/ML model and training patterns** cross-referenced from the Development and Architecture categories +- **10 AI/ML model and training patterns** cross-referenced from the Development, Architecture, and Operations categories - **10 agentic AI workload patterns** applying the same SCI principles to LLM-driven orchestration — model routing, retry budgets, planning depth caps, response and prompt caching, parallel tool calls, two-stage retrieval, structured output, `max_tokens` hygiene, and loop-safe tool calling Choices this implementation made, documented so readers can evaluate them against their own context: diff --git a/tools/gsf/green-code-skill/SKILL.md b/tools/gsf/green-code-skill/SKILL.md index d875513..34f1ca3 100644 --- a/tools/gsf/green-code-skill/SKILL.md +++ b/tools/gsf/green-code-skill/SKILL.md @@ -126,7 +126,7 @@ Detailed pattern entries (problem, detection signals, fix approach, SCI term, ca |---|---|---| | `references/patterns-development.md` | 23 Development patterns | Code-level: data handling, media, web performance, async, compression, TLS termination | | `references/patterns-architecture.md` | 17 Architecture patterns | System-level: region selection, containers, serverless, statelessness, scaling | -| `references/patterns-ai-ml.md` | 9 AI/ML model and training patterns + 10 agentic AI workload patterns | Model selection, edge inference, agentic loops, LLM caching, retry budgets, planning depth | +| `references/patterns-ai-ml.md` | 10 AI/ML model and training patterns + 10 agentic AI workload patterns | Model selection, edge inference, agentic loops, LLM caching, retry budgets, planning depth | | `references/sci-formula.md` | SCI specification primer | Whenever the user asks "how does this reduce carbon?" or wants the math | | `references/trade-offs.md` | When NOT to apply a pattern | Always consult before applying anything aggressive | diff --git a/tools/gsf/green-code-skill/references/patterns-ai-ml.md b/tools/gsf/green-code-skill/references/patterns-ai-ml.md index cbf2734..09f130a 100644 --- a/tools/gsf/green-code-skill/references/patterns-ai-ml.md +++ b/tools/gsf/green-code-skill/references/patterns-ai-ml.md @@ -47,6 +47,15 @@ This file consolidates AI/ML patterns across two families: model and training pa - **Trade-off:** Shifts M to client devices. Name this explicitly. - **URL:** https://patterns.greensoftware.foundation/architecture/system-topology/energy-efficent-ai-edge +### Optimize power-performance for edge AI inference +- **Problem:** Edge AI deployment strategies optimize for accuracy and latency without treating power as a first-class metric, which drains batteries, triggers thermal throttling, and shortens device lifespan. +- **Detection signals:** Inference always run at peak frequency/precision regardless of battery level or temperature; no DVFS, thermal-aware throttling, or battery-tier policies; no linked energy/latency/accuracy measurement per model variant or power mode. +- **Fix:** Add power-aware workload scheduling (defer non-urgent inference, thermal-aware throttling, tiered battery-level policies, DVFS, placement across heterogeneous CPU/GPU/NPU/non-von-Neumann accelerators); expose a small set of named modes (Performance/Balanced/Power Saver) calibrated from the energy-latency-accuracy triplet rather than raw knobs; close the loop with fleet telemetry. +- **SCI:** Reduces E (operational energy, longer battery life) and M (fewer battery/device replacements from reduced thermal stress); can reduce I via time-shifting workloads to cleaner-grid or charging windows. +- **Trade-off:** Deferring or throttling inference adds latency; scheduling, mode calibration, and telemetry add engineering and testing complexity. Start with simple battery-tier policies before adaptive systems. +- **Pillar:** Energy efficiency, hardware efficiency. +- **URL:** https://patterns.greensoftware.foundation/operations/optimize-edge-ai-power-performance + ### Select a more energy-efficient AI/ML framework - **Problem:** Frameworks differ significantly in throughput per watt for the same model on the same hardware. - **Detection signals:** Eager-mode PyTorch in production inference, no consideration of optimized runtimes.