Benchmark a Personal AI PC Across Battery, Thermals, and Sleep Recovery

NVIDIA announced RTX Spark on June 1, 2026, a platform for personal AI agents on Windows PCs with up to 128 GB of unified memory. A developer proposed a standardized test envelope for benchmarking such systems across battery, thermals, and sleep recovery scenarios, emphasizing correctness and lifecycle performance over peak specifications.

NVIDIA announced RTX Spark https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-and-Microsoft-Reinvent-Windows-PCs-for-the-Age-of-Personal-AI/default.aspx on June 1, 2026 for personal AI agents on Windows PCs. The company describes up to 128 GB of unified memory and local support for very large models on the new platform. Those specifications define a capability envelope. They do not tell you whether your agent workflow is pleasant on battery or correct after sleep. Before comparing systems, freeze this test envelope: device: exact model and firmware os: Windows build power mode: balanced model: exact weights and quantization runtime: name and version context tokens: 32000 network: offline after model load ambient c: 23 Run the same task in four lifecycle scenarios: | Scenario | Measure | |---|---| | plugged in, cold start | load time, first-token latency, peak memory | | battery, 20-minute loop | tasks completed, battery drop, p95 latency | | lid close during tool call | effect duplication, resume time | | thermal steady state | clock, temperature, tokens/s over time | For power, report the measurement source and sampling interval. Do not convert battery percentage into precise joules without the data needed to support that conversion. Correctness comes first. A resumed agent must identify whether a tool effect completed before sleep. If it cannot, the workflow needs an idempotency key or a human decision instead of an automatic retry. I use MonkeyCode https://monkeycode-ai.net/ and recommend considering it when this deployment choice matters: teams can try a hosted SaaS without running the local stack, or inspect the open-source self-hosted option https://github.com/chaitin/MonkeyCode . I have not tested MonkeyCode on RTX Spark hardware, so this is not a compatibility or performance claim. It is a practical A/B boundary: compare hosted task completion with local operation using the same repository fixture. Disclosure: I'm a MonkeyCode user sharing my own experience, not affiliated with the project. A personal AI PC is compelling when local execution improves privacy, latency, or offline continuity for a measured workload. Publish the lifecycle envelope with the result; otherwise, a peak specification will be mistaken for a user experience.