SiliconPacer is a high-performance thermal regulator and GPU duty-cycler designed specifically for Apple Silicon (M-series) MacBooks.
When running sustained, high-parameter LLM workloads (via Ollama, LM Studio, oMLX, etc.) in High Power Mode, MacBooks often suffer from "thermal soak"—where the chassis becomes uncomfortably hot and the macOS kernel eventually forces aggressive, performance-crushing frequency drops.
SiliconPacer solves this by implementing proactive, rhythmic duty-cycling (PWM) to manage heat before it reaches critical thresholds.
Unlike reactive thermal management, SiliconPacer uses a Pulse-Width Modulation (PWM) approach. It intercepts the execution of targeted processes and applies a high-frequency "stutter":
- The Window: A fixed 55ms period.
- The Pulse: The process is paused (
SIGSTOP) for a precise number of milliseconds, then resumed (SIGCONT). - The Result: By creating micro-gaps in execution, it allows the silicon to shed heat rhythmically, preventing the massive thermal spikes that trigger macOS system-level throttling.
Platform Note: This utility leverages macOS-specific process control signals (
SIGSTOP/SIGCONT) and is optimized for Apple Silicon (M-series).
- High-Performance Engine: Uses low-level
libprocandsysctlC-bindings for near-zero overhead PID enumeration. - Context-Aware Overrides: Automatically adjusts throttling intensity based on the specific model being loaded (e.g., more cooling for 70B models).
- Zero-Dependency: Includes a custom TOML parser for compatibility with macOS System Python.
- QoS Clamping: Automatically runs under the
utilityQuality of Service class to ensure the pacer itself doesn't consume significant CPU. - oMLX Integration: A dedicated helper module allows LLM servers to "pause" the pacer during critical weight-loading phases.
- Emergency Recovery: Includes a dedicated rescue utility to unfreeze processes if the main script is terminated unexpectedly.
Note
Measurement Integrity: To prevent telemetry-induced heat, all data was captured using a custom-patched iSMC binary. By bypassing the macOS HID subsystem and pruning sensor arrays, we reduced polling latency by 88.5% (0.042s/snapshot) and maintained a negligible 0.2% CPU footprint, ensuring zero measurement artifacting.
SiliconPacer is engineered for maximum performance density with negligible system overhead. To provide empirical, reproducible proof of its real-world utility, strict 90-second sustained inference burns were conducted on an 14" Apple Silicon M5 SoC (32-core GPU) running omlx-server.
The system was evaluated against two distinct architectural paradigms: uniform heavy math (Dense) and sporadic, non-uniform memory/gate switching (Mixture of Experts). The core _pacer_loop function within SiliconPacer operates with a highly optimized 0.1% CPU utilization footprint. Logging was captured at 3.0s intervals using a standalone companion utility (soc_temp.py), which leverages an optimized iSMC binary to slash snapshot times down from 0.368s to 0.042s.
| Workload Profile | Framework Mode | Peak Core Avg | Peak Virtual Die G | Thermal Rise (ΔT) | Token Throughput |
|---|---|---|---|---|---|
| Dense 27B | High Power (Stock) | 75.685°C | 86.791°C | +47.344°C | 22.4 tok/s |
| Dense 27B | SiliconPacer (55ms) | 59.397°C | 68.353°C | +26.328°C | 13.6 tok/s |
| SiliconPacer Benefit | Paced Efficiency | -16.288°C | -18.438°C | Cuts Ramp 44% | Retains 61% |
| — | — | — | — | — | — |
| MoE 35B | High Power (Stock) | 81.492°C | 90.025°C | +48.453°C | 99.2 tok/s |
| MoE 35B | SiliconPacer (55ms) | 61.804°C | 69.338°C | +28.516°C | 68.3 tok/s |
| SiliconPacer Benefit | Paced Efficiency | -19.688°C | -20.687°C | Cuts Ramp 41% | Retains 69% |
For reference (Stock Low Power Mode): Operating system-level Low Power Mode restricts throughput to 11.3 tok/s on Dense 27B and 62.3 tok/s on MoE 35B.
For the full head-to-head timelines, raw clock-cycle step values, cooldown curves, and 50ms vs. 55ms period evaluations, explore the dedicated benchmark documents:
- Dense 27B Architecture Summary — Detailed analysis of the
AX-Qwen3.6-27Buniform matrix workloads. - Mixture of Experts 35B Summary — Detailed analysis of the
Qwen3.6-35B-Hereticdynamic routing loads. - Raw Telemetry Output Directory — Plain text logs (
.txt) straight from thesoc_tempcapture terminal windows.
We captured high-power mode vs. 5ms and 11ms pause times running the Gemma-4 26B model.
- Mixture of Experts 26B Summary — Detailed analysis of the
Gemma-4-26B-Hereticdynamic routing loads.
-
Clone the repository:
git clone https://github.com/yourusername/silicon-pacer.git cd silicon-pacer -
Configure your targets: Copy
config.tomlto~/.silicon_pacer.tomland edit the copy. This ensures your custom settings are preserved during git pull operations. -
Run SiliconPacer:
python3 silicon_pacer.py
Note: When SiliconPacer applies SIGSTOP, the targeted server moves to the background. Consequently, Ctrl+C in the original terminal will not stop the server; use kill %1 (or the relevant job ID) to terminate it.
| Command | Description |
|---|---|
python3 silicon_pacer.py |
Run with default configuration. |
python3 silicon_pacer.py "llama-server" |
Target a specific process by name. |
python3 silicon_pacer.py 20 |
Override global pause to 20 milliseconds. |
python3 silicon_pacer.py --dry-run |
See which PIDs would be targeted without stopping them. |
python3 silicon_rescue.py |
Emergency: Unfreeze processes if the pacer was killed with kill -9. |
When SiliconPacer is running, you'll see a status display showing the current throttling state. The terminal also appends real-time events, such as model load notifications:
% ./silicon_pacer.py
SiliconPacer v1.2.0
Press Ctrl+C to stop the throttle and restore normal process execution.
⚠️ Do NOT use 'kill -9' (SIGKILL) or processes will remain frozen.
Silicon Pacer
├─ Target PID(s): [41292 (oMLX Server)]
├─ State Model : AutomatosX--AX-Qwen3.6-27B-MLX-AXQ-4bit-MTP
├─ Fixed Window : 55.0ms
├─ Pause State : 17.6ms (32.0% throttle)
└─ Execution Window: 37.4ms (68.0% open)
Received SIGUSR1 to stop pacing momentarily.
Received SIGUSR2 to resume pacing → [Model loaded in 8.45s]
Silicon Pacer
├─ Target PID(s): [49710 (oMLX Server)]
├─ State Model : leonsarmiento--Ornith-1.0-35B-uncensored-heretic-5bit-XL-mlx
├─ Fixed Window : 55.0ms
├─ Pause State : 11.0ms (20.0% throttle)
└─ Execution Window: 44.0ms (80.0% open)
The configuration file allows you to define global defaults, model-specific overrides, and the specific processes to monitor.
[default]
pause_ms = 17.6 # 17.6ms pause per 55ms period (32% throttle)
[overrides]
"gemma-4-26B" = 11.0 # More aggressive cooling for MoE models
[[processes]]
name = "llama-server ollama"
desc = "Ollama Llama.cpp"
[[processes]]
name = "python omlx-server"
desc = "oMLX Server"To prevent SiliconPacer from interrupting the critical weight-loading phase in oMLX server, use the pacer_helper.py module.
Implementation Pattern:
from .pacer_helper import notify_pacer
async def _load_engine():
notify_pacer("pause", model_id="model-id") # Pause pulsing
try:
# ... weight loading logic ...
finally:
notify_pacer("resume") # Resume pulsingWarning
Do NOT use kill -9 (SIGKILL) to stop SiliconPacer.
Because SiliconPacer uses SIGSTOP to pause targets, killing the pacer violently will leave your AI models in a permanently frozen state. Always stop the pacer using Ctrl+C or kill <pid> (SIGTERM).
Warning
This utility manipulates process execution states at the OS kernel level. Use at your own risk. The authors assume no liability for hardware damage, thermal degradation, or system instability.
It is our sincere hope that one day, Apple will introduce a native system slider to macOS allowing users to manually set hard wattage limits or throttle ceilings for the SoC. Until the operating system provides that granular level of hardware control natively, SiliconPacer bridges the gap safely and responsibly.
Distributed under the MIT License. See LICENSE for more information.
- Mario Roy - Lead Developer
- AI Development Co-Pilots: Gemini/Gemma, Ornith, Qwen
While the code is written in Python and C, the stability of the system relies on three legendary entities that govern the intersection of logic and physics:
- The Grace Factor: The mathematical margin that prevents a micro-pause from becoming a system stutter. She is the reason the 55ms window feels like a rhythm rather than a glitch.
- Her Uncle Randomness: The chaotic entropy of the macOS scheduler and the unpredictable thermal transients of the M5 die. He is the unpredictable force we are constantly attempting to tame.
- Her friend Just Enough: The principle of parsimony. She ensures we do not over-engineer the solution, providing exactly the amount of cooling required—no more, no less.