{"slug": "overclocking5060ti", "title": "overclocking5060ti", "summary": "A developer published a headless GPU overclocking and power-tuning method for Linux that relies purely on NVML, avoiding X11/Xorg and nvidia-settings. Tested on an RTX 5060 Ti 16GB running MoE inference with Strata, the setup applied a +5300 MHz memory offset, +450 MHz GPC offset and a 182W power limit via a systemd service, reaching roughly 108–118 tok/s with an 85.1% expert cache hit rate.", "body_md": "Pure-NVML, completely headless GPU overclocking and power tuning for Linux without any X11/Xorg or `nvidia-settings` display server hacks.\n\nTested on **RTX 5060 Ti 16GB** hitting **~108–118 tok/s** in MoE inference using [Strata](https://github.com/Niko1221/Strata) by [@coldniko](https://x.com/coldniko).\n\n- **Model Context:** 30,000 tokens (thinking budget capped at 22,500)\n- **Reasoning / Thinking:** 66 → 91.5 tok/s\n- **Tool Calling / Generation:****107.9 tok/s sustained** (1996 tokens in 19s)\n- **Expert Cache Hit Rate:****85.1%** (only 1.4% routed over PCIe)\n- **Memory OC:****+5300 MHz** via`nvmlDeviceSetMemClkVfOffset`\n- **Core (GPC) OC:****+450 MHz** via`nvmlDeviceSetGpcClkVfOffset`\n- **Power Limit:** 182W (Persistence mode ON)\n\nRun the following bash script with `sudo` to set up the isolated Python venv, deploy all config files, and enable the systemd service:\n\n``` js\n#!/usr/bin/env bash\nset -euo pipefail\n\necho \"==> 1. Setting up Python venv with nvidia-ml-py...\"\nmkdir -p /opt/nvidia-nvml\npython3 -m venv /opt/nvidia-nvml\n/opt/nvidia-nvml/bin/pip install --upgrade pip\n/opt/nvidia-nvml/bin/pip install nvidia-ml-py\n\necho \"==> 2. Creating /etc/default/nvidia-llama-power...\"\ncat << 'EOF' > /etc/default/nvidia-llama-power\nPOWER_LIMIT=182\nMEMORY_OFFSET=5300\nGRAPHICS_OFFSET=450\nEOF\n\necho \"==> 3. Creating /usr/local/sbin/nvidia-llama-oc.py...\"\ncat << 'EOF' > /usr/local/sbin/nvidia-llama-oc.py\nimport os\nfrom pynvml import *\n\nmemory = int(os.environ[\"MEMORY_OFFSET\"])\ngraphics = int(os.environ[\"GRAPHICS_OFFSET\"])\n\nnvmlInit()\n\ntry:\n    h = nvmlDeviceGetHandleByIndex(0)\n\n    nvmlDeviceSetMemClkVfOffset(h, memory)\n    nvmlDeviceSetGpcClkVfOffset(h, graphics)\n\n    mem_actual = nvmlDeviceGetMemClkVfOffset(h)\n    gpc_actual = nvmlDeviceGetGpcClkVfOffset(h)\n\n    if mem_actual != memory:\n        raise RuntimeError(\n            f\"memory offset mismatch: requested={memory}, actual={mem_actual}\"\n        )\n\n    if gpc_actual != graphics:\n        raise RuntimeError(\n            f\"GPC offset mismatch: requested={graphics}, actual={gpc_actual}\"\n        )\n\n    print(f\"memory offset: {mem_actual} MHz\")\n    print(f\"GPC offset: {gpc_actual} MHz\")\nfinally:\n    nvmlShutdown()\nEOF\nchmod +x /usr/local/sbin/nvidia-llama-oc.py\n\necho \"==> 4. Creating /usr/local/sbin/nvidia-llama-power.sh...\"\ncat << 'EOF' > /usr/local/sbin/nvidia-llama-power.sh\n#!/bin/bash\nset -e\n\nsource /etc/default/nvidia-llama-power\nexport POWER_LIMIT MEMORY_OFFSET GRAPHICS_OFFSET\n\n/usr/bin/nvidia-smi -pm 1\n/usr/bin/nvidia-smi -pl \"$POWER_LIMIT\"\n/usr/bin/nvidia-smi -rgc\n\nexec /opt/nvidia-nvml/bin/python /usr/local/sbin/nvidia-llama-oc.py\nEOF\nchmod +x /usr/local/sbin/nvidia-llama-power.sh\n\necho \"==> 5. Creating /etc/systemd/system/nvidia-llama-power.service...\"\ncat << 'EOF' > /etc/systemd/system/nvidia-llama-power.service\n[Unit]\nDescription=Set NVIDIA GPU profile for llama.cpp / Strata\nAfter=multi-user.target\nConditionPathExists=/usr/bin/nvidia-smi\n\n[Service]\nType=oneshot\nExecStart=/usr/local/sbin/nvidia-llama-power.sh\nRemainAfterExit=yes\n\n[Install]\nWantedBy=multi-user.target\nEOF\n\necho \"==> 6. Enabling and starting systemd service...\"\nsystemctl daemon-reload\nsystemctl enable --now nvidia-llama-power.service\n\necho \"==> All done! Check status with: systemctl status nvidia-llama-power.service\"\n```\n\nCheck service status:\n\n```\nsystemctl status nvidia-llama-power.service\n```\n\nExpected output:\n\n```\n● nvidia-llama-power.service - Set NVIDIA GPU profile for llama.cpp / Strata\n     Loaded: loaded (/etc/systemd/system/nvidia-llama-power.service; enabled; preset: enabled)\n     Active: active (exited)\n...\nnvidia-llama-power.sh: Enabled persistence mode via daemon for GPU 00000000:23:00.0.\nnvidia-llama-power.sh: Power limit for GPU 00000000:23:00.0 was set to 182.00 W from 180.00 W.\nnvidia-llama-power.sh: memory offset: 5300 MHz\nnvidia-llama-power.sh: GPC offset: 450 MHz\n```\n\nTo change power limit or clock offsets, simply edit `/etc/default/nvidia-llama-power`:\n\n```\nsudo nano /etc/default/nvidia-llama-power\n```\n\nAnd re-apply immediately:\n\n```\nsudo systemctl restart nvidia-llama-power.service\n```\n\n", "url": "https://wpnews.pro/news/overclocking5060ti", "canonical_source": "https://gist.github.com/thisisspartan/b1d398bd4c3898cc99a96a44d5af1b4c", "published_at": "2026-11-02 12:47:45+00:00", "updated_at": "2026-11-03 00:50:06.522703+00:00", "lang": "en", "topics": ["ai-infrastructure", "ai-tools", "mlops", "developer-tools"], "entities": ["NVIDIA", "RTX 5060 Ti", "Strata", "NVML", "nvidia-smi", "coldniko"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/overclocking5060ti", "markdown": "https://wpnews.pro/news/overclocking5060ti.md", "text": "https://wpnews.pro/news/overclocking5060ti.txt", "jsonld": "https://wpnews.pro/news/overclocking5060ti.jsonld"}}