145 lines
5.2 KiB
Markdown
145 lines
5.2 KiB
Markdown
# Tips and Tricks
|
||
|
||
Practical tips, workflows, and techniques for getting the most out of NVCurve.
|
||
|
||
## Understanding the V/F Curve
|
||
|
||
The voltage-frequency curve defines how your GPU clocks at different voltage levels. Each point on the curve represents a (frequency, voltage) pair. NVCurve lets you add a frequency offset to each point, effectively shifting where your GPU operates.
|
||
|
||
- **Positive offsets** increase clock speed at a given voltage (overclocking).
|
||
- **Negative offsets** reduce clock speed at a given voltage, which can enable undervolting (lower voltage for the same performance, or same voltage at lower clocks for efficiency).
|
||
|
||
## Start Small, Test Often
|
||
|
||
When experimenting with offsets:
|
||
|
||
1. Use `--dry-run` to preview changes before applying them:
|
||
```bash
|
||
nvcurve write --global --delta 25 --dry-run
|
||
```
|
||
2. Apply small increments (10–25 MHz) and test stability between each step.
|
||
3. Use snapshots to roll back: `nvcurve snapshot restore`.
|
||
|
||
## The Global Offset Shortcut
|
||
|
||
If all active points share the same delta, the **Global Offset** slider appears in the web UI toolbar. This is equivalent to `nvcurve write --global` but gives you interactive control. Drag the slider to stage a uniform offset across every point, then click Apply.
|
||
|
||
## Curve Flattening for Efficiency
|
||
|
||
Curve flattening is a powerful technique for efficiency-oriented tuning:
|
||
|
||
1. Select a group of points in the upper voltage range.
|
||
2. Click **Flatten to [anchor]** in the toolbar.
|
||
3. This sets each selected point to land on the same effective frequency as your anchor point.
|
||
|
||
The result is a "step" in your curve where multiple voltage points map to the same clock — useful for finding the sweet spot where your GPU delivers peak frequency with minimal voltage.
|
||
|
||
## Keyboard-Driven Workflow
|
||
|
||
For precise tuning, the keyboard shortcuts in the curve editor are your friend:
|
||
|
||
- **Arrow keys** nudge selected point(s) by ±1 MHz.
|
||
- **Ctrl/Cmd + arrow keys** nudge by ±10 MHz.
|
||
- **Tab / Shift+Tab** steps selection through points one by one.
|
||
- **Enter** opens inline input for exact values.
|
||
|
||
This lets you make surgical adjustments without reaching for the mouse.
|
||
|
||
## Profile Strategy
|
||
|
||
A good profile setup covers your common use cases:
|
||
|
||
| Profile | Purpose |
|
||
|---|---|
|
||
| `gaming` | Aggressive positive offsets for maximum performance |
|
||
| `balanced` | Mild offsets for a good performance/temperature tradeoff |
|
||
| `efficient` | Negative offsets or flattened curve for low power usage |
|
||
| `stock` | Zero offsets (baseline for comparison) |
|
||
|
||
Set your preferred profile as the default so it auto-applies on boot:
|
||
|
||
```bash
|
||
nvcurve profile default gaming
|
||
```
|
||
|
||
## LACT Conflict
|
||
|
||
If you use LACT (Linux Auto-Clock Tuner) or similar tools, they will conflict with NVCurve because both write to the same hardware registers.
|
||
|
||
**Before using NVCurve with a default profile:**
|
||
|
||
```bash
|
||
sudo systemctl disable --now lactd
|
||
```
|
||
|
||
Otherwise, `lactd` will apply its own curve on startup and overwrite NVCurve's auto-applied profile.
|
||
|
||
## Multi-GPU Considerations
|
||
|
||
In multi-GPU setups:
|
||
|
||
- Each GPU maintains its own isolated state (write lock, active profile, monitoring).
|
||
- Use `--gpu N` to target a specific GPU from the CLI.
|
||
- In the web UI, switch GPUs using the dropdown in the status bar.
|
||
- Set per-GPU default profiles:
|
||
```bash
|
||
nvcurve --gpu 0 profile default gaming
|
||
nvcurve --gpu 1 profile default efficient
|
||
```
|
||
|
||
## Diagnostics Before Troubleshooting
|
||
|
||
When something isn't working right, run diagnostics first:
|
||
|
||
```bash
|
||
nvcurve read --diag
|
||
```
|
||
|
||
This shows:
|
||
- GPU name and driver version
|
||
- VRAM totals
|
||
- All NvAPI function probe results
|
||
- Current clock offsets and memory offset ranges
|
||
- Power limits
|
||
|
||
Use `nvcurve inspect` to examine raw ClockBoostTable fields for specific points:
|
||
|
||
```bash
|
||
nvcurve inspect --point 80
|
||
nvcurve inspect --range 78-82
|
||
```
|
||
|
||
## Headless / Scripting Workflow
|
||
|
||
For headless systems or automation, the CLI is fully self-contained:
|
||
|
||
```bash
|
||
#!/bin/bash
|
||
# Example: Apply a profile, run a benchmark, then restore
|
||
|
||
nvcurve profile apply gaming
|
||
./my-benchmark.sh
|
||
nvcurve snapshot restore
|
||
```
|
||
|
||
All CLI commands that interact with hardware automatically escalate to root via `sudo`. The `--json` flag on `nvcurve read` makes it easy to parse output in scripts.
|
||
|
||
## Safety Limits
|
||
|
||
NVCurve enforces safety limits to prevent damage:
|
||
|
||
- **Max delta cap**: ±3000 MHz hard limit per point.
|
||
- **Auto-snapshot**: A snapshot is saved before every write (configurable).
|
||
- **Negative frequency warnings**: The tool warns if an offset would result in a negative effective frequency.
|
||
|
||
These safeguards are in place, but remember — you're writing to undocumented hardware registers. Always test stability after applying changes and monitor temperatures.
|
||
|
||
## Background Daemon vs. Web Server
|
||
|
||
NVCurve has two running components to understand:
|
||
|
||
- **Daemon** (`nvcurve daemon`) — Lightweight Unix socket daemon that handles auto-loading profiles on boot. Managed via `nvcurve service`.
|
||
- **Web server** (`nvcurve serve`) — FastAPI REST + WebSocket server for the web UI. Starts on demand.
|
||
|
||
The systemd service (`nvcurve service install`) manages the daemon. The web server is optional and starts separately. Use `--auto-serve` at service install time if you want the web server to auto-start on boot as well.
|