NVIDIA RTX Tuning Toolkit for Linux

Scripts for tuning and overclocking NVIDIA RTX GPUs. Originally built and tested on an MSI RTX 5090 Vanguard; now also covers an RTX PRO 6000 Blackwell (Workstation Edition) after a card swap.

Every script that carries GPU-specific defaults (power limits, clock offsets) comes in two versions, picked by filename suffix:

Suffix	Card	Status
-5090	RTX 5090 (MSI Vanguard)	Validated — offsets/profiles below were tested on this card
-pro6000	RTX PRO 6000 Blackwell (Workstation Edition)	New — defaults are conservative starting points, not yet validated

Shared, model-agnostic files (install-driver.sh, nvidia-settings/install.sh, monitor.sh, nvidia-settings/config/10-coolbits.conf) have no suffix and work with either card.

GPU currently installed: RTX PRO 6000 Blackwell (Workstation Edition) · GB202 · GDDR7 96 GB ECC Run nvidia-smi to confirm your driver version — swapping the card doesn't change it automatically.

---

Quick Start (RTX PRO 6000 — currently installed)

If nvidia-smi already works and the card is up, you can skip the driver bootstrap and start at step 2.

# 1. Install the NVIDIA driver + CUDA toolkit first
sudo bash install-driver.sh

# 2. Run the one-time setup (requires sudo for xorg config + display restart)
bash nvidia-settings/install.sh

# 3. Set power limit (stock/default TGP: 600 W; board min: 150 W, confirmed via nvidia-smi)
./nvidia-settings/power-limit-pro6000.sh 500       # 500 W — reduced power profile
./nvidia-settings/power-limit-pro6000.sh default   # 600 W — stock TGP
./nvidia-settings/power-limit-pro6000.sh status    # show current limits (no sudo)

# 4. After your session restarts, apply a memory overclock — UNVALIDATED,
#    start low (see "Memory Offset" below for why this card gets a lower cap)
./nvidia-settings/oc-memory-pro6000.sh 250    # very conservative start
./nvidia-settings/oc-memory-pro6000.sh 500    # moderate — test thoroughly

# ECC is disabled by default on this install — check or toggle it:
./nvidia-settings/ecc-pro6000.sh status   # no sudo needed
./nvidia-settings/ecc-pro6000.sh on       # pending — needs a reboot to apply
./nvidia-settings/ecc-pro6000.sh off      # pending — needs a reboot to apply

# 5. Monitor in real time (watch ECC error counters, not just temps —
#    only meaningful while ECC is on)
./monitor.sh

# 6. Revert to stock
./nvidia-settings/oc-reset-pro6000.sh

# Wayland-friendly method: nvidia-tuner presets
nvidia-tuner --core-clock-offset -50 --memory-clock-offset 500 --power-limit 500

# Or reset clocks to stock and unlock max power with the bundled helper
./nvidia-tuner/apply-default-pro6000.sh        # core 0, mem 0, PL 600W

# Undervolt profile helpers (UNVALIDATED starting points — see caveat below)
./nvidia-tuner/apply-balanced-pro6000.sh       # core -50, mem +500, PL 500W
./nvidia-tuner/apply-efficient-pro6000.sh      # core -75, mem +500, PL 450W
./nvidia-tuner/apply-max-efficiency-pro6000.sh # core -75, mem +1000, PL 400W
./nvidia-tuner/apply-low-power-pro6000.sh      # core -50, mem +500, PL 300W

If Secure Boot is enabled, install-driver.sh will prompt for a MOK enrollment password and the key will be enrolled on the next reboot. If the driver is already installed and the GPU is working, install-driver.sh is optional unless you want the CUDA toolkit or need to refresh the Secure Boot enrollment path.

RTX PRO 6000 caveat: none of the offsets/profiles above have been validated on real hardware yet — they're conservative starting points, not "reported working" numbers. This card's memory is ECC-capable, but ECC is currently disabled on this install (confirmed via nvidia-smi --query-gpu=ecc.mode.current,ecc.mode.pending) — check with ./nvidia-settings/ecc-pro6000.sh status and toggle it with on/off if you want it back. With ECC off there's no error counter and no correction layer at all, so an unstable memory offset can silently corrupt compute results with zero warning — validate with real workload output (not just a crash-free stress test) before trusting a profile. The power-limit constants (150–600 W) were confirmed via nvidia-smi --query-gpu=power.min_limit,power.max_limit,power.default_limit on an actual Workstation Edition card, so apply-low-power-pro6000.sh's 300 W is safely within range. If you have a Max-Q (300 W hard cap) or Server edition (450–600 W depending on power cable) instead, run power-limit-pro6000.sh status and adjust the LIMIT_* constants in that script to match your card.

---

Quick Start (RTX 5090 — previous card, kept for reference)

Keep this around in case you reinstall a 5090, or as a template for tuning another 5090.

# 1. Install the NVIDIA driver + CUDA toolkit first
sudo bash install-driver.sh

# 2. Run the one-time setup (requires sudo for xorg config + display restart)
bash nvidia-settings/install.sh

# 3. Set power limit (daily default: 500 W; stock: 575 W; max: 600 W)
./nvidia-settings/power-limit-5090.sh 500       # 500 W — daily default profile
./nvidia-settings/power-limit-5090.sh default   # 575 W — stock TDP
./nvidia-settings/power-limit-5090.sh max       # 600 W — board maximum
./nvidia-settings/power-limit-5090.sh status    # show current limits (no sudo)

# 4. After your session restarts, apply a memory overclock (daily default: +2500)
./nvidia-settings/oc-memory-5090.sh 2500   # daily default (RTX 5090 MSI Vanguard)
./nvidia-settings/oc-memory-5090.sh 500    # conservative start
./nvidia-settings/oc-memory-5090.sh 1000   # moderate
./nvidia-settings/oc-memory-5090.sh 1500   # aggressive (test for stability first)
./nvidia-settings/oc-memory-5090.sh 3000   # extreme (reported working on RTX 5090 MSI Vanguard)

# 5. Monitor in real time
./monitor.sh

# 6. Revert to stock
./nvidia-settings/oc-reset-5090.sh

# Wayland-friendly method: nvidia-tuner presets
nvidia-tuner --core-clock-offset -125 --memory-clock-offset 2500 --power-limit 500
nvidia-tuner --memory-clock-offset 3000 --power-limit 500

# Or reset clocks to stock and unlock max power with the bundled helper
./nvidia-tuner/apply-default-5090.sh        # core 0, mem 0, PL 600W

# Undervolt profile helpers (increasing efficiency, all validated on this card)
./nvidia-tuner/apply-balanced-5090.sh       # core -75,  mem +2500, PL 500W
./nvidia-tuner/apply-efficient-5090.sh      # core -125, mem +2500, PL 450W
./nvidia-tuner/apply-max-efficiency-5090.sh # core -125, mem +3000, PL 400W

---

How It Works

Coolbits (xorg)

NVIDIA locks overclocking in the driver by default. Setting Coolbits=28 in /etc/X11/xorg.conf.d/10-coolbits.conf unlocks:

Bit	Value	Feature
2	4	Clock rate adjustment
3	8	GPU overclocking
4	16	Fan speed control

This config is generic — it's not tied to either GPU model, only to the PCI BusID of the slot the card is in (see Troubleshooting if BusID needs updating).

Memory offset (GPUMemoryTransferRateOffsetAllPerformanceLevels)

nvidia-settings exposes this attribute once Coolbits is set. The value is a transfer-rate offset in MHz applied on top of the stock GDDR7 frequency.

Important: this control is exposed only from an NVIDIA-managed Xorg session. A Wayland session that only provides Xwayland compatibility is not enough.

Both cards use the same 512-bit bus, 28 Gbps GDDR7 signalling (GB202 die), so stock nvidia-smi memory clock reads the same on both: 14001 MHz. What differs is validated headroom and risk profile:

RTX 5090 (oc-memory-5090.sh) — validated on this card:

Offset	Effective memory clock	Risk
0	14001 MHz (stock)	None
+500	~14501 MHz	Very low
+1000	~15001 MHz	Low
+1500	~15501 MHz	Medium — test thoroughly
+2000	~16001 MHz	High
+2500	~16501 MHz	Very high — validate with long stress tests
+3000	~17001 MHz	Extreme — may be unstable on many cards

Reported working on one sample: RTX 5090 MSI Vanguard at +2500 and +3000.

RTX PRO 6000 (oc-memory-pro6000.sh) — not yet validated:

Offset	Effective memory clock	Risk
0	14001 MHz (stock)	None
+250	~14251 MHz	Unknown — untested, very conservative start
+500	~14501 MHz	Unknown — untested
+750	~14751 MHz	Unknown — untested
+1000+	—	Blocked above +1500 by the script's safety cap until a stable point is established

This card has ECC-capable memory and denser, clamshell-mounted modules (96 GB vs. 32 GB), which typically means less OC headroom than a gaming card. ECC is currently disabled on this install (see ./nvidia-settings/ecc-pro6000.sh status) — with it off there's neither an error counter nor a correction layer, so an unstable offset risks silent data corruption in compute results rather than an obvious crash or visual artifact. If you re-enable ECC (ecc-pro6000.sh on, then reboot), corrected single-bit errors get counted instead of silently landing in your data — see the Stability Testing section below — but uncorrected multi-bit errors can still slip through unnoticed either way. Increase in small steps and validate against known-good workload output, not just uptime.

Note: GDDR7 uses PAM4 signalling. Small offsets can yield meaningfully higher bandwidth. If you see visual artifacts, GPU resets, incorrect compute results, or crashes → reduce the offset.

---

Power Limit

Max is 600 W on both cards via nvidia-smi -pl, but the stock/default value and board minimum differ (RTX 5090: 400–600 W; RTX PRO 6000: 150–600 W, confirmed live on the installed card):

Card	Command	Watts	Notes
RTX 5090	power-limit-5090.sh status	—	Read current limits (no sudo)
RTX 5090	power-limit-5090.sh default	575 W	Stock TDP, best performance/cooling balance
RTX 5090	power-limit-5090.sh max	600 W	+4% headroom over default (partner-card OEM headroom)
RTX 5090	power-limit-5090.sh 500	500 W	Daily default profile used on this card
RTX 5090	power-limit-5090.sh min	400 W	Board minimum
RTX PRO 6000	power-limit-pro6000.sh status	—	Read current limits (no sudo)
RTX PRO 6000	power-limit-pro6000.sh default	600 W	Stock TGP (Workstation Edition — no extra headroom above stock)
RTX PRO 6000	power-limit-pro6000.sh max	600 W	Same as default on this edition
RTX PRO 6000	power-limit-pro6000.sh min	150 W	Board minimum (confirmed via nvidia-smi power.min_limit on a Workstation Edition card)

Power limit changes take effect immediately and persist until the next driver unload or reboot. To make a limit persistent, add the relevant power-limit-*.sh call to a boot-time service (see below).

---

Stability Testing

After applying an offset, stress-test before committing to it:

# Simple VRAM bandwidth test (requires cuda-samples or pytorch)
python3 -c "
import torch, time
a = torch.randn(8192, 8192, device='cuda')
b = torch.randn(8192, 8192, device='cuda')
torch.cuda.synchronize()
t = time.time()
for _ in range(200):
    c = a @ b
torch.cuda.synchronize()
print(f'Matmul throughput test: {200/(time.time()-t):.1f} iter/s')
"

# Or use gpu-burn (install separately)
# gpu-burn 120   # 2-minute stress test

Watch ./monitor.sh in a second terminal for temperature spikes or ECC error counts climbing. On the RTX PRO 6000, a rising corrected-ECC count is a signal to back off the memory offset even if nothing has crashed — correction means the offset is already causing bit errors, just ones the memory can still fix.

This only works while ECC is enabled — it's currently disabled on this install, so monitor.sh will just show [N/A] for both counters and give you no early warning at all. Turn it on for OC testing sessions with ./nvidia-settings/ecc-pro6000.sh on (then reboot — the mode change is pending until the GPU resets), and back off with off afterward if you'd rather not keep the small VRAM/bandwidth overhead ECC carries day to day.

---

Making the Overclock Persistent on Boot

Create a systemd service that re-applies the offset after each login. Point ExecStart at the script suffix matching your card.

# /etc/systemd/system/gpu-oc.service
[Unit]
Description=GPU Memory Overclock
After=graphical.target

[Service]
Type=oneshot
Environment=DISPLAY=:0
ExecStart=/home/matt/development/nvidia-tuning-toolkit/nvidia-settings/oc-memory-pro6000.sh 250
RemainAfterExit=yes

[Install]
WantedBy=graphical.target

sudo systemctl enable --now gpu-oc.service

(For the RTX 5090, swap the ExecStart line for nvidia-settings/oc-memory-5090.sh 2500.)

---

Wayland-friendly method: nvidia-tuner

If you prefer nvidia-tuner (works well on Wayland), current daily-driver presets:

# RTX PRO 6000 (unvalidated starting point)
nvidia-tuner --core-clock-offset -50 --memory-clock-offset 500 --power-limit 500

# RTX 5090 (validated on this card)
nvidia-tuner --core-clock-offset -125 --memory-clock-offset 2500 --power-limit 500

Install nvidia-tuner

# Download latest release binary
curl -fL https://github.com/WickedLukas/nvidia-tuner/releases/latest/download/nvidia-tuner -o nvidia-tuner

# Install system-wide
sudo install -m 0755 nvidia-tuner /usr/local/bin/nvidia-tuner

# Verify
nvidia-tuner --help

On this machine nvidia-tuner actually ended up in ~/.local/bin (not /usr/local/bin) — check which nvidia-tuner before copying any systemd unit below, and fix the Environment=PATH=... / ExecStart= line if your location differs.

Make preset persistent after reboot (systemd)

Currently provided: RTX PRO 6000 low-power profile, enabled at boot. systemd/nvidia-tuner-low-power-pro6000.service in this repo runs nvidia-tuner/apply-low-power-pro6000.sh (core -50, memory +500, power 300 W) as a oneshot service on every boot, targeting multi-user.target so it doesn't wait on a display session. Install it with:

sudo cp systemd/nvidia-tuner-low-power-pro6000.service /etc/systemd/system/
sudo systemctl daemon-reload
sudo systemctl enable --now nvidia-tuner-low-power-pro6000.service

Then verify it actually applied (don't just trust "active (exited)" — confirm the clocks/power actually changed):

systemctl status nvidia-tuner-low-power-pro6000.service
sudo journalctl -u nvidia-tuner-low-power-pro6000.service -n 50 --no-pager
nvidia-smi --query-gpu=power.limit,clocks.current.graphics,clocks.current.memory --format=csv,noheader

Unverified in this session: installing/enabling this service requires sudo, which needs interactive authentication — something only you can do from a real terminal, not something run through this toolkit's automation. The ExecStart command itself also wasn't dry-run as root here for the same reason. It follows the exact pattern already used by nvidia-tuner services elsewhere in this repo (root-run oneshot, no manual sudo prefix needed since the service already runs privileged), but if systemctl status shows a failure after enabling, check the journal output above first — it will show whether nvidia-tuner itself rejected something (e.g. a stale PATH) rather than a driver/hardware issue.

To disable it later:

sudo systemctl disable --now nvidia-tuner-low-power-pro6000.service

Other profiles / a different card: to persist any other apply-*.sh script instead, copy the unit above, change ExecStart to point at the script you want (or inline the nvidia-tuner flags directly), and give it a distinct filename/Description so it doesn't collide with the low-power one:

[Unit]
Description=NVIDIA tuner preset
After=multi-user.target
Wants=multi-user.target

[Service]
Type=oneshot
ExecStart=/home/matt/development/nvidia-tuning-toolkit/nvidia-tuner/apply-balanced-pro6000.sh
RemainAfterExit=yes

[Install]
WantedBy=multi-user.target

Undervolt profiles (Linux / nvidia-tuner)

These profiles trade a deeper core undervolt and lower power limit for efficiency while keeping a memory overclock. Going down each table lowers the power limit and increases the memory offset.

RTX 5090 — all values validated on this card (MSI Vanguard); test for stability on your own hardware before relying on them:

Profile	Core offset	Memory offset	Power limit
Balanced	-75	+2500	500 W
Efficient	-125	+2500	450 W
Max efficiency	-125	+3000	400 W

./nvidia-tuner/apply-balanced-5090.sh
./nvidia-tuner/apply-efficient-5090.sh
./nvidia-tuner/apply-max-efficiency-5090.sh

RTX PRO 6000 — unvalidated starting points, deliberately more conservative than the 5090 profiles (no headroom data yet, ECC memory raises the cost of guessing wrong):

Profile	Core offset	Memory offset	Power limit
Balanced	-50	+500	500 W
Efficient	-75	+500	450 W
Max efficiency	-75	+1000	400 W
Low power	-50	+500	300 W

./nvidia-tuner/apply-balanced-pro6000.sh
./nvidia-tuner/apply-efficient-pro6000.sh
./nvidia-tuner/apply-max-efficiency-pro6000.sh
./nvidia-tuner/apply-low-power-pro6000.sh

If a profile is unstable on your card, lower the memory offset in -250 steps and re-test.

---

Files

nvidia-tuning-toolkit/
├── install-driver.sh                    # Driver + CUDA bootstrap (NVIDIA open driver + Secure Boot) — either card
├── monitor.sh                           # Live GPU stats (1s refresh) — either card, name detected automatically
├── nvidia-settings/
│   ├── install.sh                       # One-shot setup (Coolbits + permissions) — either card
│   ├── config/
│   │   └── 10-coolbits.conf             # xorg.conf.d snippet — unlock OC controls (shared, generic)
│   ├── power-limit-5090.sh              # RTX 5090: set GPU power limit (400–600 W)
│   ├── power-limit-pro6000.sh           # RTX PRO 6000: set GPU power limit (150–600 W)
│   ├── oc-memory-5090.sh                # RTX 5090: apply memory transfer-rate offset
│   ├── oc-memory-pro6000.sh             # RTX PRO 6000: apply memory transfer-rate offset (conservative cap)
│   ├── oc-reset-5090.sh                 # RTX 5090: reset offset to stock (0)
│   ├── oc-reset-pro6000.sh              # RTX PRO 6000: reset offset to stock (0)
│   └── ecc-pro6000.sh                   # RTX PRO 6000: check/toggle ECC mode (pending until GPU reset)
├── nvidia-tuner/
│   ├── apply-default-5090.sh            # RTX 5090: stock clocks, max power (0 core, 0 mem, 600 W)
│   ├── apply-default-pro6000.sh         # RTX PRO 6000: stock clocks, max power (0 core, 0 mem, 600 W)
│   ├── apply-balanced-5090.sh           # RTX 5090: -75 core, +2500 mem, 500 W (validated)
│   ├── apply-balanced-pro6000.sh        # RTX PRO 6000: -50 core, +500 mem, 500 W (unvalidated)
│   ├── apply-efficient-5090.sh          # RTX 5090: -125 core, +2500 mem, 450 W (validated)
│   ├── apply-efficient-pro6000.sh       # RTX PRO 6000: -75 core, +500 mem, 450 W (unvalidated)
│   ├── apply-max-efficiency-5090.sh     # RTX 5090: -125 core, +3000 mem, 400 W (validated)
│   ├── apply-max-efficiency-pro6000.sh  # RTX PRO 6000: -75 core, +1000 mem, 400 W (unvalidated)
│   └── apply-low-power-pro6000.sh       # RTX PRO 6000: -50 core, +500 mem, 300 W (unvalidated clock offsets; wattage confirmed in-range)
└── systemd/
    └── nvidia-tuner-low-power-pro6000.service  # Boot unit: runs apply-low-power-pro6000.sh at startup (see "Make preset persistent")

---

Troubleshooting

Symptom	Likely cause	Fix
No targets match 'gpu:0'	Wayland/Xwayland session or Coolbits not active	Log into an Xorg session, then restart display manager after installing config
Screen goes black / GPU reset	Offset too high	Run the matching oc-reset-*.sh from a TTY (Ctrl+Alt+F2)
nvidia-settings not found	Package missing	sudo apt install nvidia-settings
Offset applied but clock unchanged	Driver ignoring offset at idle	Run a GPU load first
Coolbits installed but OC controls still missing	Wrong BusID in 10-coolbits.conf	Run lspci | grep -i nvidia, convert the address to PCI:bus:device:function, and update the BusID line (or omit it on single-GPU systems)
Ran the wrong -5090 / -pro6000 script for your card	Script targets a different GPU's tuned defaults	Check nvidia-smi --query-gpu=name --format=csv,noheader and use the matching suffix; power-limit/OC values are model-specific
RTX PRO 6000: power-limit-pro6000.sh rejects a value you expected to work	Constants assume Workstation Edition (600 W); you may have Max-Q (300 W) or Server (450–600 W)	Run power-limit-pro6000.sh status and edit LIMIT_DEFAULT/LIMIT_MAX/LIMIT_MIN in the script to match your edition