HP OMEN Ubuntu 26.04 AI + Gaming Build Book

Machine: HP OMEN MAX Gaming Laptop 16-ak0xxx
OS: Ubuntu 26.04.1 LTS (Resolute Raccoon)
Kernel observed: 7.0.0-31-generic
Purpose: Reproducible Linux workstation for CUDA/AI development, local LLMs, image generation, containers, and Windows/Linux gaming.

This is a record of the known-good configuration built and validated on this machine. It deliberately distinguishes required setup from optional tuning. After a future reinstall, package/model versions may have advanced; prefer current compatible versions unless reproducing this exact environment.

1. Known-good hardware and architecture

CPU/platform: AMD platform using amd-pstate-epp
Integrated GPU: AMD Radeon 890M (Strix)
Discrete GPU: NVIDIA GeForce RTX 5070 Ti Laptop GPU (Blackwell GB205M)
NVIDIA compute capability: 12.0
NVIDIA VRAM: ~12 GB nominal; applications see about 11.5 GiB
RAM: 32 GB
Storage: ~1 TB NVMe
Secure Boot: enabled
Graphics design: hybrid graphics
- AMD iGPU remains the normal desktop/battery GPU.
- NVIDIA dGPU is offloaded to for CUDA, AI, and games.
- Do not globally switch PRIME/MUX unless a specific problem requires it.

Displays

Internal AUO panel - Native: 1920x1200, 16:10 - Preferred/max refresh: 165 Hz - Alternate native timing: 60 Hz - Adaptive Sync advertised: 60-165 Hz, with an additional 48-60 Hz descriptor - Seamless dynamic video timing: supported - Maximum luminance advertised: 400 cd/m² - Connected as card2-eDP-1 to the AMD GPU

External Samsung C24F390 - Native: 1920x1080 @ 60 Hz - EDID range: 50-72 Hz - Connected as card1-HDMI-A-1 on the NVIDIA GPU - 1080p60 is the correct native HDMI mode; this is not a 120/144/165 Hz display.

2. Base operating-system checks

Update the system first:

sudo apt update
sudo apt full-upgrade

Check firmware:

fwupdmgr get-devices
fwupdmgr get-updates

Useful baseline commands:

uname -a
lsb_release -a
free -h
lsblk
lspci -nnk | grep -A3 -E 'VGA|3D|Display'
mokutil --sb-state

3. NVIDIA driver and hybrid graphics

The validated driver was the Ubuntu-recommended NVIDIA 595 open kernel driver (nvidia-driver-595-open), version 595.84.

Check Ubuntu's recommendation:

ubuntu-drivers devices

Install the recommended open driver if needed:

sudo apt install nvidia-driver-595-open
sudo reboot

Validate:

nvidia-smi

Known-good observations: - NVIDIA driver: 595.84 - nvidia-smi CUDA compatibility: 13.2 - Secure Boot works with the NVIDIA kernel driver.

PRIME OpenGL offload

Normal desktop rendering should remain on AMD:

glxinfo -B

Explicit NVIDIA offload:

__NV_PRIME_RENDER_OFFLOAD=1 \
__GLX_VENDOR_LIBRARY_NAME=nvidia \
glxinfo -B

The renderer should report NVIDIA / RTX 5070 Ti.

A simple live test:

__NV_PRIME_RENDER_OFFLOAD=1 \
__GLX_VENDOR_LIBRARY_NAME=nvidia \
glxgears

While it runs:

nvidia-smi

glxgears should appear as an NVIDIA process.

Vulkan

vulkaninfo --summary

The system can expose AMD, NVIDIA, and software (llvmpipe) devices. Seeing the other devices is not itself a problem.

For games using Vulkan/Proton, the useful NVIDIA selection variable is:

__VK_LAYER_NV_optimus=NVIDIA_only

4. Development workstation packages

Install the core toolchain:

sudo apt install \
  build-essential gcc g++ clang clangd cmake ninja-build \
  gdb lldb valgrind git git-lfs pkg-config \
  python3 python3-dev python3-venv \
  curl wget rsync unzip zip jq tree htop nvtop tmux ripgrep

Initialize Git LFS:

git lfs install

Validated toolchain included: - GCC 15.2.0 - CMake 4.2.3 - System Python 3.14.4

Create working directories:

mkdir -p ~/Projects/{cpp,python,ai}
mkdir -p ~/AI/{models,datasets,outputs}

Python policy

Use project virtual environments. Do not use sudo pip.

Install/use uv as the Python environment/package workflow. The validated installation used uv 0.12.10.

Example project:

mkdir -p ~/Projects/ai/example
cd ~/Projects/ai/example
uv venv
source .venv/bin/activate

5. CUDA Toolkit

The machine used NVIDIA's CUDA repository for Ubuntu 26.04 and CUDA Toolkit 13.3.

Important distinction:

nvidia-smi reports the driver's maximum CUDA compatibility.
nvcc --version reports the installed host CUDA Toolkit.
PyTorch may bundle yet another CUDA runtime.

These versions do not need to be identical.

Use NVIDIA's CUDA repository/keyring for the current Ubuntu release rather than Ubuntu's legacy nvidia-cuda-toolkit package.

After installing CUDA 13.3, ensure the toolkit binaries are on PATH, for example:

export PATH=/usr/local/cuda-13.3/bin:$PATH

Persist as appropriate in the shell configuration.

Validate:

nvcc --version

Known-good result: CUDA 13.3, nvcc V13.3.73.

Avoid: - Ubuntu nvidia-cuda-toolkit for this setup - CUDA .run installer - replacing the working Ubuntu NVIDIA driver unnecessarily

6. PyTorch GPU validation

Test project used:

~/Projects/ai/pytorch-test

Validated: - PyTorch 2.14.0+cu130 - NVIDIA RTX 5070 Ti detected - ~11.50 GiB VRAM visible - compute capability (12, 0) - large FP16 matrix multiplication succeeded.

Minimal validation:

import torch

print(torch.__version__)
print(torch.cuda.is_available())
print(torch.cuda.get_device_name(0))
print(torch.cuda.get_device_capability(0))
print(torch.cuda.get_device_properties(0).total_memory / 1024**3)

Remember that PyTorch's bundled CUDA runtime can differ from the host CUDA 13.3 toolkit.

7. Docker + NVIDIA Container Toolkit

Docker CE was installed from Docker's official Ubuntu repository, including Buildx and Compose.

Install/configure NVIDIA Container Toolkit using NVIDIA's repository, then configure Docker:

sudo nvidia-ctk runtime configure --runtime=docker
sudo systemctl restart docker

Validate Docker:

docker version
docker compose version

Validate GPU access with an NVIDIA CUDA container. The known-good test used:

nvidia/cuda:13.2.0-base-ubuntu24.04

For example:

docker run --rm --gpus all \
  nvidia/cuda:13.2.0-base-ubuntu24.04 \
  nvidia-smi

The RTX 5070 Ti should be visible inside the container.

Avoid: legacy nvidia-docker2 and Docker Desktop unless specifically needed.

8. Ollama

Ollama 0.33.3 was validated.

Check service:

systemctl status ollama --no-pager

Check NVIDIA usage while running a model:

watch -n 1 nvidia-smi

A tested model used roughly 3.2 GB VRAM and achieved approximately 95-104 tokens/s.

Model-size expectations for ~12 GB VRAM: - 3-4B: effortless - 7-8B: excellent - 12-14B quantized: very good - ~20B: constrained - 30B+: likely significant offload - 70B: impractical for GPU-resident use

Ollama's observed model storage was under:

/usr/share/ollama/.ollama/models/

9. llama.cpp

Repository:

~/Projects/ai/llama.cpp

Useful dependency:

sudo apt install libssl-dev

Clone/build:

cd ~/Projects/ai
git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp

cmake -S . -B build -G Ninja \
  -DGGML_CUDA=ON \
  -DCMAKE_BUILD_TYPE=Release

cmake --build build -j

The build automatically selected:

CMAKE_CUDA_ARCHITECTURES=120a-real

This is appropriate for the Blackwell GPU. Do not force architecture 120 manually unless there is a demonstrated reason.

NCCL warnings can be ignored for a single-GPU workstation.

Check devices:

./build/bin/llama-cli --list-devices

Known-good Qwen3 14B command

./build/bin/llama-cli \
  -hf Qwen/Qwen3-14B-GGUF:Q4_K_M \
  -ngl 99 \
  -fa on \
  -ctk q8_0 \
  -ctv q8_0 \
  -c 8192

Local GGUF:

./build/bin/llama-cli \
  -m ~/AI/models/llm/my-model.gguf \
  -ngl 99 \
  -fa on \
  -ctk q8_0 \
  -ctv q8_0 \
  -c 8192

Add -cnv for interactive conversation mode.

Benchmarks observed

Qwen3 8B Q4_K_M: - prompt processing pp512: ~2957 tok/s - generation tg128: ~64.8 tok/s

Qwen3 14B Q4_K_M: - pp512: ~1604 tok/s - tg128: ~36.7 tok/s

14B with Q8 KV, 8192 context: - pp512: ~1158 tok/s - tg128: ~29.6 tok/s

14B with Q8 KV, 16384 context: - pp512: ~812 tok/s - tg128: ~24.7 tok/s - VRAM remained below 10 GB in the observed run.

10. Shared AI model layout

A useful central model hierarchy is:

mkdir -p ~/AI/models/{llm,checkpoints,diffusion_models,text_encoders,vae,loras,controlnet,upscale_models}

This prevents each application from maintaining unnecessary duplicate model trees.

11. ComfyUI

Repository:

~/Projects/ai/ComfyUI

The validated environment used: - Python 3.13.15 in a uv virtual environment - PyTorch 2.14.0+cu130 - ComfyUI 0.34.0 - comfy-aimdo 0.5.2 - comfy-kitchen 0.2.33 - frontend 1.51.9 - templates 0.11.55

Start:

cd ~/Projects/ai/ComfyUI
source .venv/bin/activate
python main.py --enable-manager

Open:

http://127.0.0.1:8188

Known-good startup characteristics: - ~11774 MB VRAM - NORMAL_VRAM - cudaMallocAsync - asynchronous weight offloading - pinned memory - PyTorch attention - DynamicVRAM enabled - native NVFP4/FP8-related comfy-kitchen operations available.

Warnings about Manager falling back to uv, missing matrix-nio, optional OpenGL_accelerate, or torch JIT future behavior did not prevent operation.

Shared model paths

Create extra_model_paths.yaml with:

shared_models:
  base_path: /home/rmk/AI/models
  checkpoints: checkpoints
  diffusion_models: diffusion_models
  text_encoders: text_encoders
  vae: vae
  loras: loras
  controlnet: controlnet
  upscale_models: upscale_models

Adjust /home/rmk if restoring under a different username.

12. SDXL validation

Model:

~/AI/models/checkpoints/sd_xl_base_1.0.safetensors

The built-in SDXL workflow succeeded at 1024x1024.

Observed: - generation ~13.42 s - peak VRAM ~7134 MB - no CUDA/Blackwell compatibility problem.

13. FLUX.1 Dev

Separate-file workflow used:

~/AI/models/diffusion_models/flux1-dev.safetensors
~/AI/models/text_encoders/clip_l.safetensors
~/AI/models/text_encoders/t5xxl_fp8_e4m3fn_scaled.safetensors
~/AI/models/vae/ae.safetensors

The full BF16 FLUX model is much larger than GPU VRAM, but ComfyUI DynamicVRAM/offloading allowed it to run.

Observed BF16 run: - ~22.7 GB staged model data - 20 sampling steps ~46 s - total ~57.16 s - VRAM reached 100% - generation succeeded.

The warning:

clip missing: ['text_projection.weight']

did not prevent generation in the tested workflow.

14. FLUX NVFP4

Repository/environment:

~/Projects/ai/comfy-quants

A separate Python 3.13 uv environment was used. PyTorch needed to be installed into this environment:

uv pip install torch torchvision \
  --extra-index-url https://download.pytorch.org/whl/cu130

Quantization command:

comfy-quants export-model-nvfp4 \
  --config configs/flux_nvfp4.yaml \
  --source ~/AI/models/diffusion_models/flux1-dev.safetensors \
  --out ~/AI/models/diffusion_models/flux1-dev-nvfp4.safetensors \
  --device cuda:0 \
  --hash-output \
  --json

Successful ComfyUI loading reported quantization metadata and mixed-precision operations.

Observed NVFP4 run: - Flux staged ~6476 MB - 20 steps ~16 s - ~1.21 it/s - total ~24.56 s - VRAM remained below the BF16 saturation behavior.

Compared with BF16, observed total latency fell by roughly 57%, with sampling approaching 3x faster.

Practical default: use flux1-dev-nvfp4.safetensors; retain BF16 as a reference/fallback.

15. Gaming base

Install:

sudo apt install \
  steam-installer steam-devices \
  gamemode mangohud \
  nvidia-prime

Ensure Vulkan libraries for both 64-bit and 32-bit gaming are installed as required by Steam/Proton.

If Steam reports pactl: not found:

sudo apt install pulseaudio-utils

32-bit NVIDIA userspace

The matching 32-bit NVIDIA GL libraries are important for older/32-bit paths.

Check i386 architecture:

dpkg --print-foreign-architectures

If necessary:

sudo dpkg --add-architecture i386
sudo apt update

Validated matching packages included:

libnvidia-gl-595:amd64 595.84-0ubuntu0.26.04.1
libnvidia-gl-595:i386  595.84-0ubuntu0.26.04.1

Install the i386 side if missing:

sudo apt install libnvidia-gl-595:i386

16. GameMode

Validate:

gamemoded -t

The functional tests passed, except for the legacy CPU-governor expectation.

This system uses:

amd-pstate-epp

with the powersave governor and platform power profiles. This is normal. Do not force the old performance CPU governor globally merely to make the GameMode test green.

During a game:

gamemoded -s

Expected:

gamemode is active

17. MangoHud

Test:

mangohud glxgears

The Ubuntu build may warn that NVML/NVCTRL NVIDIA telemetry is unavailable. This does not mean NVIDIA rendering failed.

Useful overlay metrics: - FPS - frametime graph - GPU utilization - temperature - power - VRAM - GPU clock - CPU utilization/temperature - RAM

Interpretation: - high GPU utilization + low FPS -> GPU-limited - low GPU utilization + saturated CPU thread(s) -> CPU-limited - clocks constrained near power ceiling -> potentially power-limited - VRAM saturation/stutter -> memory pressure

18. Steam native Linux games

Known-good launch options for native OpenGL games:

__NV_PRIME_RENDER_OFFLOAD=1 __GLX_VENDOR_LIBRARY_NAME=nvidia gamemoderun mangohud %command%

Portal 2 native was validated: - portal2_linux appeared in nvidia-smi - ~816 MiB VRAM in the observed run - GameMode active - NVIDIA PRIME offload confirmed.

19. Steam Windows games with Valve Proton

Known-good launch options:

__NV_PRIME_RENDER_OFFLOAD=1 __VK_LAYER_NV_optimus=NVIDIA_only gamemoderun mangohud %command%

Portal 2 Windows/Proton was validated: - portal2.exe appeared as an NVIDIA C+G process - DXVK/Vulkan -> RTX 5070 Ti confirmed.

Keep Valve Proton as the default. Do not globally add random DXVK/VKD3D/Wine environment variables.

20. GE-Proton / ProtonUp-Qt

ProtonUp-Qt can be installed through Flatpak and used to install GE-Proton for Steam.

GE-Proton is best treated as a per-game compatibility alternative, not a universal replacement for Valve Proton.

Portal 2 Windows was also validated under GE-Proton, with portal2.exe appearing on the RTX 5070 Ti.

21. Heroic Games Launcher

Heroic was installed via Flatpak for Epic/GOG/Amazon and standalone Windows-game workflows.

For Windows games, use Heroic's Wine/GE runner management rather than manually sharing Steam's Proton runner.

Useful per-game environment variables:

__NV_PRIME_RENDER_OFFLOAD=1
__VK_LAYER_NV_optimus=NVIDIA_only

For OpenGL paths also use:

__GLX_VENDOR_LIBRARY_NAME=nvidia

Enable GameMode and MangoHud per game where useful.

A lightweight Windows ScummVM game was validated through Heroic and appeared as an NVIDIA C+G process.

Standalone Windows installer

For a setup.exe: 1. Add a game in Heroic. 2. Select a current Wine-GE/compatible runner. 3. Use Heroic's installer/run-installer facility to execute setup.exe inside the game's prefix. 4. Complete the Windows installer. 5. Change the game's executable to the installed game's real .exe. 6. Add the NVIDIA PRIME/Vulkan environment variables above. 7. Enable GameMode/MangoHud if desired.

Do not add Lutris merely for completeness; install another launcher only when a game actually needs it.

22. Power and performance configuration

Available profiles:

powerprofilesctl list

Validated:

low-power
balanced
performance

The ACPI platform profile exposes the same high-level choices:

cat /sys/firmware/acpi/platform_profile
cat /sys/firmware/acpi/platform_profile_choices

Known-good gaming setting while plugged in:

powerprofilesctl set performance

Normal everyday setting:

powerprofilesctl set balanced

Check:

powerprofilesctl get

NVIDIA laptop power behavior

Observed:

Default Power Limit: 80 W
Current Power Limit: 80 W (in detailed nvidia-smi output)
Maximum reported limit: 140 W

nvidia-powerd was active:

systemctl status nvidia-powerd --no-pager

The log contained:

Client (presumably SBIOS) has requested to disable Dynamic Boost DC controller

Interpret the 140 W value only as a reported upper programmable boundary, not as proof that this laptop should run at 140 W.

Do not: - force nvidia-smi -pl 140 - globally enable persistence mode - force an old CPU governor - disable the iGPU merely for benchmark numbers - override OEM power limits without verified model-specific support.

The stable known-good setup is preferable to speculative laptop power tuning.

23. Display diagnostics

List DRM connectors:

ls -l /sys/class/drm/

Show connection state:

for f in /sys/class/drm/card*-*/status; do
    echo "$f: $(cat "$f")"
done

Install EDID decoder:

sudo apt install edid-decode

Decode a connector:

edid-decode /sys/class/drm/card2-eDP-1/edid

External HDMI:

edid-decode /sys/class/drm/card1-HDMI-A-1/edid

Kernel-advertised modes:

cat /sys/class/drm/card2-eDP-1/modes
cat /sys/class/drm/card1-HDMI-A-1/modes

Internal panel gaming strategy

The internal panel is the preferred gaming display when high refresh matters: - 1920x1200 - 165 Hz - Adaptive Sync - use VRR when correctly exposed/enabled by GNOME - for a 165 Hz VRR setup, an FPS cap slightly below the ceiling (for example ~160-162 FPS) can be useful.

External Samsung strategy

The C24F390 is native 1080p60. For it: - target a stable 60 FPS - prioritize image quality over rendering unused 100+ FPS - DLSS Quality is generally preferable to aggressive upscaling at 1080p - frame generation is usually unnecessary when native/base performance already holds 60 FPS.

24. Graphics-quality starting points

For modern single-player games on the RTX 5070 Ti Laptop:

Textures: High/Ultra while VRAM permits
Geometry/detail: High/Ultra
Anisotropic filtering: 16x
Shadows: High as a starting point
Ray tracing: Medium/High, tune per game
DLSS at 1080p/1200p: start with Quality
DLSS Balanced: use when heavier RT needs more performance
DLSS Performance: generally reserve for heavier workloads/higher resolutions
Frame Generation: most useful when the base frame rate is already healthy; less desirable as a substitute for a very low base frame rate.

Always tune based on frametime and actual GPU/VRAM behavior rather than presets alone.

25. Post-update health check

After major Ubuntu kernel, NVIDIA-driver, CUDA, or graphics-stack updates, rerun these checks:

uname -r
nvidia-smi
nvcc --version
powerprofilesctl get
systemctl status nvidia-powerd --no-pager
vulkaninfo --summary

OpenGL PRIME:

__NV_PRIME_RENDER_OFFLOAD=1 \
__GLX_VENDOR_LIBRARY_NAME=nvidia \
glxinfo -B

GPU process test:

__NV_PRIME_RENDER_OFFLOAD=1 \
__GLX_VENDOR_LIBRARY_NAME=nvidia \
glxgears

Then in another terminal:

nvidia-smi

Docker GPU:

docker run --rm --gpus all \
  nvidia/cuda:13.2.0-base-ubuntu24.04 \
  nvidia-smi

GameMode:

gamemoded -t

AI: - run a small PyTorch CUDA test - run llama-cli --list-devices - run one known ComfyUI SDXL/NVFP4 workflow.

This catches regressions much faster than discovering them during real work.

26. Recommended clean-reinstall order

Use this order after a future clean Ubuntu installation:

1. Ubuntu 26.04 LTS
2. System updates + firmware
3. NVIDIA recommended open driver
4. Reboot and validate hybrid graphics
5. Development packages + Git LFS
6. Directory hierarchy
7. uv + Python project environments
8. NVIDIA CUDA repository + CUDA Toolkit
9. PyTorch CUDA validation
10. Docker CE
11. NVIDIA Container Toolkit + GPU-container validation
12. Ollama
13. llama.cpp CUDA build + benchmark
14. Shared AI model hierarchy
15. ComfyUI + shared model configuration
16. SDXL validation
17. FLUX BF16 validation
18. comfy-quants + FLUX NVFP4
19. Steam + GameMode + MangoHud
20. Matching NVIDIA i386 userspace libraries
21. Valve Proton validation
22. ProtonUp-Qt / GE-Proton if needed
23. Heroic
24. Native/Proton/Heroic NVIDIA offload tests
25. Display EDID / refresh / VRR checks
26. Set balanced/performance power policy
27. Final health-check suite

Do not copy old virtual environments across a clean OS install. Recreate them from dependency definitions or known commands. Large model files can be restored separately.

27. Known-good design decisions

These choices are intentional:

Area Decision

Desktop GPU AMD Radeon 890M AI/gaming GPU RTX 5070 Ti via offload Secure Boot Keep enabled NVIDIA driver Ubuntu open-kernel recommended driver CPU power amd-pstate-epp + platform profiles Gaming power powerprofilesctl set performance while plugged in Everyday power balanced NVIDIA laptop TGP Leave OEM/firmware control intact Python uv + per-project venvs System pip Never sudo pip CUDA NVIDIA repository/toolkit Containers Docker CE + NVIDIA Container Toolkit Local LLM Ollama and native CUDA llama.cpp Image generation ComfyUI FLUX default NVFP4; BF16 retained as fallback/reference Steam compatibility Valve Proton default Difficult Steam games GE-Proton per game Epic/GOG/etc. Heroic Additional launchers Only when a real game requires one

28. Quick command card

Gaming on NVIDIA

Native OpenGL:

__NV_PRIME_RENDER_OFFLOAD=1 __GLX_VENDOR_LIBRARY_NAME=nvidia gamemoderun mangohud %command%

Proton/Vulkan:

__NV_PRIME_RENDER_OFFLOAD=1 __VK_LAYER_NV_optimus=NVIDIA_only gamemoderun mangohud %command%

Power

powerprofilesctl set performance   # plugged-in gaming/heavy work
powerprofilesctl set balanced      # normal use

GPU

watch -n 1 nvidia-smi

CUDA

nvcc --version

ComfyUI

cd ~/Projects/ai/ComfyUI
source .venv/bin/activate
python main.py --enable-manager

llama.cpp

cd ~/Projects/ai/llama.cpp

./build/bin/llama-cli \
  -hf Qwen/Qwen3-14B-GGUF:Q4_K_M \
  -ngl 99 -fa on -ctk q8_0 -ctv q8_0 -c 8192

29. Final known-good state

At the end of this build:

Ubuntu is stable under Wayland.
Secure Boot remains enabled.
AMD drives the normal desktop/internal-panel path.
NVIDIA PRIME offload works.
CUDA workloads run on the RTX 5070 Ti.
PyTorch Blackwell CUDA computation is validated.
Docker GPU passthrough works.
Ollama GPU acceleration works.
llama.cpp builds natively for Blackwell and performs well.
ComfyUI runs SDXL and FLUX.
FLUX NVFP4 provides a major performance/memory improvement over the tested BF16 path.
Steam native Linux games use the RTX through PRIME.
Valve Proton games use the RTX.
GE-Proton games use the RTX.
Heroic Windows games use the RTX.
GameMode works.
MangoHud works, with a known Ubuntu packaging limitation around some NVIDIA telemetry.
The internal panel is 1920x1200 @ 165 Hz with Adaptive Sync capability.
The external Samsung C24F390 correctly operates at native 1920x1080 @ 60 Hz.
HP's standard Linux platform profile is already settable to performance.
NVIDIA laptop power remains under OEM/firmware control rather than unsafe manual overrides.

Principle: preserve the known-good hybrid architecture and change one component at a time. Validate after every driver, kernel, CUDA, or gaming-stack change.

HP OMEN Ubuntu 26.04 AI + Gaming Build Book

1. Known-good hardware and architecture

Displays

2. Base operating-system checks

3. NVIDIA driver and hybrid graphics

PRIME OpenGL offload

Vulkan

4. Development workstation packages

Python policy

5. CUDA Toolkit

6. PyTorch GPU validation

7. Docker + NVIDIA Container Toolkit

8. Ollama

9. llama.cpp

Known-good Qwen3 14B command

Benchmarks observed

10. Shared AI model layout

11. ComfyUI

Shared model paths

12. SDXL validation

13. FLUX.1 Dev

14. FLUX NVFP4

15. Gaming base

32-bit NVIDIA userspace

16. GameMode

17. MangoHud

18. Steam native Linux games

19. Steam Windows games with Valve Proton

20. GE-Proton / ProtonUp-Qt

21. Heroic Games Launcher

Standalone Windows installer

22. Power and performance configuration

NVIDIA laptop power behavior

23. Display diagnostics

Internal panel gaming strategy

External Samsung strategy

24. Graphics-quality starting points

25. Post-update health check

26. Recommended clean-reinstall order

27. Known-good design decisions

28. Quick command card

Gaming on NVIDIA

Power

GPU

CUDA

ComfyUI

llama.cpp

29. Final known-good state

More writing

Foundational papers in AI and ML

Simple things in life

Creating this website