WSL2 routes GPU access through the Windows NVIDIA driver via passthrough, so installing a separate Linux CUDA toolkit inside WSL2 does not help and can actively break GPU detection. The fix is to ensure only the Windows host driver is installed and up to date.
You ran ollama ps, saw 100 percent CPU, and the obvious move feels like installing the NVIDIA CUDA toolkit inside your Ubuntu WSL2 distro. Do not. That single step is the most common reason Ollama will not use the GPU in WSL2, because WSL2 does not want a Linux GPU driver at all. It borrows the GPU from the Windows host through a passthrough layer, and a second driver installed inside Linux fights with it.
WSL2 exposes your NVIDIA GPU through the Windows host driver, not a Linux one. The fix is to install or update the Windows NVIDIA driver only, never a Linux driver or CUDA toolkit inside WSL2. The host driver must be reasonably current, then nvidia-smi should work inside WSL2 and Ollama will detect the card.
In WSL2 the GPU is shared from Windows. NVIDIA publishes the CUDA-enabled libraries straight into the distro at /usr/lib/wsl/lib/, mounted there automatically when a supported Windows driver is present. That is the whole supply chain: Windows driver up top, WSL libraries underneath, your app on top of those. Install a separate Linux NVIDIA driver or full CUDA toolkit inside the distro and you stack a second, conflicting driver onto a passthrough that was already complete, which is exactly when GPU detection stops working. If you have built a local AI box around this, the AI PC range at Evetech shows the kind of NVIDIA hardware this passthrough is designed for.
Work through these on the Windows side first, then check inside WSL2.
wsl -l -v and check the VERSION column reads 2. WSL1 has no GPU passthrough at all.nvidia-smi. If it lists your card, the passthrough is healthy.ollama ps while a model is loaded. You should now see GPU usage instead of 100 percent CPU.Order matters: get nvidia-smi working before you touch Ollama, because Ollama only finds the GPU once the layer beneath it is sound.
nvidia-smi inside WSL2 is your first signal, it proves the host driver and passthrough are talking. The second is ollama ps during inference: a GPU-accelerated model reports GPU rather than pure CPU, and tokens stream noticeably faster. If nvidia-smi works but Ollama still sits on CPU, restart the Ollama service so it re-scans for CUDA on a clean start. Local LLM work leans hard on the card, and the most popular PC builds list is a quick way to gauge what configurations local builders are running inference on.
No. The CUDA runtime libraries are mounted into WSL2 automatically once a supported Windows driver is installed. Adding a separate Linux CUDA toolkit is the step that usually breaks detection.
Ollama scans for CUDA at start-up. If the driver became healthy after Ollama launched, restart the service so it picks up the GPU on a fresh start.
Windows 11, or Windows 10 21H2 and later, with a WSL2-capable NVIDIA driver on the host. Older builds lack the passthrough plumbing entirely.
The driver rule is the same, but containerised Ollama also needs the NVIDIA container toolkit registered as the Docker runtime inside WSL2. That is a separate step on top of a working host driver.
Running local models well starts with the right NVIDIA card and a clean driver setup. Explore the AI PC range at Evetech to build a machine that drives Ollama on the GPU from day one.
No. The CUDA runtime libraries are mounted into WSL2 automatically once a supported Windows driver is installed. Adding a separate Linux CUDA toolkit is the step that usually breaks detection.
Ollama scans for CUDA at start-up. If the driver became healthy after Ollama launched, restart the service so it picks up the GPU on a fresh start.
Windows 11, or Windows 10 21H2 and later, with a WSL2-capable NVIDIA driver on the host. Older builds lack the passthrough plumbing entirely.
The driver rule is the same, but containerised Ollama also needs the NVIDIA container toolkit registered as the Docker runtime inside WSL2. That is a separate step on top of a working host driver.
Low FPS after an ultrawide monitor upgrade usually means the GPU is pushing more pixels. Lower resolution scale, tune graphics presets, check drivers, and retest the same game scene on your SA gaming setup.
Read more
AM5 is a smart long-term platform because Kimberley gamers get CPU upgrade room and DDR5 support. SA builders should still check motherboard value, BIOS support, cooler fit, and GPU budget before committing.
Read more
Ryzen 7 9800X3D interest comes from its gaming-first X3D cache, strong frame-rate reputation, and clear upgrade appeal. In SA, the real question is whether your GPU, monitor, board, and budget can make the CPU advantage visible.
Read more
Wi-Fi 7 routers make sense for SA fibre when your network has enough fast devices to use the upgrade. SA buyers should check fibre speed, router placement, mesh needs, gaming latency, and whether Wi-Fi 6 is still enough.
Read more
240Hz OLED panels are still expensive in SA because OLED tech and high refresh remain premium. Check panel size, warranty, burn-in care, GPU demands, and whether your games can reach 240 frames before judging value.
Read more
ITX builds are gaining popularity with Kimberley gamers because compact PCs save desk and room space. SA builders should compare case airflow, GPU clearance, PSU size, cooling noise, and upgrade limits before downsizing.
Read more
Tests reported audio capture and local-network scanning on a specific LG television with its screen off. LG says supported standby modes can keep voice and connected-device features active, so owners should review model-specific settings.
PCIe 5.0 and PCIe 4.0 differ most in lane support, compatibility, and upgrade value for video editing in 4k. SA buyers should match the choice to their actual hardware and games.
Real world difference between b850 b650 mmo live service games sits in stability, networking, storage, and upgrade room. SA players should compare LAN quality, USB layout, RAM support, and SSD slots for always-updated games.
Dead Pixel on Wi-Fi 7 Router is likely a display-path or source-title issue, so check the screen, cable, driver state, and related system symptoms before blaming the device itself.
Industry reports point to another roughly 10 percent Intel processor increase from 5 October, but Intel has not published a consumer price list confirming the scope. South African buyers should treat it as a risk signal, not a guaranteed Rand change.
Put the headphones in pairing mode, open Bluetooth and devices in Windows 11, then add the headset. If it fails, remove the old pairing and check charge, range and drivers.
Microsoft says upcoming Windows 11 work will target memory use on 8GB PCs and improve boot and sign-in responsiveness. There is no public release date or measured uplift yet.
Windows 11 does not show CPU temperature in its standard Settings pages. Use your CPU maker's supported utility or a trusted monitor, then read idle and loaded temperatures in context.
Low FPS after an ultrawide monitor upgrade usually means the GPU is pushing more pixels. Lower resolution scale, tune graphics presets, check drivers, and retest the same game scene on your SA gaming setup.
Gen 5 NVMe vs Gen 4 NVMe for MMO and Live-Service Games comes down to load behaviour, capacity, motherboard lanes, heat, and real game or workflow needs. SA buyers should match the choice to their setup instead of assuming one option always wins.
AM5 is a smart long-term platform because Kimberley gamers get CPU upgrade room and DDR5 support. SA builders should still check motherboard value, BIOS support, cooler fit, and GPU budget before committing.
Wi-Fi 6E vs Wi-Fi 6 for Streaming on Twitch and YouTube comes down to router support, latency, client devices, band support, and home network layout. SA buyers should match the choice to their setup instead of assuming one option always wins.