GPU Not Detected, Slow, or Crashing

1. Symptom Lookup — Go to the Right Section

2. Quick Diagnostic Commands — Run These First

Run these on the affected system before contacting support. Paste the output into your support ticket.

Linux

nvidia-smi -q                          # full GPU status, temps, power, ECC

nvidia-smi --query-gpu=index,name,temperature.gpu,power.draw,clocks.sm,clocks.gr,pstate,ecc.errors.uncorrected.aggregate.total --format=csv

lspci | grep -i nvidia                 # confirm GPUs visible on PCIe bus

dmesg | grep -i 'nvidia\|nvrm\|xid'   # kernel GPU errors

cat /proc/driver/nvidia/gpus/*/information 2>/dev/null

Windows

nvidia-smi -q                          # PowerShell or CMD

Get-PnpDevice | Where-Object {$_.FriendlyName -like '*NVIDIA*'} | Select Name, Status

Get-EventLog -LogName System -Newest 200 | Where-Object {$_.Source -like '*nvlddmkm*'} | Format-List

3. GPU Not Detected (nvidia-smi returns no GPUs)

Step 1 — Confirm PCIe visibility first

If the GPU does not appear in lspci (Linux) or Device Manager (Windows), this is a hardware problem — driver reinstall will not help.

1. lspci | grep -i nvidia

If no output: the GPU is not seen by the CPU. Proceed to hardware checks below.

If output is present but driver says no GPU: skip to Section 7 (Windows driver) or Section 8 (Linux driver).

Step 2 — Reseat the GPU

1. Power off completely. Remove AC power cord and hold the power button 5 seconds to discharge capacitors.
2. Remove the GPU from its PCIe slot. Inspect the connector edge for dust, corrosion, or bent pins.
3. Reseat firmly until the slot latch clicks. On multi-GPU systems, try moving the card to a different PCIe slot.
4. Reconnect all PCIe power connectors (6-pin, 8-pin, or 16-pin). Confirm they are fully seated and not reversed.
5. Power on and recheck lspci.

Step 3 — Check PCIe power connectors and slot assignment

High-wattage GPUs (300 W+) require dedicated PCIe power cables — daisy-chained connectors from a single PSU rail are a common failure point.

• Verify cable routing: each GPU should ideally have its own PSU rail or cable run.
• For 4-GPU workstations or GPU servers, confirm the PSU rated wattage covers peak load: 4 × 300 W GPU + CPU + storage ≈ 1800 W minimum; use the next tier up for headroom.
• For Supermicro servers: confirm PCIe riser board seating and verify slot assignment in BIOS (Advanced > PCIe Configuration).

4. GPU Overheating, Fan Not Spinning, or Throttling

Identify throttle cause

nvidia-smi -q -d PERFORMANCE          # look for HW Slowdown, HW Thermal, HW Power Brake flags

watch -n 1 'nvidia-smi --query-gpu=temperature.gpu,power.draw,clocks.throttle_reasons.active,clocks.sm --format=csv,noheader'

Thermal thresholds (RTX A6000 / typical NVIDIA professional GPUs)

Cooling checklist

1. Verify airflow direction. Exxact 1U/2U GPU servers use front-to-back airflow. Reverse-airflow cards will overheat in standard chassis.
2. Check fan health via BMC/IPMI: ipmitool sdr type Fan — any fan reading 0 RPM is a failure.
3. Inspect for blocked airflow: blanking panels missing, cables draped over GPU heatsinks, inadequate rack spacing.
4. Check thermal paste on GPU heatsink if fan is spinning but temperatures are abnormally high (>85 °C at idle).
5. On tower/workstation systems with blower-style cards: ensure at least 1U of clearance between GPU exhaust and the next component.

Power cap as a thermal workaround

If hardware changes cannot be made immediately, a temporary power cap reduces heat and prevents shutdowns:

nvidia-smi -pl 220     # example: cap at 220W (adjust per GPU TDP)

nvidia-smi -q -d POWER | grep -i limit   # verify new limit applied

5. Severe GPU Slowdown on Intel Systems (IOMMU / VT-d)

Observed on Intel Xeon systems: deep learning workloads running at 1–3% GPU utilization. Benchmark shows 100–190× slowdown vs. AMD equivalents. Root cause: Intel VT-d (IOMMU) DMA translation overhead during concurrent CPU+GPU memory operations.

Diagnosis

dmesg | grep -i iommu

cat /proc/cmdline | grep iommu

If you see DMAR: IOMMU enabled in dmesg and no passthrough option in cmdline, IOMMU is likely the cause.

Fix — Try in order

Apply via GRUB (Ubuntu / RHEL):

sudo nano /etc/default/grub

GRUB_CMDLINE_LINUX_DEFAULT="quiet splash iommu=pt"

sudo update-grub && sudo reboot

Also check BIOS:

Advanced > CPU Configuration > Intel Virtualization Technology for Directed I/O (VT-d) → Disabled for bare-metal GPU compute workloads. Confirm with your system admin if virtualization (SR-IOV, KVM pass-through) is needed before disabling.

6. BIOS Update and Power Delivery

When to update BIOS

• System lockups under GPU load that cannot be explained by thermals or PSU capacity.
• PCIe link width unexpectedly downgrades (nvidia-smi showing PCIe Gen 1 when Gen 4 expected).
• BMC becomes unresponsive during GPU stress — may indicate BIOS-level power management bug.

Check current BIOS version

dmidecode -t bios | grep -E 'Version|Release'   # Linux

Get-WmiObject Win32_BIOS | Select-Object Name, Version, ReleaseDate  # Windows PowerShell

BIOS update by platform

7. NVIDIA Driver Reinstall — Windows

Use this procedure when: driver install fails, GPU shows 'Code 43' in Device Manager, or after a corrupted driver update pushed the system into repair mode.

Procedure

1. Download DDU (Display Driver Uninstaller) from guru3d.com and the target NVIDIA driver from nvidia.com. Store both on a USB drive.
2. Boot into Safe Mode: Settings > System > Recovery > Advanced Startup > Troubleshoot > Advanced Options > Startup Settings > Restart > press 4.
3. Run DDU: Select GPU type = 'GPU', Device = 'NVIDIA'. Choose 'Clean and restart in Safe Mode'. DDU will remove all NVIDIA components including audio, PhysX, and USB.
4. After DDU reboot (still in Safe Mode): verify Device Manager shows no NVIDIA devices.
5. Boot to normal Windows. Disconnect from the network. Run the NVIDIA installer as Administrator.
6. If install still fails: open Event Viewer (eventvwr.msc) > Windows Logs > System. Look for WindowsUpdateClient or SetupAPI errors in the 30 minutes before the failure.
7. If Windows update loop blocks driver install: try attaching a USB NIC — some NIC firmware versions block Windows authentication for update packages. Alternatively update BIOS (newer BIOS often includes updated NIC firmware).

Verify successful install

nvidia-smi                             # GPU should appear with correct driver version

Get-PnpDevice | Where {$_.FriendlyName -like '*NVIDIA*'} | Select Name, Status

8. NVIDIA Driver Reinstall — Linux

Use when: driver install fails, GPU disappears after kernel update, or nvidia.ko module fails to load.

Step 1 — Disable Secure Boot or sign the module

NVIDIA drivers are out-of-tree kernel modules. Secure Boot requires module signing.

• Recommended for pure GPU compute: disable Secure Boot in BIOS (Advanced > Security > Secure Boot).
• Alternative: enroll MOK key — see NVIDIA documentation for mokutil procedure.

Step 2 — Full driver purge

sudo systemctl stop gdm lightdm sddm 2>/dev/null; true

sudo apt-get purge --autoremove 'nvidia-*' 'cuda-*' 'libcuda*' -y    # Ubuntu/Debian

sudo dnf remove 'nvidia*' 'cuda*' -y                                 # RHEL/Rocky

sudo rm -f /etc/modprobe.d/blacklist-nouveau.conf

sudo update-initramfs -u    # Ubuntu   OR   sudo dracut -f    # RHEL

sudo reboot

Step 3 — Blacklist nouveau and install

After reboot, confirm nouveau is not loaded:

lsmod | grep nouveau    # should return empty

Install via package manager (recommended — handles DKMS automatically):

# Ubuntu — via graphics-drivers PPA

sudo add-apt-repository ppa:graphics-drivers/ppa -y

sudo apt-get update

sudo apt-get install nvidia-driver-550 -y    # substitute target version

sudo reboot

Or install from .run file (for specific versions):

sudo bash NVIDIA-Linux-x86_64-550.163.01.run --dkms

Step 4 — Verify

nvidia-smi

lsmod | grep nvidia    # nvidia, nvidia_modeset, nvidia_uvm should all be present

cat /proc/driver/nvidia/version

9. Xid 79 — GPU Reset Required (PCIe AER Error)

Xid 79 indicates the GPU detected a PCIe AER (Advanced Error Reporting) fault and initiated a reset. Possible causes: PCIe signal integrity, power delivery, or defective GPU.

Identify

dmesg | grep -i 'Xid\|NVRM\|nvrm' | tail -40

nvidia-smi --query-gpu=ecc.errors.uncorrected.aggregate.total --format=csv   # check for aggregate ECC errors

Triage steps

1. Check PCIe link width and generation under load — if downgrading, replace cable or switch to a different slot.
2. Reseat GPU and reconnect power. Xid 79 is frequently caused by marginal power delivery.
3. Update BIOS — BIOS updates often include PCIe error handling improvements.
4. If Xid 79 is intermittent and tied to a specific GPU: run DCGM diag (Section 10) and consider RMA if test fails.
5. Check VBIOS version: nvidia-smi --query-gpu=vbios_version --format=csv — outdated VBIOS can cause AER errors; contact Exxact for VBIOS update guidance.

10. DCGM Diagnostic (Linux Only)

NVIDIA's DCGM tool runs comprehensive hardware diagnostics including memory, compute, and bandwidth tests. Required for warranty RMA eligibility.

Install DCGM

# Ubuntu

sudo apt-get install datacenter-gpu-manager -y

# RHEL / Rocky

sudo dnf install datacenter-gpu-manager -y

dcgmi --version    # verify install

Run diagnostic levels

11. When to Escalate to Exxact Support

Contact Exxact Support

Appendix A — Key nvidia-smi Fields Explained

© 2026 Exxact Corporation. Internal use and customer-facing knowledge base article.