Summary
Nine BSODs (seven 0x3B SYSTEM_SERVICE_EXCEPTION, two 0x1E KMODE_EXCEPTION_NOT_HANDLED) over five days of sustained CUDA training on a Dell Pro Max 18 Plus laptop with RTX PRO 5000 Blackwell 24GB GDDR7. All crashes share exception code 0xc0000005 (STATUS_ACCESS_VIOLATION) at different faulting addresses in ntoskrnl address space. Crashes occur on both driver 595.79 (R595 U2) and 595.97 (R595 U3). The same PyTorch RL training workloads run indefinitely without issues on NVIDIA DGX Spark GB10 GPUs (sm_121).
System Configuration
-
Laptop: Dell Pro Max 18 Plus
-
GPU: NVIDIA RTX PRO 5000 Blackwell Generation Laptop GPU, 24GB GDDR7 (sm_120)
-
CPU: Intel Core Ultra (with Intel Arc integrated graphics and Intel AI Boost NPU)
-
OS: Windows 11 Pro
-
Drivers tested: R595 U2 (595.79), R595 U3 (595.97) — clean install via Custom > “Perform a clean installation”
-
CUDA Version: 13.2
-
Framework: PyTorch (RL training, sustained GPU compute at 75–100% utilization)
Reproduction Steps
-
Run a PyTorch reinforcement learning training workload that sustains 75–100% GPU utilization over hours
-
System will BSOD with 0x3B or 0x1E, exception code 0xc0000005
-
Crash latency varies: minutes under active user work, hours when system is unattended
-
Network state changes (WiFi switch, mobile hotspot activation) are a repeated immediate trigger
Crash Data
Nine crashes total across Mar 23–27, 2026:
| # | Date/Time | Stop Code | Driver | Faulting Address | Minidump | Trigger |
|---|---|---|---|---|---|---|
| 1 | Mar 23, 07:06 AM | 0x3B | 595.79 | — | 032326-30281-01.dmp | Active user work |
| 2 | Mar 23, 07:15 AM | 0x3B | 595.79 | — | 032326-31609-01.dmp | Active user work |
| 3 | Mar 23, 11:31 PM | 0x3B | 595.79 | 0xfffff80262c5effd | 032326-31390-01.dmp | Unattended |
| 4 | Mar 24, 05:41 AM | 0x3B | 595.79 | 0xfffff8077d6c364c | 032426-30828-01.dmp | Unattended |
| 5 | Mar 26, 2:09 PM | 0x3B | 595.79 | 0xfffff805170b2c23 | 032626-25250-01.dmp | Active user work |
| 6 | Mar 27, 10:34 AM | 0x1E | 595.79 | — | 032726-31203-01.dmp | WiFi switch + screen lock |
| 7 | Mar 27, 2:29 PM | 0x1E | 595.79 | — | 032726-30718-01.dmp | None (user away >1 hour) |
| 8 | Mar 27, ~4:30 PM | 0x3B | 595.97 | 0xfffff802e88c3657 | 032726-31500-01.dmp | Active user work |
| 9 | Mar 27, ~5:03 PM | 0x3B | 595.97 | 0xfffff80559902c23 | 032726-30343-01.dmp | Phone hotspot activation |
All nine share exception code 0xc0000005 at different faulting addresses, indicating a race condition or memory corruption rather than a single deterministic fault.
Key Findings
1. nvidia-smi confirms Disp.A = On — dGPU handles display compositing alongside CUDA compute. dwm.exe and explorer.exe run on the RTX PRO 5000 simultaneously with CUDA training. The internal display is physically wired through the dGPU (TCC mode fails: “TCC can’t be enabled for device with active display”). NVIDIA Control Panel “Preferred graphics processor = Integrated graphics” does not affect Disp.A status. BIOS “Hybrid Graphics / Advanced Optimus” and “Discrete Direct Output Mode” options do not provide iGPU-only display routing.
2. Minidump analysis (crash #5) identified nvlddmkm.sys and npu_kmd.sys involvement.
-
nvlddmkm.sysloaded at crash time -
npu_kmd.sys(Intel NPU driver, build 32.0.100.4512) logged bugcheck callbacks with TDR references and all-zero fence IDs on the Compute node -
Faulting RIP in ntoskrnl address space — a kernel-mode driver passed a bad pointer
3. Intel NPU disablement was a partial mitigation. Disabling the Intel AI Boost NPU in Device Manager eliminated fast crashes under active use (a 5-hour training run with 3 hours of active work succeeded — previously crashed within minutes). However, crashes continued (crashes #6–9), proving the NVIDIA driver is unstable independently.
4. Thermals are not a factor. GPU temperatures across all crashes ranged from 56–78°C, well below the 83°C throttle threshold for RTX PRO 5000.
5. Identical workloads run stably on DGX Spark. The same PyTorch RL training code runs for days without issues on NVIDIA DGX Spark GB10 GPUs (sm_121, 128GB unified memory), confirming the workload is not inherently problematic.
Mitigations Attempted (None Resolved)
| Mitigation | Result |
|---|---|
| BIOS: Disabled Adaptive C-States for Discrete Graphics | Crashes continued |
| Dell Optimizer: Ultra Performance thermal mode | Crashes continued |
| Windows: Best Performance power mode, PCI Express Link State PM off | Crashes continued |
| Intel NPU disabled in Device Manager | Partial — eliminated fast crashes under active use, but crashes continued under sustained load |
| Display compositing routed to iGPU (NVIDIA Control Panel) | Never took effect — Disp.A = On persisted |
| Screen/standby timeouts disabled (powercfg) | Crashes continued |
TCC mode (nvidia-smi -dm 1) |
Failed — “TCC can’t be enabled for device with active display” |
| Driver update: 595.79 → 595.97 clean install | Two 0x3B crashes within hours of install |
Minidumps
All nine minidumps are available in C:\Windows\Minidump\. I can provide them upon request or via file sharing.
Expected Behavior
Sustained CUDA compute workloads should not cause kernel-mode access violations, particularly on a professional workstation GPU designed for extended compute tasks.
Request
This appears to be a driver-level instability in nvlddmkm.sys specific to the RTX PRO 5000 Blackwell laptop GPU under concurrent WDDM display compositing + CUDA compute. The display cannot be decoupled from the dGPU on this hardware (TCC blocked, no BIOS MUX switch). A driver fix that stabilizes the WDDM kernel-mode path during sustained CUDA workloads would resolve this issue.