My gaming PC occasionally freezes, so far only while exiting Halo Campaign Evolved. The latest freeze ended in a KMODE_EXCEPTION_NOT_HANDLED (0x1E) BSOD. After restarting, the system entered BIOS because Windows Boot Manager was missing.
The Samsung 990 Pro 2TB was still detected by BIOS and the Windows installer, but DiskPart showed no partitions. From Fedora, TestDisk and gdisk reported that the main GPT header and partition table were invalid while the backup GPT was valid. The backup still contained the expected 100 MB EFI partition and the Windows partition. Rebuilding and writing the main GPT from the backup restored the drive and Windows booted normally.
This has happened more than once. The SSD passes Samsung's full diagnostic scan, has the latest firmware, and reports healthy SMART status. MemTest86 completed four passes both at 4800 MT/s and with 6000 MT/s XMP enabled without errors. BIOS and Intel ME firmware are updated, and DISM, SFC, Cinebench, OCCT 3D, and OCCT VRAM testing have not revealed a problem. BitLocker is disabled.
The crash dump showed two slow I/O timeout records from !blackboxntfs, including one on the same thread that crashed, although Event Viewer has no storNVMe, storport, or disk errors. The system uses an MSI B760 Gaming Plus WiFi, Intel i7-14700K, RTX 4080 Super, and the 990 Pro is currently installed in the CPU-connected M2_1 slot. I am planning to test it in the chipset-connected M2_2 slot to see whether the PCIe path makes a difference.
What could cause intermittent NVMe hangs or timeouts that corrupt only the primary GPT while leaving the backup intact, even when SMART and drive diagnostics look normal?
2 Answers
The 100 MB EFI partition is smaller than current Windows recommendations. A 260–300 MB EFI partition is generally safer for future boot files and updates, especially on systems using native 4K sectors. It probably does not explain the GPT header corruption by itself, but rebuilding the installation with a larger EFI partition would eliminate one possible boot-maintenance issue.
Since BitLocker is disabled, encryption is unlikely to be involved. I would also check whether the crashes are tied to a particular game, graphics or chipset driver, and whether Windows logs any WHEA PCIe errors even if the usual disk and storNVMe events are absent.
The slow I/O timeout in the crash data is the most useful clue here. A healthy SMART report and a passing vendor scan do not rule out an intermittent controller, PCIe link, motherboard slot, power, or driver problem. The crash may have interrupted a write to the primary GPT while the backup copy remained untouched, rather than the SSD having bad flash cells.
Moving the drive to M2_2 is a reasonable isolation test. Also compare behavior with the chipset and CPU PCIe settings at default, check Windows and BIOS storage-related settings, and inspect the Windows dump with WinDbg for the exact failing module and storage request. Keep current backups and avoid repeatedly rewriting the partition table until the underlying issue is identified.
That matches what I found in the dump: !blackboxntfs recorded two slow I/O timeouts, one on the same thread that eventually crashed. I am going to test the other M.2 slot and continue checking the PCIe and storage settings.