According to reports, overseas user Ok-Dimension-741 recently posted that their RTX 4090's GDDR6X VRAM experienced uncorrectable errors, reducing available VRAM from 24GB to approximately 22.5GB. The manufacturer received the card three times but only reapplied thermal paste before returning it, refusing to replace the VRAM or provide a new unit.
Windows Event Viewer recorded uncorrectable GDDR memory errors in three separate partitions, while OCCT's VRAM test was even more extreme, reporting 4.7 million errors within 15 seconds.
nvidia-smi's row remapping diagnostics show that this RTX 4090 has accumulated 33 uncorrectable errors, but the number of remapped rows is 0, with 192 spare banks still remaining.
GDDR6X VRAM has built-in ECC (Error Correction Code) mechanisms that can automatically detect and fix single-bit errors, marked as "correctable errors." When VRAM chips suffer from physical aging or manufacturing defects, causing multiple bits to flip simultaneously beyond ECC's correction capability, "uncorrectable errors" occur, and the corresponding data cannot be repaired.
The GPU features a row remapping mechanism as a last line of defense: memory chips come with a portion of spare banks reserved from the factory. When a specific row of memory cells repeatedly fails, the GPU maps the data from that row to a spare bank, shielding the damaged area.
This RTX 4090 still has 192 spare banks remaining, indicating ample resources. However, the number of remapped rows is 0, which means the row remapping mechanism did not respond to these 33 uncorrectable errors.
Possible reasons include these errors being scattered across different physical locations, not yet reaching the threshold to trigger remapping, or the error locations falling outside the address range eligible for remapping.
Regardless of the scenario, 33 uncorrectable errors indicate physical damage to the VRAM chips, which will continue to worsen over time.
The main reason is that the manufacturer's testing method involves running FurMark stress tests. FurMark primarily stresses the GPU core and cooling system without triggering VRAM errors, so each test reported that the "GPU started and ran normally."
The user requested that the manufacturer check the row remapping data during the RMA process, but all three return-to-factory attempts were rejected. The manufacturer simply reapplied thermal paste and sent the card back as is.
The warranty for this RTX 4090 expires this November, leaving the user with little time to assert their rights. With GPU prices continuing to rise, buying a new card directly is not realistic. The user has filed a complaint with consumer protection agencies, requesting only that the manufacturer acknowledge the 33 uncorrectable errors recorded in the row remapping data and provide a replacement based on this evidence.