NVIDIA Issues Hotfix for GPU Driver’s Overheating Issue

Yesterday NVIDIA rushed out a critical hotfix to contain the fallout from a previous driver release that had triggered alarm across AI and gaming communities by causing systems to falsely report secure GPU temperatures – whilst cooling demands quietly climbed toward potentially critical levels.

In NVIDIA’s official post across the hotfix release, though only third within the list of stated fixes, the difficulty is cited as ‘.

Shortly after the affected Game Ready driver 576.02 was rolled out, a pinned thread on the Stable Diffusion sub-Reddit, titled , became a resource for anecdotal issues and user-reported updates in regards to the latest driver. From these, and other reports around the online, some time-line of emergent problems could be established.

The primary Reddit report of the bug seems to have occurred late Friday afternoon UTC, on the ZephyrusG14 subreddit, where the user fricy81 cited a post at NVIDIA forums (archived):

The user at NVIDIA forums reported that after installing the driving force update, tools like MSI Afterburner and in-game monitors comparable to the one in (which generally access native system readings, much as Task Manager’s GPU panel does in Windows) stopped updating GPU temperature readings, freezing at around 35-36°C.

Restarting the monitoring software had no effect, the user stated, and only a full system reboot would restore accurate readings. Tools like HWInfo and NVIDIA’s own monitoring app continued to report temperatures accurately. The user emphasized that the difficulty occurred during normal use, not only after waking the system from sleep.

User feedback across various forums highlighted a general disruption of normal fan curve behavior and an alteration of core thermal regulation, leading to graphics processing units idling at unexpectedly high temperatures, and alarmingly overheating under what would typically be considered standard operational loads, as detailed on this comment:

Sub-Optimal

The official release PDF for the 576.02 driver update offers some clues about changes that will have contributed to the brand new issues. In section 5.5, NVIDIA acknowledges that GPU temperature could be reported incorrectly on NVIDIA systems, specifically showing zero degrees when no applications are running.

Source: https://us.download.nvidia.com/Windows/576.02/576.02-win11-win10-release-notes.pdf

The discharge states:

NVIDIA Optimus is a GPU switching technology that toggles between integrated and discrete graphics based on application demands, in an effort to robotically balance performance and power consumption, designed to conserve battery life and reduce power consumption. For tasks comparable to gaming or HD video playback, Optimus prompts the discrete GPU for higher performance; during lighter activities comparable to web browsing, it reverts to integrated (onboard) graphics.

The update appears to have prolonged a behavior previously limited to Optimus systems, allowing the affected GPU to enter a low-power state while idle, even when not hosted on an Optimus system, in turn disrupting temperature reporting in third-party tools.

Risk Adjustment

In most scenarios, it’s fair to say that the graphics card’s VBIOS would likely have prevented everlasting GPU damage. VBIOS enforces thermal and power limits on the firmware level, independently of the driving force.

Due to this fact even when a driver were to cause improper fan behavior or misreport temperatures, the VBIOS should still throttle performance, ramp up fan activity, or else shut down the GPU to stop hardware failure.

That doesn’t mean the chance was trivial – sustained high temperatures can degrade performance over time or stress adjoining components; moreover, absent a standard understanding that an updated driver caused an issue (not least in systems where drivers update ‘silently’), a difficulty of this nature could mislead a big proportion of affected users, who may attempt remedies for non-existent problems, and even potentially cause damage to their systems by applying non-relevant ‘fixes’.

The errant behavior brought on by update 576.02 was particularly alarming for those engaged in artificial intelligence workflows, where high-performance hardware is routinely pushed to its thermal limits for prolonged durations.

The problematic 576.02 driver inspired a broader rash of complaints after its release in mid-April, despite initial reports that it offered some helpful performance improvements. Notwithstanding the supply of the hotfix, and the extent of disruption that 576.02 seems to have caused, on the time of writing it stays available for download* at NVIDIA’s site.

Afterglow

When it comes to the fallout from the faulty update, there are many kinds of damage and or inconvenience reported: user Frankie_T9000 reported that his GPU crashed on boot resulting from heat buildup under the fault update, and only stabilized after undervolting. He commented ‘‘

Yesterday one other user in the identical thread stated:

Though NVIDIA (because it states persistently in each hotfix release) often provides hotfixes for particular video-games or platforms, the chance of warmth damage to or around a GPU is higher for AI practitioners than for videogamers, since intensive machine learning processes comparable to training or sustained inference place a GPU under consistent long-term load – an event prone to be triggered only periodically in a game, which can ‘spike’ into high usage for a boss-battle or a very demanding map section, but which is otherwise designed as a compromise between GPU exploitation and system stability.