Yesterday NVIDIA rushed out a important hotfix to include the fallout from a previous driver launch that had triggered alarm throughout AI and gaming communities by inflicting techniques to falsely report secure GPU temperatures – whilst cooling calls for quietly climbed towards doubtlessly important ranges.
In NVIDIA’s official put up across the hotfix launch, although solely third within the record of acknowledged fixes, the problem is cited as ‘GPU monitoring utilities could cease reporting the GPU temperature after PC wakes from sleep’.
Shortly after the affected Sport Prepared driver 576.02 was rolled out, a pinned thread on the Secure Diffusion sub-Reddit, titled Learn to Save Your GPU!, turned a useful resource for anecdotal points and user-reported updates in regards to the new driver. From these, and different stories across the internet, some time-line of emergent issues could be established.
The primary Reddit report of the bug appears to have occurred late Friday afternoon UTC, on the ZephyrusG14 subreddit, the place the person fricy81 cited a put up at NVIDIA boards (archived):
A person at NVIDIA boards finds points after the 576.02 replace. Supply: https://www.nvidia.com/en-us/geforce/boards/game-ready-drivers/13/563010/geforce-grd-57602-feedback-thread-released-41625/3524072/
The person at NVIDIA boards reported that after putting in the driving force replace, instruments like MSI Afterburner and in-game displays such because the one in Name of Responsibility (which usually entry native system readings, a lot as Process Supervisor’s GPU panel does in Home windows) stopped updating GPU temperature readings, freezing at round 35-36°C.
Restarting the monitoring software program had no impact, the person acknowledged, and solely a full system reboot would restore correct readings. Instruments like HWInfo and NVIDIA’s personal monitoring app continued to report temperatures appropriately. The person emphasised that the problem occurred throughout regular use, not simply after waking the system from sleep.
Person suggestions throughout numerous boards highlighted a normal disruption of regular fan curve habits and an alteration of core thermal regulation, leading to graphics processing items idling at unexpectedly excessive temperatures, and alarmingly overheating underneath what would usually be thought of normal operational masses, as detailed on this remark:
‘I may inform one thing was off. The climate exterior was in all probability round 55°F / 12°C, however I used to be cooking alive in my room. My window was open, and but I couldn’t really feel any distinction. All of the followers have been working at max, and temps regarded positive at first—round 68°C to 72°C after gaming for some time.
‘At first, that appeared regular—till the subsequent morning, after I realized these aren’t idle temps, and the followers have been nonetheless [kicking].
‘I had accomplished some AI overclocking after fixing a number of issues recently, so I wasn’t positive if the values had simply spiked too excessive. It’s occurred as soon as earlier than after putting in ASUS AI Suite 3 – the BIOS settings wouldn’t even work correctly due to it.
‘Anyway, I went forward and rolled again to an older driver for now.’
Sub-Optimum
The official launch PDF for the 576.02 driver replace affords some clues about adjustments which will have contributed to the brand new points. In part 5.5, NVIDIA acknowledges that GPU temperature could be reported incorrectly on NVIDIA Optimus techniques, particularly displaying zero levels when no functions are working.
Part 5.5 of the official 576.02 replace notes addresses temperature-monitoring points that appear to have affected a wider variety of techniques than the Optimus system. Supply: https://us.obtain.nvidia.com/Home windows/576.02/576.02-win11-win10-release-notes.pdf
The discharge states:
5.5 GPU Temperature Reported Incorrectly on Optimus Techniques
5.5.1 Subject
On Optimus techniques, temperature-reporting instruments corresponding to Speccy or GPU-Z report that the NVIDIA GPU temperature is zero when no functions are working.
5.5.2 Rationalization
On Optimus techniques, when the NVIDIA GPU just isn’t getting used then it’s put right into a low-power state. This causes temperature-reporting instruments to return incorrect values. Waking up the GPU to question the temperature would end in meaningless measurements as a result of the GPU temperature change because of this.
These instruments will report correct temperatures solely when the GPU is awake and working.
NVIDIA Optimus is a GPU switching know-how that toggles between built-in and discrete graphics primarily based on utility calls for, with the intention to routinely stability efficiency and energy consumption, designed to preserve battery life and scale back energy consumption. For duties corresponding to gaming or HD video playback, Optimus prompts the discrete GPU for higher efficiency; throughout lighter actions corresponding to internet looking, it reverts to built-in (onboard) graphics.
The replace seems to have prolonged a habits beforehand restricted to Optimus techniques, permitting the affected GPU to enter a low-power state whereas idle, even when not hosted on an Optimus system, in flip disrupting temperature reporting in third-party instruments.
Threat Adjustment
In most situations, it’s truthful to say that the graphics card’s VBIOS would doubtless have prevented everlasting GPU injury. VBIOS enforces thermal and energy limits on the firmware stage, independently of the driving force.
Subsequently even when a driver have been to trigger improper fan habits or misreport temperatures, the VBIOS ought to nonetheless throttle efficiency, ramp up fan exercise, or else shut down the GPU to forestall {hardware} failure.
That doesn’t imply the chance was trivial – sustained excessive temperatures can degrade efficiency over time or stress adjoining elements; moreover, absent a typical understanding that an up to date driver brought about an issue (not least in techniques the place drivers replace ‘silently’), a problem of this nature may mislead a big proportion of affected customers, who could try cures for non-existent issues, and even doubtlessly trigger injury to their techniques by making use of non-relevant ‘fixes’.
The errant habits attributable to replace 576.02 was significantly alarming for these engaged in synthetic intelligence workflows, the place high-performance {hardware} is routinely pushed to its thermal limits for prolonged durations.
The problematic 576.02 driver impressed a broader rash of complaints after its launch in mid-April, regardless of preliminary stories that it supplied some useful efficiency enhancements. However the availability of the hotfix, and the extent of disruption that 576.02 appears to have brought about, on the time of writing it stays obtainable for obtain* at NVIDIA’s website.
Afterglow
By way of the fallout from the defective replace, there are quite a few kinds of injury and or inconvenience reported: person Frankie_T9000 reported that his GPU crashed on boot as a result of warmth buildup underneath the fault replace, and solely stabilized after undervolting. He commented ‘appears like its not completely harmed however must repaste asap (I’ve pads coming wednesday) suspect the previous thermal paste was aged extra by the warmth buildup so im placing new paste pads.‘
Yesterday one other person in the identical thread acknowledged: ‘Im utilizing a customized fan curve wit msi afterburner, and it stored displaying that my gpu temps have been continually at 27°C, so the followers did not activate, which led to overheating points. I believed it was a me subject however after putting in the earlier driver all of it labored out positive once more. Additionally, the temps arent displayed appropriately in taskmanager.’
Although NVIDIA (because it states persistently in every hotfix launch) usually offers hotfixes for specific video-games or platforms, the chance of warmth injury to or round a GPU is greater for AI practitioners than for videogamers, since intensive machine studying processes corresponding to coaching or sustained inference place a GPU underneath constant long-term load – an occasion more likely to be triggered solely periodically in a recreation, which can ‘spike’ into excessive utilization for a boss-battle or a very demanding map part, however which is in any other case designed as a compromise between GPU exploitation and system stability.
* Archive: https://archive.ph/ylVR1
First revealed Tuesday, April 22, 2025