Changes
Every release of the client, most recent first. Certificates record the version that produced them, so an older one can be read against what the tool did at the time.
Runs from superseded versions are refused rather than accepted quietly. A certificate is a claim about a card, and one produced by a build known to measure the wrong thing is worse than no certificate at all. That is why the tool sometimes asks you to download it again.
0.7.2
14 Aug 2026 changes what results meanWithdraws the fan check on NVIDIA cards, which 0.7.1 should not have claimed. The documented interface reports the fan speed the driver intends rather than the speed measured at the fan, and states that a physically blocked fan will not change that number, so the reading cannot show the failure it was there to find. The first card to run 0.7.1 reported 0 RPM for five minutes at 220 W with its fans plainly turning. There is a separate interface that does read a real tachometer, but it covers one fan of the several most cards have, and the one that covers them all is not published, so neither can say whether the cooler as a whole is working. No certificate was affected, because a reading of zero never met the threshold to fail anything. The check stands on AMD, where the fan speed is measured rather than intended.
0.7.1
14 Aug 2026Memory temperature is now read on NVIDIA as well as AMD, where the card exposes it.
0.7.0
14 Aug 2026 changes what results meanReads the junction temperature the card actually throttles on, not just the edge sensor, and the safety cutoff moves with it. Memory temperature and fan speed are recorded for the first time: memory at its throttling point means worn thermal pads, and an AMD fan being asked to spin that is not turning has failed.
0.6.2
13 Aug 2026A run that stalls while preparing the render test now says so instead of going quiet.
0.6.0 to 0.6.1
13 Aug 2026 changes what results meanThe stress test now works memory as well as compute. Its working set was small enough to sit entirely in cache, so it never touched VRAM at all. A whole run is about eleven minutes instead of sixteen.
0.5.7 to 0.5.8
13 Aug 2026 changes what results meanThe stress test was not stressing. It drew 149 W on a 220 W card while reporting the GPU as 98% busy, because the load left most of the card idle. It now holds cards at their rated power.
0.5.3 to 0.5.6
13 Aug 2026New console: one screen that updates in place rather than pages of scrolling output. Cancelled runs now record that they were cancelled.
0.5.2
13 Aug 2026 changes what results meanFixed a crash on some AMD cards. The client was allocating from memory types the Vulkan specification forbids it to use without an extension it does not enable.
0.5.1
13 Aug 2026 changes what results meanFixed the bug that capped every run at 4032 MB of memory tested regardless of card size. Coverage went from 25% to 85% on a 16 GB card. Allocations now ask each buffer which memory it accepts instead of choosing once for the whole device.
0.4.0 to 0.4.2
13 Aug 2026A certificate now requires a test session opened before the run and timed by the server, so one cannot be produced without spending the time. Runs upload what machine they were on, so a failure can be diagnosed without asking anyone for their console.
0.2.0 to 0.3.1
13 Aug 2026 changes what results meanThe memory test is split across several buffers. A single descriptor can only address 4 GB, and exceeding that silently wrapped, which is why earlier runs reported billions of errors on healthy cards.
0.1.0 to 0.1.9
12 to 13 Aug 2026First working client: three tests, hardware fingerprint, signed certificate.