https://github.com/NotRexed/Hookshot/releases/tag/small-vram uses much less VRAM on cards with 8 GB or less. Your log showed the renderer going past what the card allows (5.9 GB used on a 3.4-5.1 GB budget), and then the GPU timed out. Please try it. If it still goes black or crashes, set Resolution to 75% or 67% in the Insert menu and send hookshot_log.txt.

leonidasmark commented

2026-11-07 16:03:39 +00:00

(Migrated from github.com)

RTX 2070 8 GB (Turing), i7-8700, 32 GB RAM, Windows 11, 4K desktop. Built from source with the command-line route. Results for each build, in order, in case they help with 8 GB cards:

Build	Settings	Result
`brighter` (`4756836`), driver 621.01	defaults	black screen with sound: `submit failed (-4)` after VRAM grew (2 GB textures + 1.7 GB mirrors)
`small-vram` (`c02f290`), 621.01	defaults	still black: device lost in a compute dispatch at frame 2016 with only ~1.6 GB VRAM in use
`c02f290`, 621.01	`RX_VK_VRAM_MB=3500`, 75%	no crash, but 86-96 ms of GPU time a frame in play (~10 fps)
`c02f290`, 621.01	`RX_VK_VRAM_MB=2048`, `RX_VK_MIRROR_MB=2048`	GPU 16-19 ms a frame; the replay thread became the limit
`c02f290`, 621.01	same, 67%, native matrices	device lost again (compute 53caac5f…, frame 1962) at ~1.4 GB VRAM
`c02f290`, 621.01	same + `RX_VK_MMA_NO_COOP=1`	one more GPU fault entering a scene
`c02f290` → 621.14	same + `RX_VK_ROBUST=1`, driver updated at the same time	stable
`1c7c9c7`, 621.14	defaults (menu: 50%, frame generation 2x, pacing on)	stable for ~24,600 frames on the native matrix path, GPU 13-15 ms, 33-50 fps game

So 1c7c9c7's 2 GB of mirrors on 8 GB cards matches what fixed the frame rate here (1 GB left the GPU at 86-96 ms).

Every black screen lined up with a pair of nvlddmkm event 153 entries ("Error occurred on GPUID: 100") in Windows' System log, a few seconds apart, including one where the game logged no device lost (GPU time just dropped to 0). That may be a quicker way for others to tell a GPU reset from a renderer bug.

What is still off on 1c7c9c7 with defaults: VRAM sits above what the driver allows the whole time, 7.1-7.35 GB used of 6.98-7.05 GB allowed, even at 50% resolution, and the textures run 0.3-0.6 GB over their own budget (e.g. 3895 MB on a 3383 MB budget). It did not crash this time, but that is the state the earlier crashes started from.

I can't separate the driver update from the code changes for the crash, since both came together.

RTX 2070 8 GB (Turing), i7-8700, 32 GB RAM, Windows 11, 4K desktop. Built from source with the command-line route. Results for each build, in order, in case they help with 8 GB cards: | Build | Settings | Result | |---|---|---| | `brighter` (4756836), driver 621.01 | defaults | black screen with sound: `submit failed (-4)` after VRAM grew (2 GB textures + 1.7 GB mirrors) | | `small-vram` (c02f290), 621.01 | defaults | still black: device lost in a compute dispatch at frame 2016 with only ~1.6 GB VRAM in use | | c02f290, 621.01 | `RX_VK_VRAM_MB=3500`, 75% | no crash, but 86-96 ms of GPU time a frame in play (~10 fps) | | c02f290, 621.01 | `RX_VK_VRAM_MB=2048`, `RX_VK_MIRROR_MB=2048` | GPU 16-19 ms a frame; the replay thread became the limit | | c02f290, 621.01 | same, 67%, native matrices | device lost again (compute 53caac5f…, frame 1962) at ~1.4 GB VRAM | | c02f290, 621.01 | same + `RX_VK_MMA_NO_COOP=1` | one more GPU fault entering a scene | | c02f290 → 621.14 | same + `RX_VK_ROBUST=1`, driver updated at the same time | stable | | **`1c7c9c7`, 621.14** | **defaults** (menu: 50%, frame generation 2x, pacing on) | **stable for ~24,600 frames on the native matrix path**, GPU 13-15 ms, 33-50 fps game | So 1c7c9c7's 2 GB of mirrors on 8 GB cards matches what fixed the frame rate here (1 GB left the GPU at 86-96 ms). Every black screen lined up with a pair of `nvlddmkm` event 153 entries ("Error occurred on GPUID: 100") in Windows' System log, a few seconds apart, including one where the game logged no device lost (GPU time just dropped to 0). That may be a quicker way for others to tell a GPU reset from a renderer bug. What is still off on 1c7c9c7 with defaults: VRAM sits above what the driver allows the whole time, 7.1-7.35 GB used of 6.98-7.05 GB allowed, even at 50% resolution, and the textures run 0.3-0.6 GB over their own budget (e.g. 3895 MB on a 3383 MB budget). It did not crash this time, but that is the state the earlier crashes started from. I can't separate the driver update from the code changes for the crash, since both came together.