Playable · Completable
October 3 repeat, PPSA02457 v1.022.125: the installed runner restores the Survival save and renders the snowy crash site and HUD. This run measures the current build after the shared GTA III renderer changes; it adds no new emulator fixes.
Second fresh launch, same executable and Survival save: menu 15.50 FPS; two unpaused 30-second stationary world samples 12.70 and 8.50 FPS, 10.60 FPS combined. Output 1080p, Speed preset, warm caches. No ten-second stop in these intervals, but shorter delays remain. The previous launch averaged 9.52 FPS. This is run-to-run variation, not a new optimization; 30 FPS remains unmet.
NVIDIA GeForce RTX 3070 Ti
Result
Every recorded run for this title, newest first. Entries marked as a development build were never published as a release.
Second fresh launch, same executable and Survival save: menu 15.50 FPS; two unpaused 30-second stationary world samples 12.70 and 8.50 FPS, 10.60 FPS combined. Output 1080p, Speed preset, warm caches. No ten-second stop in these intervals, but shorter delays remain. The previous launch averaged 9.52 FPS. This is run-to-run variation, not a new optimization; 30 FPS remains unmet.
Read the report →
October 3 repeat, PPSA02457 v1.022.125: the installed runner restores the Survival save and renders the snowy crash site and HUD. This run measures the current build after the shared GTA III renderer changes; it adds no new emulator fixes. Main menu: 14.67 FPS. Two unpaused 30-second stationary world samples: 7.27 and 11.77 FPS; 9.52 FPS combined. The first includes a 9.998-second stall. Output 1080p, Speed preset, warm caches. The earlier valid sample was 7.93 FPS, but this is not a controlled speedup comparison. 30 FPS remains unmet.
Read the report →
A standalone reproduction exited with 0x40010006 when Windows debug output ran on the firmware stack. The stack switch now updates Windows bounds, and guest escapes restore the enclosing HLE state. ANSI/Unicode output, 8 stack tests and 9 native bridge tests pass. The installed build starts without a debugger or a temporary-directory override and restores the save. Unpaused gameplay records 7.93 FPS over 30.02 seconds. This is not a controlled performance comparison; 30 FPS, complete rendering correctness and long-session stability remain unverified.
Read the report →
Large indirect-image tables no longer occupy ordinary draw/dispatch stacks. Scalar pointer recovery reuses an immutable register snapshot and initializes only the eligible bitmap range. 72 scalar/resource tests and targeted Vulkan probes pass. A fresh process restores the save and records 11.96 FPS over 30.01 seconds in unpaused gameplay. Different weather and warm-up prevent a controlled speedup claim; 30 FPS remains unmet. A bounded trace locates the remaining scalar-resource warning in a stencil-only draw with colour writes disabled. Full rendering correctness and long-session stability remain unverified. A separate trace finds repeated 4 MiB mappings holding the memory lock. Reusing physical page commitment and removing duplicate native queries reduces the mapping microbenchmark median by 18%; a gameplay gain is not established.
Read the report →
Sampled-image descriptor updates reuse temporary arrays, reducing their stack frame from 753,720 to 56 bytes without changing the Vulkan batch. A separate regression fixes unmap failures that left removed native pages marked readable. 154 renderer/index, 29 memory and 60 submission tests pass, along with Vulkan checks including 4352 mixed texture views. The first fresh load restores the saved world; a stationary unpaused sample records 12.03 FPS over 30.01 seconds. Different weather and warm-up prevent a controlled speedup claim. One scalar resource remains unresolved. 30 FPS, reliable loading and a full playthrough are not verified.
Read the report →
Reusable scalar storage reduces the draw function stack from 447,424 to 32,640 bytes. A further loading crash in hashing exposed a check-to-read race: GPU input copies and hashes now hold the mapping until they finish. A concurrent-unmap regression passes, along with 34 memory/lifetime/index tests and 60 submission tests. The first follow-up restores the saved world and measures 8.63 FPS over 30.01 seconds without pausing. Different weather and warm-up prevent a direct comparison with the earlier 10.00 FPS sample. One shader resource remains unresolved; 30 FPS and long-session stability are still unverified. A fresh repeat with the same executable still crashes during loading in hash + 0xf0. The protected-read regression is valid, but this observed crash is not fixed; the failing caller is under investigation.
Read the report →
A fresh process restores the same saved world again after the buffer ownership fix. Bounded overlap queries replace full scans without changing writer order or the 4096-entry cache limit. One paused ABBA comparison gives 11.93–12.23 FPS for linear scans and 12.30–12.43 for indexed scans; this small difference is not a proven general gain. A separate unpaused, fixed-camera sample records 300 flips in 30.01 seconds: 10.00 FPS, with changing weather and frost. 154 backend/index tests and five targeted Vulkan checks pass. 30 FPS, long-session stability and complete rendering correctness remain open.
Read the report →
A traced 6 MiB GPU readback overlapped a corrupted game object during loading. Buffer caches now track mapping ownership so a new allocation at an old address cannot receive a retired result. The first follow-up restores the saved snowy world and responds to walking input. This is one successful recovery, not proof of long-session stability. 29 memory/retirement tests, 60 submission tests and four targeted Vulkan probes pass. Increasing cache capacity alone measured 9.93 to 9.10 FPS in the same paused scene; the default remains unchanged. 30 FPS is still unmet.
Read the report →Normal Save now writes a 213,388-byte archive, commits it and returns to gameplay. A fresh process recognizes the slot with the correct date and duration, without the damaged-save label. Shared fixes cover POSIX file writes and the complete save-parameter structure; 52 save/filesystem tests pass. Restoring the world still encounters memory corruption, so reliable save recovery is not claimed. A separate released-command-arena check passes 60 submission tests; its connection to the loading failure is unproven.
Read the report →
CPU resource walks bypass vector-only interpretation while retaining dependency checks and GPU instructions. A separate correction invalidates both words of vector mask outputs. All 71 scalar tests and nine GPU probes pass. Isolated walks take 15–36% less time; the latest 30-second world sample records 7.00 FPS, without a controlled before/after gain. Copying, resource preparation and submissions remain expensive. Keyboard input now follows foreground focus. Loading corruption, one unresolved binding and the 30 FPS target remain open.
Read the report →
The captured G-buffer and final output requested sRGB but used UNORM attachments, darkening colours when read back as sRGB. The renderer now encodes colour writes and preserves encoded bytes during display transfer. A GPU regression verifies sampling, alpha and scanout, including Vulkan validation; 129 focused tests pass. Mip backing checks also retain their calculated byte ranges. See the report for live runs and remaining limits. The new run reaches the snowy world with visibly brighter materials and records 9.49 FPS over 30.05 seconds (baseline: 9.13 FPS). Different weather and effects prevent a controlled speedup claim; 30 FPS, the unresolved scalar binding and intermittent loading failures remain open.
Read the report →
A real-game trace found ten RG32F mip readbacks per frame (2730 KiB) followed by repeated uploads. Shared-memory validation now distinguishes a verified GPU publication from a CPU replacement. All four Vulkan cases with linear/packed mips and memory tracking pass, alongside 129 focused tests. In the new game run, menu render-target uploads and readbacks fall to zero and the ten-level pyramid stays on the GPU. Dark lighting, intermittent loading failures and the 30 FPS target remain open. The candidate reaches the snowy world; a 30.01-second stationary sample records 260 flips (8.66 FPS), so an overall FPS gain is not established.
Read the report →Resident texture lookup now keeps mip levels and array slices distinct. Complete colour pyramids can be assembled on the GPU. A six-level RG32F probe verifies each level, a same-frame rewrite and sampling without extra target readbacks or texture uploads; 128 focused tests pass. The game candidate still faults during loading, including diagnostic synchronous and 8192-entry-cache runs. A world FPS gain and the final lighting result remain unverified.
Read the report →Three additional runs of the previous build failed during loading or the world transition, after two earlier runs had reached gameplay. A worker can stop while audio continues and the window stays black. The old log also confirms a 1×1 mip incorrectly refreshing the 512×512 base texture. These failures are recorded separately from the earlier successful run; stable loading is not yet established.
Read the report →
October 1 development build, PPSA02457 v1.022.125: Survival passes loading and the intro, then renders the snowy crash site and HUD. Fixes protect CPU memory from stale GPU writeback, retire released storage buffers and preserve colour during depth-only draws. RG32F attachments, mip views and D16 shadow bias are supported. This validation did not include a full playthrough, save recovery or audio correctness.
Read the report →
Index staging, pipeline lookup, scalar register snapshots, queue probing and scratch initialisation were each made cheaper, all without any title-specific condition. The menu settled at roughly 17 FPS — still well short of the 30 FPS target.
Read the report →
Presentation moved to native 1920×1080 and the shared graphics path stopped copying decoded instruction structures around during resource preparation and scalar interpretation. Play, Options and Credits are readable; water and lighting artifacts remain.
Read the report →Early runs stopped dead at the same guest address. Unity's shader reader had taken four bytes of mesh data as a signed string length and written a terminator into unmapped memory. Fixing the file descriptor behaviour behind it got the title past startup.
Read the report →
The animated background was already running, but no menu appeared: the platform initialisation coroutine had stopped inside FMOD, leaving the services the start screen waits for null forever. These were the last 4K-era measurements before the switch to native 1080p.
Read the report →Captures
Frames produced by the emulator itself during the runs above.
Second fresh-process performance repeat
October 3, 2026
Performance repeated on the current runner
October 3, 2026
Windows stack bounds fix diagnostic-output startup failure
October 2, 2026
Less resource preparation overhead; stencil warning located
October 2, 2026
Descriptor workspace reuse and safer failed unmaps
October 2, 2026
Smaller draw stack and protected guest reads
October 2, 2026
Second save recovery and buffer overlap measurements
October 2, 2026
Saved world loads after buffer ownership correction
October 2, 2026
Cheaper CPU shader walks; world sample remains at 7 FPS
October 2, 2026
sRGB colour writes corrected; mip checks reuse their ranges
October 1, 2026
Tracked mip memory: repeated transfers removed
October 1, 2026
New Game reaches the opening world
October 1, 2026
A run of CPU-side reductions across the shared draw path
September 28, 2026
Native 1080p output and cheaper resource preparation
September 27, 2026
The menu was missing because audio initialisation stalled
September 26, 2026
19
Recorded runs
NVIDIA GeForce RTX 3070 Ti
0.3.3
Current build
October 4, 2026
Playable
Result
Playable · Completable
Playthrough reports carry the date the maintainer confirmed them. Every timing refers to the NVIDIA GeForce RTX 3070 Ti development machine and will differ on other hardware. Game content is supplied locally and is never distributed here.
