Skip to content

HeavenBench 1.1.1

Choose a tag to compare

View all tags
@dev-zetta dev-zetta released this 04 Nov 23:48
v1.1.1
7930b9e

HeavenBench 1.1.1 fixes GPU selection, cached geometry integrity and renderer resource lifetimes, and reduces avoidable rain and ZIP-loading work.

Videos

Cinematic sequences are offline rendered; playback frame rate is not measured benchmark performance. Retained UI, browser and tablet inserts in the showcase are labelled Archive 1.0.0.

Fixes and improvements

  • Explicit Linux Vulkan GPU requests now verify the actual selected device and, for index/PCI requests, its physical PCI identity. OpenGL retains its documented best-effort preference and verification limits.
  • Tessellation caches distinguish original source-surface ordinals, preventing same-name surfaces from sharing geometry. ZIP models and manifests validate logical EOF/CRC before acceptance. Neutral ZIP cache semantics advance to 37; incompatible entries regenerate automatically.
  • Binary asset readers use bounded 64 KiB read-ahead. A controlled warm fort.mesh ZIP workload uses about 59% less elapsed time per load; this is a specific CPU-loader result, not a general startup or rendering-FPS claim.
  • Effective-zero rain removes invisible draw work and rain-only depth demand while retaining wetness and persistent geometry. Shared depth refreshes retained individual/instanced placements once when rendering resumes.
  • Temporary OpenGL height-sampling surfaces hand retiring current GLX bindings to the existing dummy drawable. The fix resolves the reproduced Intel teardown crash and clears the retired swap-chain cache pointer.

The reviewed source passes 54 native cases with five disclosed skips on Intel OpenGL, AMD OpenGL and AMD Vulkan, 39/34 SDK-on/off host tests, nine GPU-selection controls, a 433-source Clang-Tidy gate and a final-source ASan/UBSan reader oracle. Native negative controls reproduce the original cache, rain, depth-resume and Intel GLX failures. The dated review records exact evidence and scope.

Packages and compatibility

Standard packages target Linux x86-64 portable/Zen1/Zen5, Windows x64, Apple Silicon macOS 15.0+ and Android ARM64. Choose the portable Linux build unless the CPU-specific build matches your processor. Original Heaven assets remain external; see Getting started.

Separate experimental Linux FSR 3.1 and Windows FSR 3.1/DLSS-adapter packages retain generation Off by default. Proprietary NVIDIA runtime plugins are not bundled and native NVIDIA execution remains untested. Generated submissions are not confirmed display FPS; requested generation makes tours diagnostic and unscored.

Android versionCode is 5 and the APK is development-signed. Its signing certificate differs from the published 1.1.0 APK, so it cannot update that installation in place: back up imported assets and user data, uninstall the old application, then install 1.1.1. Host signature/alignment checks pass; physical installation is not claimed. macOS remains ad hoc signed without Developer ID notarization. Package/CI success does not establish unrestricted visual quality or physical rendering, HDR, latency and display-cadence acceptance on every platform.

Extraction tools remain 0.3.1, result schema 2 and score version 3. Compare benchmark results only with matching application/score versions, workloads and effective settings.

All five fresh workflows and eight package jobs pass at release commit 7930b9eaae5feef8dfae7396e479d534848d721f. Independent audits verify all eight application packages, two exact-source archive variants, package contents/notices and checksums. Both portable and experimental Linux AppImages pass relocated CLI checks; the portable package also passes Intel OpenGL and AMD Vulkan procedural height-sampling/teardown controls.

The experimental Linux package passes two bounded Intel Vulkan/X11 full-scene controls at 960×540 with pinned Vulkan validation: Off records 497 real sources; FSR records 301 real and 296 generated native submissions, with zero presentation failures and zero validation errors or warnings. Both processes exit cleanly with unchanged package inputs. These are startup/submission checks, without a new performance, generated-image quality, physical display or latency claim.

Use the release manifest, provenance archive and SHA256SUMS to verify the downloadable bytes and their exact source/evidence bindings.

Assets 26
Loading