Alloyium started from a simple bet: most of what Nvidium does for GPU-driven terrain is algorithm, so the non–Mesh-Shader pieces should be reproducible on OpenGL 4.3 (Compute + SSBO + Indirect Draw) for any vendor—not only NVIDIA.
That bet was only half right. Culling, compaction, and Indirect command generation transferred cleanly. Emulating Mesh Shader vertex export with Compute did not: without an on-chip Shader Export path, writing attributes to VRAM and reading them again in a later stage added cost Mesh Shaders never pay. Near the end of the first architecture, Alloyium could look correct and still run slower than Vanilla + Embeddium.
The rewrite did not drop Nvidium’s mindset. It kept the few optimizations that actually carry the idea—GPU culling, compaction, Indirect MultiDraw—and cut everything that only existed to imitate Mesh Shader hardware. GPU work is culling and Indirect commands only; draws always use Embeddium’s static VBO/IBO (baseVertex / firstIndex). That form is Compute B'—current Alloyium.
| Target | Minecraft 1.20.1 / Forge |
| Required | Embeddium |
| Optional | Oculus (Iris) — see below |
| License | LGPL-3.0-or-later |
| Artifact | alloyium-{SemVer}-Forge+1.20.1.jar |
Embeddium static VBO / IBO
│ read-only
▼
Compute A — frustum / cone / (optional) Hi-Z cull + compaction
▼
Compute B' — emit DrawElementsIndirect commands only
(no vertex-attribute writes)
▼
glMultiDrawElementsIndirect
→ draw Embeddium buffers via baseVertex / firstIndex
Hard rules:
- Do not emit position / normal / UV (or similar) from Compute.
- Final draws always reference Embeddium’s static buffers.
- Goal is a realistic uplift over Vanilla + Embeddium, not parity with Nvidium’s Mesh Shader path.
On this target (Forge 1.20.1), shader packs come through Oculus—the Forge port of Iris—not a separate Iris jar. Alloyium does not hard-disable when Oculus is installed. Solid terrain can still use Alloyium’s Indirect path with Oculus (Iris) program / vertex-format bridging; shadows stay on Oculus (Iris).
That said, several features are limited or inactive for shader compatibility, and Oculus alone can tax FPS even with no pack loaded. Same advice as the in-game warning: if you do not need Oculus and are not planning to use shader packs, removing Oculus is recommended for better performance.
Escape hatch: oculus.irisFallback in config/alloyium-client.toml (while a pack is active, leave terrain to Embeddium / Oculus).
Numbers depend on hardware, settings, and whether you are standing still or moving.
- Standing still, region-level Indirect caching and similar pieces can push very high FPS in favorable setups.
- Moving play is the honest metric; chase idle peaks and you will ship “optimizations” that regress real sessions.
- With Oculus (Iris) and a shader pack, Oculus’s baseline often dominates; Alloyium’s gain is usually small.
- Design target band: about 1.3–1.6× vs Vanilla + Embeddium. Beating Nvidium is explicitly out of scope.
The fix that mattered was not “better culling math”—it was stopping the VRAM round-trip for vertex data.
Old Alloyium tried to stand in for Nvidium across vendors: GPU-driven terrain end-to-end, including a Compute stage that rewrote terrain vertex payloads like a Mesh Shader output stage (Compute B). The Nvidium-shaped pieces that are portable—frustum / cone culling, stream compaction, Indirect draw lists—already worked. The collapse was structural. Mesh Shaders can stream primitives on-chip to the rasterizer; Compute on OpenGL 4.3 must park results in SSBO memory. That Write→barrier→Read path (and the frontend load of oversized Indirect lists) could wipe out every win from culling.
New Alloyium is not a retreat from Nvidium’s spirit. It is the same lineage after keeping only that elite core of ideas and stripping the imitation that had become the bottleneck. Compute no longer owns geometry; it decides what to draw and emits Indirect commands (Compute B'). Embeddium still owns meshing and the static VBO/IBO; Alloyium only supplies offsets and MultiDraw. The goal remains GPU-driven terrain in Nvidium’s sense—cull hard, draw little, stay Indirect—without paying Mesh Shader taxes that OpenGL 4.3 cannot waive. Ceiling stays below true Mesh Shader paths; the self-inflicted bandwidth trap does not.