VKD3D-Proton translates Direct3D 12 to Vulkan, DXVK covers Direct3D 11 down to 8, and MetalSharp's MoltenVK carries both to Metal, with Apple's Metal Shader Converter compiling most shaders straight to Metal libraries. Version 5.4 runs on both MetalSharp 1.3.0 runtimes.
Two translators feed one Vulkan driver. Shaders take the short path through Apple's converter wherever it can carry them.
DXBC shaders reach the converter through DirectXShaderCompiler's dxilconv. If the converter can't take a shader, or METALSHARP_MSC=0 is set, that pipeline is built from SPIR-V as before. Geometry and tessellation shaders stay on SPIR-V.
Since MetalSharp 1.0.0
Five releases in a week.
MetalSharp 1.0.0 shipped 4.5 and 1.1.0 shipped 5.3. MetalSharp 1.2.0 and 1.3.0 ship 5.4, inside MetalSharp Wine 6.5 and 7.0.
4.5Oct
Shipped in MetalSharp 1.0.0.
5.0Nov 3
DLSS Frame Generation on MetalFX, background pipeline compiles, and a pipeline cache that hits.
5.1Nov 6
Minecraft Bedrock's ray tracing renders, and Unreal Engine 5 Nanite games like Subnautica 2 run upright.
5.2Nov 7
Most remaining pipelines move to Metal Shader Converter, with accuracy fixes found on the way.
5.3Nov 7
Black 3D in DXVK games fixed, and DLSS in an x86_64 Wine. Shipped in MetalSharp 1.1.0.
5.4Nov 9
CONTROL Resonant renders: sparse BC textures, mesh shaders and ray tracing tables. Ships in MetalSharp 1.2.0 and 1.3.0.
5.4
Streamed materials, not block noise.
CONTROL Resonant streams its materials through about 5,000 reserved BC textures. Compressed sparse textures now stay on Metal's tier 2, so they render as materials instead of colored blocks, and texture shapes where Metal would lose the mip tail fall back to an ordinary texture.
5.4
Mesh shaders and ray tracing.
SPIRV-Cross writes a mesh shader's array outputs the way Metal accepts, so all 49 of CONTROL Resonant's mesh shaders compile, and ray tracing shader tables with a stride of 0 no longer hang the GPU.
5.4
ExecuteIndirect, and cheaper timestamps.
Command signatures that change a root CBV run instead of failing with E_NOTIMPL. Timestamps ride the next render or compute pass instead of a blit pass each: Subnautica 2's API insights went from 689 to 112.
Performance
Cheaper draws, faster shaders.
Measured with the project's tests/perf scenes and reported in each release.
−23%CPU time, geometric mean of 147 test scenes, 5.2 against 5.1 (Apple M6).
−21%GPU time across the same scenes.
1–4 msPer large shader through Metal Shader Converter, against 0.4–1.7 s through SPIR-V and MSL text.
5.2 against 5.1, geometric mean of 147 scenes
Reduction, higher is better. x86_64 under FEX on an Apple M6.
CPU time−23%
GPU time−21%
Peak memory−9%
0%5%10%15%20%25%
Show as a table
Measure
Change, 5.2 vs 5.1
CPU time
−23%
GPU time
−21%
Peak memory
−9%
Scene
Release
Before
After
Faster by
Constant buffer rows out of range (cbv_rows)
5.0
410 ms
20 ms
20×
Markers batched in a render pass (draw_markers)
5.0
186 ms
1.5 ms
124×
Counted ExecuteIndirect dispatches
5.0
13.7 ms
2.4 ms
5.7×
2000 predicated draws, CPU
5.2
53 ms
0.49 ms
108×
Contended 64-bit atomics, CPU (atomics64_contended)
5.2
384 ms
0.08 ms
4800×
Geometry shader stream-out, GPU (xfb_gs_dxbc)
5.2
11.3 ms
1.0 ms
11×
From the tests/perf scenes in each release's notes: 5.0 on an Apple M4, 5.2 on an Apple M6. All release notes.
Capabilities
What games see.
Measured through VKD3D-Proton and this MoltenVK on Apple silicon.
SM 6.8Shader Model 6.8.
FL 12_1Feature levels 12_1 down to 11_0 by default.
FL 12_2With MoltenVK's ray tracing pipelines turned on.
DXR 1.1Ray tracing tier 1.1, with that same setting.
Shading
Modern pipeline features
Variable-rate shading tier 2
Mesh shader tier 1 (Nanite)
Sampler feedback tier 0.9
Barycentrics and ROVs
Resources
Binding and memory
Tiled resources tier 4
Resource binding tier 3
Fully typed format casting
Depth bounds and min/max sampler filtering
Upscaling
DLSS on MetalFX
Super Resolution and Ray Reconstruction through the lane's own NGX core and NVAPI, and Frame Generation on macOS 27 with the game's own nvngx_dlssg.dll. METALSHARP_DLSS_SCALE sets the render scale.
In MetalSharp 1.3.0
On both runtimes.
FEX runtime
Inside MetalSharp Wine (Arm64)
vkd3d-proton-v5.4-arm64
D3D12
VKD3D-Proton as ARM64X, loaded by x86_64 games as ARM64EC
D3D11–8
DXVK's x86_64 dxgi, d3d11 and d3d10core
DLSS
The lane's NGX core and NVAPI, ARM64X
x86_64 runtime
Inside MetalSharp Wine (x86_64)
vkd3d-proton-v5.4-x86_64
D3D12
VKD3D-Proton for x86_64 Windows programs
D3D11–8
DXVK's dxgi, d3d11 and d3d10core
DLSS
Since 5.3: nvngx.dll and nvapi64.dll as x86_64 Wine builtins
Both run on the universal MoltenVK package (arm64 + x86_64) with Apple's Metal Shader Converter library and DirectXShaderCompiler's DXBC converter. In MetalSharp it's the VKD3D-Proton + DXVK route.
On its own
Install it into any prefix.
MetalSharp sets this up for you. For another Wine prefix:
Download two packages
From the 5.4 release: the VKD3D-Proton package for your Wine (x86_64, or arm64 for an arm64 Wine with FEX) and moltenvk-v5.4-universal. Extract them side by side.
tar -xf vkd3d-proton-v5.4-x86_64.tar.zst
tar -xf moltenvk-v5.4-universal.tar.zst
This repository is AGPL-3.0. VKD3D-Proton is LGPL-2.1, DXVK is zlib, MoltenVK and SPIRV-Cross are Apache-2.0, and Apple's Metal Shader Converter ships unmodified under Apple's license.