Void Compute

GPU Hardware Operating Software Translator

Project maintained by Void-Compute Hosted on GitHub Pages — Theme by mattgraham

GHOST

GPU Hardware Operating Software Translator

Breaking hardware monopolies through low-level systems architecture.

Documentation • Roadmap • Devlog • Discord

![Isaac Sim Full (Newton Physics) — live viewport on AMD Radeon RX 7800 XT](/assets/isaac_sim_newton_amd.png) **Isaac Sim Full (Newton Physics) — live viewport, 92 FPS, on AMD Radeon RX 7800 XT.** *No NVIDIA components in the render path.*

Overview

GHOST is a runtime interoperability environment that intercepts and routes compute instructions across GPU architectures. It enables standard, high-performance execution of proprietary ML workloads on non-native silicon — no virtualization, no source access required.

Core Features

CUDA on AMD: CUDA-only software runs on RDNA/Vega via ZLUDA, orchestrated by the Ghost driver-identity layer.
Newton Physics on Warp: NVIDIA’s next-generation physics engine initializes and executes through the Ghost compute stack.
Smart Failover: Applications launch natively via ROCm first; on crash or incompatibility, Ghost automatically relaunches with the ZLUDA translation layer.
Runtime Identity Profiles: Per-series device identity mapping (Vega through RDNA3) keeps capability checks coherent across GPU generations.
JIT Compatibility Stubs: Hardware-specific interceptors are generated and compiled on the fly — no static binaries, tailored per machine.
Waiting Room TUI: While the AI loads: live VRAM/temperature/load telemetry, background music, and a fully working version of DOOM in your terminal.

The Milestone

NVIDIA Isaac Sim — the industry-standard robotics simulator, historically locked to NVIDIA GPUs — running its full application with a live viewport, Newton physics, and robot workloads on an AMD Radeon RX 7800 XT.

Official requirements state an NVIDIA RTX GPU is mandatory. Our compatibility research demonstrates otherwise. The full technical write-up is in progress — the method stays close to the chest while it matures. Community members get the details first.

Architecture & Roadmap

GHOST is built on a modular, layer-separated philosophy: identity, translation, and orchestration are independent components with explicit contracts. Our primary focus is platform hardening before horizontal expansion.

Phase 1: The AMD Core (Current)

Windows port — full Rust rewrite (v1.56), hardened runtime daemon
ROCm / HIP execution logic with smart ROCm↔ZLUDA failover
Per-series identity profiles: Vega through RX 9000 mapping
Waiting Room TUI: live telemetry + DOOM + music
Isaac Sim (Newton) compatibility milestone — live viewport on RDNA3
Trace infrastructure — per-call diagnostics as the community bug-report format
Multi-GPU orchestration and device-specific targeting
MDL material fidelity for the alternate viewport path

Phase 2: Heterogeneous Expansion (R&D)

Linux build — feature parity with the Windows daemon
Intel Track: oneAPI / Level Zero integration (SYCL)
Apple Track: Metal Performance Shaders / MLX routing for unified memory
Ray-traced rendering delegate research for the compatibility viewport

Phase 3: Silicon Re-Architecting (Future)

Core-Stripped NPU: Repurposing standard GPU Compute Units for dedicated, direct-to-ALU AI inference.

Installation & Quick Start

Windows 10/11 (primary platform):

Download the latest build from the Releases page.
Right-click ghost_amd.exe → Run as Administrator (Ghost manages registry configuration and deploys system stubs — elevated rights are required).
On first launch, Ghost auto-downloads its dependencies (ZLUDA, HIP SDK components) and detects your hardware. First-time setup also needs the Visual Studio Build Tools (C++) for the JIT stub compiler — doctor verifies this for you.
Inside the (ghost-amd)> shell: run <your_script.py> launches your AI workload with smart failover, or explore doctor, benchmark, translate, clean.

Linux: work-in-progress, not yet at feature parity with the Windows build.

Full requirement list and troubleshooting: Technical-Docs

Developer Log

We believe in building in public — including the unglamorous parts.

v1.56 (current) — The Rust Edition: Complete rewrite from the PowerShell prototype into a hardened Rust runtime daemon: interactive shell, smart failover, Waiting Room TUI, JIT stub compiler, per-series identity profiles (Vega → RX 9000).
2026-09 — The Isaac Sim milestone: After mapping NVIDIA’s rendering-architecture boundaries across the full Isaac Sim install, Ghost’s alternate rendering path was validated end-to-end: full application, live viewport at ~92 FPS, Newton physics on the Ghost compute stack — on an AMD RX 7800 XT. Technical write-up in progress.
Warp validation: NVIDIA Warp 1.13 (runtime CUDA kernel generation via NVRTC) initializes fully through the Ghost stack on RDNA hardware.

Community & Support

GHOST is a community-driven ecosystem with an active beta-testing program across Vega and RDNA hardware.

Discord — updates, support, beta testing, and the place where compatibility questions get answered first.
GitHub Issues — bug reports and hardware requests; include your GPU model, driver version, and ghost_trace.log.
GitHub Discussions — architectural questions and PR submissions.

Explore the GHOST Ecosystem on GitHub