CUDA for AMD on Windows

This GitHub project unveils a reproducible Windows setup that allows CUDA-targeted applications to run on AMD GPUs using ZLUDA and ROCm/HIP. It directly challenges NVIDIA's CUDA ecosystem dominance, offering a pathway for broader GPU compatibility in AI and compute workloads. Hacker News celebrates this open-source effort as a significant step towards breaking vendor lock-in and fostering hardware agnosticism.

Score

Comments

Highest Rank

13h

on Front Page

First Seen

Sep 13, 3:00 PM

Last Seen

Sep 14, 3:00 AM

Rank Over Time

The Lowdown

Speedstu's GitHub project, "CUDA for AMD on Windows," presents a significant step towards liberating compute workloads from NVIDIA's proprietary CUDA ecosystem. By leveraging ZLUDA and AMD's ROCm/HIP, this setup enables CUDA-targeted Windows applications to execute on AMD GPUs, primarily validated on the RX 9060 XT (gfx1200) for LibTorch AI workloads.

Core Functionality: Allows CUDA-facing compute applications to run on AMD GPUs through a compatibility layer.
Key Components: Built around ZLUDA (a CUDA implementation layer) and AMD's HIP SDK and libraries (rocBLAS, hipBLASLt, rocSPARSE).
Validated Hardware: Successfully tested on an AMD Radeon RX 9060 XT (gfx1200) using CUDA-enabled LibTorch for PPO network training and inference.
API Coverage: Critical CUDA components like nvcuda, cuBLAS, cuBLASLt, cuSPARSE, and cuFFT pass cuda_check via their AMD counterparts.
Known Limitations: The project acknowledges that ZLUDA is not a complete CUDA implementation, and key libraries like cuDNN are unavailable with the current stable Windows HIP SDK, potentially limiting convolution-heavy AI models.
Performance: Benchmarks indicate the upstream ZLUDA path performs slightly better than historical custom overlays for the validated PPO workload.

This project not only offers a practical solution for AMD GPU owners but also stands as a beacon for the open-source community, calling for broader hardware compatibility testing and contributing to the ongoing effort to diminish vendor-specific hardware lock-in in the high-performance computing and AI spheres.

The Gossip

Cracking the CUDA Code: NVIDIA's Moat Under Siege

The main discussion revolves around this project's potential to dismantle NVIDIA's long-standing CUDA ecosystem dominance. Commenters express hope that AI's growth will force a shift towards more open standards and hardware agnosticism, transforming CUDA from a lock-in mechanism into an intermediate representation. There's a strong sentiment against proprietary hardware/software stacks in the ML space, with calls for unified open standards.

Generational GPU Gaps: Compatibility Conundrums

Users eagerly inquire about support for older AMD GPUs, particularly RDNA1 cards like the 5700XT. The discussion quickly highlights the significant hardware limitations of these older generations, such as the absence of matrix multiply (WMMA), BF16, or INT8 support, which severely impacts their suitability for modern AI workloads. This underscores the practical challenges of retrofitting CUDA compatibility onto diverse and aging hardware architectures.

The Broader Battle for Open Compute

Beyond the immediate project, the conversation branches into the larger struggle for open compute standards. Participants lament the fractured state of alternative ecosystems like Apple's MLX/MPS and AMD's ROCm. There's a perceived difficulty for organizations like Khronos to unite vendors such as Apple, AMD, and Intel to create a truly competitive and unified open-source platform akin to CUDA, which continues to limit broader hardware adoption.