一个开源项目通过整合ZLUDA和AMD ROCm/HIP技术栈，实现了在Windows系统中于AMD显卡上直接运行CUDA目标应用的兼容性方案1。该方案已在AMD Radeon RX 9060 XT上进行了验证1，支持cuBLAS、cuSPARSE、cuFFT等多个CUDA库的调用1，并通过包含2,216,347个参数的PPO神经网络完成前向推理和学习任务的测试1。
该实现采用了ZLUDA v6-preview.69、AMD HIP SDK 6.4和LibTorch 2.3.0 + cu118等核心版本组件1。性能对比数据显示，上游路径达到13,278 median SPS，而自定义覆盖方案的性能为12,876 SPS1。不过该方案仍存在明显局限，目前仅在RX 9060 XT上经过验证1，且不支持cuDNN/MIOpen、NCCL、TensorRT以及某些自定义CUDA扩展1。
An open-source project has successfully implemented a compatibility solution that allows CUDA-targeted applications to run on AMD graphics processors under Windows, leveraging ZLUDA and AMD's ROCm/HIP technology stack 1. The initiative has been validated on AMD Radeon RX 9060 XT hardware and demonstrates support for core CUDA libraries including cuBLAS, cuBLASLt, cuSPARSE, and cuFFT 1.
The implementation was tested using ZLUDA v6-preview.69, AMD HIP SDK 6.4, and LibTorch 2.3.0 with cu118 1. Functionality was verified through a PPO neural network containing 2,216,347 parameters, which successfully completed forward inference and training operations 1. Performance benchmarking conducted on September 13, 2026, showed the upstream path achieving a median throughput of 13,278 SPS compared to 12,876 SPS with custom overrides 1.
The solution carries significant limitations, with current validation restricted to the AMD Radeon RX 9060 XT 1. The implementation does not support cuDNN/MIOpen, NCCL, TensorRT, or certain custom CUDA extensions 1.
评论
还没有评论，欢迎留下第一条。