IdeaWell

Project ideas from Hacker News discussions.

CUDA for AMD on Windows

Original Article

Hacker News Discussion

📝 Discussion Summary (Click to expand)

Theme 1 – Feasibility and limits of running CUDA on AMD hardware
Many commenters discuss the promise (and current limits) of projects like ZLUDA that let CUDA‑targeted code run on AMD GPUs.

“RDNA1 isn't good for a whole lot, even flagship RDNA2 cards are a stretch for many things. The lack of WMMA/matrix multiply/BF16 is too severe of a penalty.” – monster_truck
“I wish there were a way to use RDNA1 cards with CUDA for AMD. My 5700XTs are sitting in a drawer.” – system2

Theme 2 – Nvidia’s CUDA moat and ecosystem inertia
The discussion repeatedly points to CUDA’s entrenched advantage, suggesting that even if translation becomes easy, the surrounding stack keeps Nvidia ahead.

“Nvidia's moat is (and will remain for the foreseeable future) the entire stack. you cannot fathom the pain and misery of working on literally any other stack.” – mathisfun123
“AI will take down Nvidia’s moat. When it becomes trivial to translate CUDA/PTX to HIP, SYCL or Metal, CUDA is no longer the moat, it becomes the intermediate representation.” – swerner

Theme 2 – Need for and frustrations with open standards
Several users argue that a unified, vendor‑neutral approach (HIP, SYCL, OpenCL, OneAPI) is essential but hampered by fragmentation and poor developer experience.

“Off-topic and somewhat of a rant, but I'd far prefer us all focusing on open standards like HIP, SYCL, OpenCL, etc.” – linuxhansl
“I would too, but sadly that's Khronos' job to organize, and they've had trouble getting American vendors to work together.” – bigyabai
“The problem with OneAPI is naming. It leads people to believe that is another competing standard where in fact is is simply just an implementation of a standard compliant SYCL compiler.” – swerner

🚀 Project Ideas

Generating project ideas…

ZLUDA Compatibility Hub

Summary

  • Automated testing, submission, and a public database for CUDA‑on‑AMD Windows compatibility across AMD GPUs.
  • Core value: lowers the barrier for users to run CUDA workloads on AMD hardware and accelerates ZLUDA adoption through community‑driven validation.

Details

Key Value
Target Audience Windows developers/researchers with AMD GPUs who want to run CUDA applications
Core Feature CI matrix that builds ZLUDA, runs a benchmark suite (LibTorch, TensorFlow, custom kernels), auto‑detects GPU, logs success/failure, aggregates results into a searchable compatibility table
Tech Stack GitHub Actions (Windows runners), PowerShell/Bash test scripts, Python/Flask for result aggregation, React frontend for UI
Difficulty Medium
Monetization Hobby

Notes

  • HN users like system2 wished to use older RDNA1 cards; this hub would give them concrete compatibility data.
  • Provides a centralized place for users to submit reports (as requested in the repo), fostering discussion and broader testing.

Unified GPU Compute Abstraction Layer (UGUL)

Summary

  • A portable library that lets developers write CUDA‑style kernels and have them compiled to HIP, SYCL, Metal, or Vulkan Compute with automatic fallback.
  • Core value: write once, run anywhere on GPU, weakening the CUDA moat and enabling AMD adoption without rewriting code.

Details

Key Value
Target Audience GPU programmers seeking portable compute code (ML, HPC, graphics)
Core Feature API mirroring CUDA runtime + kernel launch macros; backend selection via env var; uses LLVM‑based translators (ZLUDA/HIP) and SPIRV‑Cross for Vulkan
Tech Stack C++17, LLVM, CMake, optional Rust bindings; SPIRV‑Cross for Vulkan backend
Difficulty High
Monetization Hobby

Notes

  • Commenters like swerner and linuxhansl lament the lack of usable open standards; UGUL offers a pragmatic path forward.
  • Generates discussion about performance tradeoffs and invites contributions to improve each backend, creating a collaborative ecosystem.

LLM‑Assisted AMD Kernel Optimizer (LAKO)

Summary

  • VS Code extension that takes CUDA source, uses an LLM to suggest HIP equivalents optimized for AMD architectures, inserting WMMA/BF16/INT8 workarounds where needed.
  • Core value: bridges the instruction‑set gap on AMD GPUs, making ported kernels performant without deep manual tuning.

Details

Key Value
Target Audience ML engineers and researchers porting models to AMD GPUs
Core Feature LLM‑prompted refactoring generates HIP code with intrinsics, provides benchmarks via ROCm profiler, shows diff and performance estimates
Tech Stack TypeScript/VS Code API, LLM backend (Ollama/OpenAI), Rust HIP compiler, ROCm‑SMI for metrics
Difficulty Medium
Monetization Hobby

Notes

  • HN users monster_truck and system2 complained about missing WMMA/BF16 on RDNA1/2; LAKO directly addresses those gaps.
  • Enables quick iteration and community discussion: users can share optimized kernels and compare performance on different AMD cards.

Read Later