Soham Kulkarni
ReportIndependent FOSS maintainer and researcher working on ROCm backporting, GPU benchmarking, and reproducible Linux GPU-compute workflows for older AMD hardware. The project is documented publicly through a Substack article series and code assets on GitHub. The expected project outcome is a generalisable pipeline for recovering ROCm functionality on unsupported AMD GPUs, along with benchmarked evidence, failure-mode documentation, and CI/CD automation for community use. Substack: https://netherquark.substack.com/
Generalisable Pipeline for porting ROCm
Convert ad hoc community workarounds into a reproducible pipeline for older AMD GPUs, starting with gfx90c (Ryzen 5 5650U / Cezanne Vega 7) and validating the process on gfx902 (Ryzen 5 2200G / Raven Ridge Vega 8). Completed milestone: Failure analysis of benchmark results from running 113 tests from rocm-examples on gfx90c. Pending: Compile rocm-examples from source and compare against Fedora-packaged binaries; Resolve each failing case for rocm-examples; Port gfx900-targeted assembly / kernels to gfx90c where needed; Write explainer articles for GPU internals. ie. GPU assembly, Vega ISA, SGPRs, scratch, memory, scheduling, LDS, LLVM AMDGPU, etc; Update rocBLAS library for gfx90c by adding the missing kernels; Generate Tensile data for gfx90c; Update any other incompatible libraries to add compatibility for gfx90c; Compile ROCm 6.4.4 with patched components and ensure native HIP execution on gfx90c; Port HIPify/runtime fixes for gfx90c; Benchmark real-world HIP workloads natively on gfx90c vs spoofing gfx900; Compare native gfx90c, Mojo, ZML, and other universal runtimes; Port PyTorch Docker images to gfx90c by modifying bundled ROCm; Port native PyTorch to gfx90c by modifying bundled ROCm; Repeat the full process on gfx902; Evaluate TheRock as a cleaner build path Test whether bleeding edge ROCm 7.1 can be ported; CI/CD pipeline for automatic ROCm / PyTorch / Dockerized PyTorch builds
Monthly stipend
Support for sustained development effort across the full duration of the project
Hardware purchase for upstream contribution
Funding for a RDNA4 GPU needed for upstream ROCm contribution. "AMD RX 9060 XT" is the most suitable model, there are two variants, 8GB and 16GB. The 16GB variant can run larger AI models, making it more useful in the long run. If it is not feasible for the foundation to fund the 16GB variant, I would like to use a part of the first month stipend to cover the difference, and proceed with the 16GB variant.
CI/CD and artifact hosting
This will be required at the end stages of the project. This funding is for cloud infrastructure required to run automated build pipelines for ROCm and PyTorch. This includes CI/CD runners, storage for build artefacts, and bandwidth for distributing binaries. Necessary if free GitHub Actions credits aren't enough.