π Kernel Olympics
Autonomous GPU Migration Platform β Starting with CUDA β ROCm. Designed for a future of portable GPU computing.
Built during the π AMD Developer Hackathon 2026 β An AI-powered platform that helps developers move CUDA applications to AMD ROCm using multi-agent reasoning.
π Why Kernel Olympics Exists
The biggest challenge preventing organizations from adopting AMD GPUs isn't hardware. It isn't performance. It isn't software quality.
It's migration.
Thousands of CUDA applications remain locked to the NVIDIA ecosystem because migrating production GPU software is expensive, risky, time-consuming, and requires highly specialized expertise.
Kernel Olympics changes that.
Instead of acting as another AI coding assistant, Kernel Olympics functions as an Autonomous GPU Migration Platform that understands an entire CUDA project, plans the migration, performs intelligent code transformations, verifies correctness, benchmarks performance, explains every change, and produces a production-ready migration report.
Reduce GPU migration from weeks of engineering work to an AI-assisted workflow that developers can trust.
The $10B Problem
AMD GPUs (MI300X) outperform NVIDIA on price/performance. Yet enterprises stay on NVIDIA because 20% of CUDA code won't port to ROCm β custom kernels, warp-sensitive logic, library-specific calls. hipify handles the easy 80%. The remaining 20% is a manual, weeks-long slog per project.
AMD's #1 adoption blocker isn't hardware β it's software migration friction.
The broader market is bigger: GPU architectures multiply (NVIDIA CUDA, AMD ROCm, Intel oneAPI, Apple Metal, custom NPUs) while the talent pool doesn't. Every hardware generation creates a $2B+ migration tax across the industry β teams rewriting kernels by hand instead of building new products.
β¨ What Makes Kernel Olympics Different?
Kernel Olympics is not:
β another AI chatbot
β another GitHub Copilot
β another wrapper around hipify
β simple prompt engineering
Instead, Kernel Olympics behaves like an experienced GPU engineering team. Multiple specialized AI agents collaborate to understand an entire repository before making any modifications. Rather than translating files one by one, the platform reasons about architecture, dependencies, compatibility, performance implications, unsupported APIs, testing strategy, documentation, and migration risks.
Every decision is transparent. Every modification is explainable. Every migration produces evidence.
π Repository Intelligence
Kernel Olympics doesn't process kernels independently. Before modifying a single line, the platform analyzes the entire project context to identify:
CUDA kernels and their call sites
Runtime API usage and device memory operations
Build configuration and dependencies
Shared utilities and include hierarchy
Unsupported CUDA features and migration complexity
Instead of blindly translating files, the platform understands the codebase as a whole β which dramatically improves migration quality.
π§ The Problem
Today, migrating GPU software is difficult because developers must manually:
Understand large CUDA codebases
Identify unsupported APIs
Rewrite kernels
Replace memory management
Verify correctness
Debug compilation failures
Benchmark performance
Write migration documentation
Even experienced GPU developers spend days or weeks doing this. The process is repetitive, error-prone, expensive, and hard to scale.
π‘ Our Solution
Kernel Olympics introduces an AI-native migration workflow. Instead of treating migration as file conversion, the platform treats it as an engineering reasoning problem. The system first understands the repository, identifies architectural patterns, analyzes dependencies, detects unsupported CUDA features, proposes migration strategies, executes intelligent transformations, validates results, benchmarks performance, and generates a comprehensive migration report explaining every decision.
This creates a migration pipeline that is explainable, repeatable, and significantly easier for developers to trust.
π¬ Demo
π Live: kernel-olympics-production.up.railway.app β Upload a CUDA kernel and watch the pipeline port it live.
π‘ Real AMD MI300X via notebooks.amd.com/hackathon β CUDA source β 4-LLM loop ports β hipcc compile β GPU run β PASSED β
π Quick Stats
Metric Value Pipeline budget 1,800s (30 min) Max iterations 10 (compile-fix loop) LLM cost per run ~$0.09 Cache hit speed ~0.2ms Tests 665 passing CI/CD β
Automated (GitHub Actions) Hardware target AMD MI300X via notebooks.amd.com/hackathon (192GB HBM3, CDNA3)
π§ Multi-Agent Architecture
π¬ Explainable AI
Every modification includes reasoning. Nothing is hidden. Every decision is transparent.
Instead of asking "What changed?", developers see why every change was made β making migration reviewable, auditable, and trustworthy.
π¬ Key Innovation: Pattern Memory Cache
Instead of calling expensive LLMs for every kernel, we cache porting patterns as trigram vectors:
Metric Without Cache With Cache Speedup Pattern lookup N/A 0.2ms β LLM call (simulated) ~12s 0.2ms 60,000Γ Verified with live API β β measured β
πΊοΈ Future Roadmap
The long-term vision extends far beyond CUDA β ROCm. Future versions can support:
CUDA β ROCm (β
currently supported)
CUDA β SYCL
CUDA β Vulkan Compute
CUDA β OpenCL
CUDA β Metal
CUDA β DirectML
The platform is designed as a universal GPU migration engine rather than a single-purpose converter. Adding new target paths is a configuration change, not a rewrite.
π Quick Start
Live Demo (no install required)
β kernel-olympics-production.up.railway.app β Upload a
.cu file, see the autonomous pipeline port it to HIP in real time.Local Setup
π For Judges: Running on AMD GPU
Full guide:
AMD_STACK_USAGE.md β but here's the 30-second version:Prerequisites
AMD GPU (MI300X, MI250, RX 7900 XTX) with ROCm 6+ installed
No AMD hardware? Use the AMD AI Notebooks portal β free MI300X instances for hackathon participants
hipcc in PATH (hipcc --version)Python 3.11+, git
Quick Run
What to Expect
Stage Time Output Port (4-LLM loop) 1β5 min HIP code in
ported_kernels/ Compile (hipcc) 2β5 sec Binary in build/ Run on AMD GPU 1β3 sec Numerical output Verify 0.5 sec PASS β or FAILA successful run ends with:
Without an AMD GPU
The pipeline still ports + verifies. Just skip compile/run:
Verify the Setup
π Makefile Targets
Target Description
help Show all available targets install Create venv + install deps port Pipeline on one kernel: make port CU_FILE=path.cu port-all Pipeline on ALL sample kernels compile hipcc proof harness + compile run Run compiled binary on AMD GPU pipeline Full cycle: port β compile β run pipeline-heavy Extended budget (1,800s) test Run 665 pytest tests demo Live demo with recording inspect Inspect spec/ported kernel/proof debug-kernel Interactive kernel explorer retry Re-run a single pipeline stageπ Debug Mode
Three levels of debugging for when things go wrong:
π§ͺ Running Tests
π§ Pipeline Architecture Details
CUDA β HIP Transformations
CUDA Intrinsic HIP Equivalent Action
__shfl_up_sync(mask, val, d, w) __shfl_up(val, d, width) Mask dropped cudaMalloc() hipMalloc() 1:1 rename cudaMemcpy() hipMemcpy() 1:1 rename findCudaDevice() hipGetDevice() SDK strip sdkCreateTimer() β Removed (NOP) threadIdx.x hipThreadIdx.x Namespace addKnown Issues Handled
NVIDIA SDK symbols β auto-detected and stripped by
verifier.pyWave64 divergence β
warpSize constant used instead of hardcoded 64SIGSEGV from host-code symbols β sanitizer in verifier catches at compile time
π₯ Team
Role Member Focus π Lead indradev_ Architecture, pipeline orchestrator βοΈ Infra cation CI/CD, AMD cloud, Jupyter integration π¬ Kernel Bromine185 CUDA kernel analysis, warp primitives π Docs _dD Documentation, demo recording π§ Infra icodemun44 Tooling, automation π§ͺ CI meteorite67 GitHub Actions, test suite π AMD Aahil-Riyaz (Satoru) AMD MI300X testing, ROCm debugging
π License
MIT
Built for the AMD Developer Hackathon ACT II Β· Track 3 (Open Innovation)
1 like Β· 6 views Β· Jul 13, 2026
Developed AI platform for CUDA to ROCm migration, reducing migration complexity and cost.
Timeline
Jul 8, 2026 - Jul 13, 2026