Show HN: Chisel – Profile AMD MI300X kernels locally

TL;DR: A CLI tool for developing AMD GPU kernels on DigitalOcean. Spins up GPU droplets, syncs code, runs/profiles HIP kernels with rocprof, and pulls results back locally. Zero GPU hardware required—just write, test, and profile AMD GPU code from your laptop.

From PyPI (recommended):

Development setup:

With uv:

With venv:

1. Configuration

   • chisel configure - Set up your DigitalOcean API credentials

   Usage:

   # Interactive configuration (recommended for first-time setup) chisel configure # Non-interactive with token chisel configure --token YOUR_DIGITALOCEAN_TOKEN

   Getting your API token:

   1. Visit DigitalOcean API Tokens
   2. Click "Generate New Token"
   3. Give it a name (e.g., "chisel-cli") and ensure it has read and write access
   4. Copy the token immediately (you won't be able to see it again)

   For detailed instructions, see the official guide.

2. Spin up GPU droplet

   • chisel up - Create or reuse a GPU-accelerated droplet

   Usage:

   What it does:

   • Checks for existing 'chisel-dev' droplet
   • If none exists, creates a new droplet with:
     • Size: gpu-mi300x1-192gb (AMD MI300X GPU)
     • Image: AMD AI/ML Ready (ROCm pre-installed)
     • Region: ATL1 (where AMD GPUs are available)
     • SSH keys: Automatically injects all keys from your DO account
   • Waits for droplet to be ready and SSH accessible
   • Displays connection information

3. List droplets

   • chisel list - Show all chisel droplets

   Usage:

   Shows:

   • All active chisel droplets with their status
   • IP addresses for SSH access
   • Region, size, and creation time
   • Current active droplet from local state

4. Sync code

   • chisel sync - Push local files to droplet (only changed files)

   Usage:

   # Sync a file to /root/chisel/ (default) chisel sync simple-mm.cpp # Sync directory contents chisel sync ./src/ # Sync to custom destination chisel sync myfile.cpp --dest /tmp/

   What it does:

   • Uses rsync for efficient file transfer
   • Only transfers changed files
   • Shows progress during transfer
   • Creates destination directory if needed

5. Run commands

   • chisel run - Execute commands remotely with live output streaming

   Usage:

   # Compile and run HIP kernel chisel run "hipcc /root/chisel/simple-mm.cpp -o /tmp/test && /tmp/test" # Run multiple commands chisel run "make && ./bench.sh" # Check GPU status chisel run "rocm-smi"

   What it does:

   • SSH exec with real-time output streaming
   • Returns actual exit codes
   • Handles both stdout and stderr
   • Works with interactive commands

6. Profile kernels

   • chisel profile - Profile commands or source files with rocprof and pull results locally

   Usage:

   # Profile source file (auto-compiles with hipcc) chisel profile simple-mm.cpp # Profile with custom compiler flags chisel profile kernel.cpp --args "-O3 -DNDEBUG" # Profile existing binary chisel profile "/tmp/my-binary" # Profile any command chisel profile "ls -la" # Custom trace options and output directory chisel profile simple-mm.cpp --trace hip,hsa,roctx --out ./results # Auto-open results in Perfetto chisel profile kernel.cpp --open

   What it does:

   • Auto-syncs source files to droplet if needed
   • Compiles source files with hipcc
   • Runs rocprof with specified trace options
   • Downloads profile results to local directory
   • Displays summary of top kernel hotspots
   • Optionally opens Chrome trace in Perfetto

7. Pull artifacts

   • chisel pull - Pull files or directories from the droplet to local machine

   Usage:

   # Pull a specific file to current directory chisel pull /tmp/results.txt # Pull to a specific local path chisel pull /tmp/logs/ --local ./local_logs/ # Pull compiled binary chisel pull /root/chisel/my-binary --local ./my-binary

   What it does:

   • Pulls any file or directory from the remote droplet
   • Automatically handles files vs directories
   • Creates local destination directories as needed
   • Shows progress and confirmation

8. Clean up old droplets

   • chisel sweep - Clean up droplets that have been running too long

   Usage:

   # Clean up droplets older than 6 hours (default) chisel sweep # Custom time threshold chisel sweep --hours 12 # Auto-confirm without prompting chisel sweep --yes

   What it does:

   • Shows all droplets with uptime and estimated costs
   • Identifies droplets running longer than specified threshold
   • Prompts for confirmation before destroying old droplets
   • Clears local state if active droplet was destroyed

9. Stop billing

   • chisel down - Destroy the droplet to stop charges

   Usage:

   What it does:

   • Prompts for confirmation before destroying
   • Completely removes the droplet (not just powered off)
   • Clears local state cache
   • Stops all billing immediately

Cost Management:

• Cost warnings appear when droplets run longer than 12 hours
• Estimated at $1.99/hour for AMD MI300X GPU droplets
• Use chisel sweep to clean up old droplets automatically
• Droplets self-destruct after 15 minutes of inactivity

• AMD's droplets ship with ROCm pre-installed, so driver stack is available instantly
• Use DigitalOcean's pydo to create / destroy nodes

1. Python CLI skeleton – typer or argparse; single main.py.

2. DigitalOcean wrapper

   import pydo client = pydo.Client(token=token) client.droplets.create(size='gpu-mi300x1-192gb', image='gpu-amd-base', ...)

3. SSH/rsync layer – Use paramiko for exec + rsync/scp shell out (simplest); later swap to async libraries if perf matters.

4. Cloud-init script – idempotent bash that:

   • apt update && apt install -y build-essential rocblas-dev …
   • Adds a /etc/profile.d/chisel.sh that exports ROCm paths.

5. State cache – tiny JSON in ~/.cache/chisel/state.json mapping project → droplet ID & IP so repeated chisel run skips the spin-up step.

6. Credential handling – ENV override CHISEL_DO_TOKEN > config file, because CI.

7. Cost guardrails – warn if droplet has been alive >N hours; chisel sweep to nuke zombies.

• concurrent runs (non-blocking sync and run)