Experiment 0001 — Triton vector add on a free T4

Status: prepared, not yet run on T4.

Purpose

Close the first real lab loop after Lesson 0001:

1. Run a simple Triton kernel on an NVIDIA GPU.
2. Check correctness against PyTorch.
3. Capture a small benchmark against torch addition.
4. Record what broke or surprised me.

This is not meant to prove performance skill yet. It proves the workflow: local notes → free GPU run → captured output → learning record.

Target environment

• Colab or Kaggle notebook
• NVIDIA T4 preferred, but any CUDA GPU is acceptable for the first loop
• Python 3
• PyTorch with CUDA
• Triton

How to run

In Colab/Kaggle, upload or paste vectoraddbenchmark.py, then run:

python vector_add_benchmark.py

If Triton is missing:

pip install triton
python vector_add_benchmark.py

What to capture after running

Paste the terminal/notebook output below.

TODO: paste actual output here after T4 run.

Notes after run

• TODO: What GPU did I get?
• TODO: Did Triton import cleanly?
• TODO: Did correctness pass?
• TODO: Was Triton faster or slower than torch for this tiny benchmark?
• TODO: What did I misunderstand from Lesson 0001?

Link to learning record

• Record 0001 — Triton program instance recall