Kernel Profiling: CUDA vs Triton vs Helion

A systematic comparison of GPU kernel performance across three frameworks on NVIDIA H200

---

This repository contains the complete source code and raw Nsight Compute profiling data for vector addition, batched matrix multiplication, and softmax — each implemented identically in CUDA, Triton, and Helion. Every kernel is a standalone script you can run, profile, and compare.

Built during our internship at the Red Hat PyTorch Team (Bangalore, India) as part of ongoing research into GPU kernel profiling with NVIDIA Nsight Tools.

---

Highlights

	CUDA	Triton	Helion
Vec-add (memory-bound)	36–68% DRAM BW	86–92% DRAM BW	86–92% DRAM BW
MatMul (compute-bound)	10–100x slower (no TC)	Near-peak (autotuned)	Near-peak (autotuned)
Softmax fwd (FP16)	Moderate	93.5% compute	Competitive
Mixed precision speedup	Minimal	10–30x via tensor cores	10–30x via tensor cores

---

Quick Start

git clone https://github.com/XAheli/Kernel-Profiling.git
cd Kernel-Profiling
pip install -r requirements.txt

# Run a kernel (correctness check against PyTorch)
python matmul/triton/kernel/matmul_triton_configG.py

# Profile it
ncu --set full -o matmul/triton/results/matmul_triton_configG \
    python matmul/triton/kernel/matmul_triton_configG.py

# View results
ncu-ui matmul/triton/results/matmul_triton_configG.ncu-rep

Requirements: Python 3.10+ | CUDA 12.8+ | NVIDIA GPU (tested on H200, 141 GB HBM3e)

---

What's Inside

<operation>/
├── cuda/
│   ├── kernel/       # Python + inline CUDA C++
│   └── results/      # .ncu-rep reports
├── triton/
│   ├── kernel/       # @triton.jit kernels
│   └── results/
└── helion/
    ├── kernel/       # @helion.kernel DSL
    └── results/

Operation	What it tests
vec-add/	Pure memory throughput — element-wise add on large 1-D buffers
matmul/	Compute throughput — batched GEMM with tiling and tensor cores
softmax/	Mixed workload — online 2-pass softmax (forward + backward)

---

File naming: <op>_<framework>_<config>.py with matching .ncu-rep results.

---

Related Resources

• NVIDIA Nsight Compute — Getting Started
• Triton Kernel Profiling with NVIDIA Nsight Tools — Red Hat Emerging Technologies Blog
• Full Research Report (PDF)

---

_{Built with care at Red Hat · PyTorch Team · Bangalore, India}