# CUDA developer — the layer where a model meets the GPU, since 2024

> Running training and inference workloads on the GPU with CUDA through PyTorch, TensorFlow and OpenAI Triton: where I have used it since 2024.

- Highlights: 2 years
- Experience: Since 2024
- Updated: 2026-11-08
- Source: https://muhammetsafak.com/expertise/cuda/
- Language: en-US
- Author: Muhammet Şafak

---
## Me in the CUDA ecosystem

CUDA is not an application area for me; it is the platform layer through which a model reaches the GPU. A training or inference workload is written inside a framework, but how it runs on the GPU decides how much the system can carry. That is why I see CUDA as the infrastructure leg of [AI systems engineering](https://muhammetsafak.com/expertise/ai-systems/).

I reach CUDA through [PyTorch](https://muhammetsafak.com/expertise/pytorch/) and [TensorFlow](https://muhammetsafak.com/expertise/tensorflow/). Next to them are OpenAI Triton for inference optimization, NVIDIA Dynamo-Triton for serving models and Kubernetes for scaling.

CUDA is, for me, the name of the platform underneath the work.

## What is the CUDA Developer role?

CUDA is NVIDIA's platform for accelerated computing and the foundation for GPU computing. A CUDA developer is the person who develops code that runs on the GPU with CUDA. I do not use CUDA by writing CUDA C++ kernels directly; I use it through PyTorch, TensorFlow and OpenAI Triton.

## What does the CUDA Developer do?

I run training and inference workloads on the GPU with CUDA through PyTorch and TensorFlow. For inference optimization I write custom GPU kernels with OpenAI Triton, serve models with NVIDIA Dynamo-Triton and scale these workloads as containers on Kubernetes.

## Technologies I pair it with

- CUDA: The NVIDIA platform I run training and inference workloads on the GPU with.
- PyTorch: The deep learning framework I use for model training, fine-tuning and adaptation with LoRA/PEFT.
- TensorFlow: The deep learning framework I use in my model-training work.
- OpenAI Triton: The language and compiler I use to write custom GPU kernels for inference optimization.
- NVIDIA Dynamo-Triton: The model serving server, formerly NVIDIA Triton Inference Server, that I use to put trained models behind an API.
- Kubernetes: The orchestration platform I use to deploy and scale model serving and inference workloads as containers.

## Frequently asked

### How long have you worked with CUDA?

Since 2024, which is 2 years.

### What does a CUDA developer do?

They develop code that runs on the GPU with CUDA. My scope is narrower: I run training and inference workloads on the GPU with CUDA through PyTorch, TensorFlow and OpenAI Triton.

### How are CUDA, PyTorch and TensorFlow related?

PyTorch and TensorFlow are the frameworks I define and train models in; CUDA is the platform that runs those workloads on NVIDIA GPUs. For the framework side see the [PyTorch](https://muhammetsafak.com/expertise/pytorch/) and [TensorFlow](https://muhammetsafak.com/expertise/tensorflow/) pages.

### Do you write CUDA C++ kernels?

No. I use CUDA through PyTorch and TensorFlow; the custom GPU kernels that inference optimization needs, I write with OpenAI Triton.

### Whose platform is CUDA?

It is NVIDIA's own platform. It is the one I use, and I make no claim about any other GPU platform.

### How do you scale GPU workloads?

I deploy model serving and inference workloads as containers and scale them with Kubernetes.