CUDA, ROCm and OpenVINO are the software layers that let AI code use NVIDIA, AMD and Intel hardware respectively. CUDA is NVIDIA's platform and the default target for almost all AI software, so things work first and best there. ROCm is AMD's open-source stack for Radeon and Instinct GPUs; it has improved quickly and now runs on Windows as well as Linux. OpenVINO is Intel's toolkit for optimising and running models on Intel CPUs, integrated and Arc GPUs and NPUs, with a focus on inference. For most teams the choice follows the work: CUDA for training and anything cutting-edge, ROCm where AMD's price per gigabyte of VRAM is attractive and the software is confirmed to work, and OpenVINO for inference on the Intel PCs you already own.
This post looks at developer compatibility. For the business risk of depending on one vendor, see is CUDA lock-in becoming a business risk?
CUDA vs ROCm vs OpenVINO at a glance
| NVIDIA CUDA | AMD ROCm | Intel OpenVINO | |
|---|---|---|---|
| Hardware | NVIDIA GPUs, from Jetson to data centre | AMD Radeon and Instinct GPUs, Ryzen AI | Intel CPUs, integrated and Arc GPUs, NPUs |
| Licence | Proprietary, free to use | Open source | Open source |
| Main strength | Everything supports it; training and inference | Open stack; large VRAM for the money | Efficient inference on common PCs |
| Training | Yes, the standard | Yes, via PyTorch | Not its purpose |
| Operating systems | Linux, Windows | Linux, Windows | Linux, Windows, macOS |
| Typical friction | Cost and supply of hardware | Some tools lag or need specific versions | Models need conversion; fewer large-model options |
Why is CUDA the default?
CUDA has been available since 2007, and nearly two decades of libraries sit on top of it: cuDNN for neural networks, TensorRT for inference, and the wider CUDA-X collection for data science, simulation and more. Researchers publish code that targets CUDA first. New model formats, such as the FP4 types on Blackwell GPUs, appear there first; see FP4 vs FP8 vs INT8 vs INT4. Production engines such as TensorRT-LLM and NVIDIA NIM are NVIDIA-only. If your team wants the least time spent on tooling, CUDA is the safe answer.
Where does ROCm stand now?
Much closer than a few years ago. AMD's ROCm 7 generation added official Windows support, extended the stack to Radeon GPUs and Ryzen AI processors, and added integration with serving frameworks including vLLM and SGLang. At CES 2026, AMD described ROCm 7.2 as a common release for Windows and Linux. PyTorch runs on it, and popular local tools such as Ollama, LM Studio and ComfyUI support AMD GPUs.
The friction that remains is in the long tail. A new research repository, a niche library or a specific optimisation may assume CUDA, and you may need a particular ROCm and framework version to make things work. Check the exact GPU model against AMD's compatibility list, because support varies by card and operating system.
Source: AMD ROCm announcements and press coverage (2025 and Jan 2026). Compiled 11 Nov 2026.
What is OpenVINO for?
Inference on Intel hardware. OpenVINO converts and optimises a trained model, then runs it on whatever Intel devices are present: the CPU, the integrated or Arc GPU, or the NPU in a recent Core Ultra PC. It is strong for vision models, speech, document processing and smaller language models, the kind of AI that runs inside business applications on ordinary PCs and edge boxes. It is not a training platform, and it is not where you run a 70-billion-parameter model. We show Intel-based local AI in Ollama on Intel Arc and compare the approaches in NUC 16 Pro vs RTX Spark.
You rarely choose CUDA, ROCm or OpenVINO directly. You choose a framework and a tool, and they choose for you. Check that the software you plan to run supports the hardware before you buy the hardware.
How much does the ecosystem matter in daily work?
- If you use PyTorch and mainstream tools, less than it used to. PyTorch has backends for CUDA, ROCm, Intel and Apple silicon, and most code runs with a device setting changed.
- If you use local model runners, very little. Engines based on llama.cpp support all three vendors and Apple's Metal.
- If you follow new research or need peak performance, a lot. Custom kernels, the newest attention optimisations and the fastest serving engines arrive on CUDA first.
- If you deploy to many ordinary PCs, OpenVINO or ONNX Runtime matters more than CUDA.
Apple's MLX is a fourth ecosystem worth knowing if your team uses Macs; see Apple's opportunity in local AI. Windows on Arm PCs add their own compatibility questions, covered in our RTX Spark app compatibility guide.
How do you stay portable?
- Write to a framework, such as PyTorch, not to a vendor API.
- Serve models behind an OpenAI-compatible API so the hardware underneath can change.
- Keep models in open formats, and export to ONNX where it helps.
- Use containers so the driver and library combination is captured; see Docker vs Kubernetes for AI.
- Before buying non-NVIDIA hardware for a specific job, run that job on a loan or cloud instance first.
Which should a business choose?
| Need | Sensible choice |
|---|---|
| Training, fine-tuning, newest models, least friction | NVIDIA and CUDA |
| Local inference on a budget, with tools confirmed to support AMD | AMD and ROCm |
| AI features on existing office PCs and edge devices | Intel and OpenVINO |
| Production serving with vendor support | NVIDIA |
| Robots and embedded vision | NVIDIA Jetson, which uses the same CUDA stack |
Buying in Singapore
For CUDA, the NVIDIA RTX PRO 4000 Blackwell (S$5,640), RTX PRO 6000 Blackwell Workstation Edition (S$29,100) and the 128 GB NVIDIA DGX Spark 4TB (S$9,450), with the Jetson Orin Nano Super 8GB (S$696) at the edge. For ROCm, the 24 GB PowerColor Red Devil Radeon RX 7900 XTX (S$1,955) (listed 11 Nov 2026; prices change; ask for a current quotation). For OpenVINO, ask us about current Intel Core Ultra mini PCs and laptops. To start, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
What is the difference between CUDA and ROCm? CUDA is NVIDIA's proprietary GPU computing platform and the default target for AI software. ROCm is AMD's open-source equivalent for Radeon and Instinct GPUs, with growing but less complete software support.
Does ROCm work on Windows? Yes. AMD added official Windows support in the ROCm 7 generation and described ROCm 7.2 as a common Windows and Linux release. Supported GPUs vary, so check AMD's compatibility list.
Is OpenVINO an alternative to CUDA? Only for inference. OpenVINO optimises and runs trained models on Intel CPUs, GPUs and NPUs. It is not designed for training large models.
Can I switch from NVIDIA to AMD later? Often, if your code targets PyTorch and your models are served behind a standard API. Custom CUDA code and NVIDIA-only tools such as TensorRT make switching harder.
