ROCm vs CUDA in 2026

[_galaxy]

Was wondering how the state of ROCm is compared to CUDA.

Also take into the account the hacks that you can makes CUDA think ROCm is CUDA

For anything CUDA related that I might need with a AMD card

[Bad 💀 Motha]

AMD ROCm™ Software can enable CUDA applications to run on AMD graphics cards, primarily by acting as a translation layer. While ROCm itself is designed to convert CUDA code to AMD's HIP language, recent developments like ZLUDA allow many CUDA-powered apps to run directly on AMD GPUs with near-native performance.


Key Takeaways for Running CUDA on AMD:

ZLUDA (Open Source): A tool that acts as a drop-in replacement for CUDA libraries on Windows and Linux, allowing apps like Blender and some AI tools to work on Radeon GPUs.

HIPify: A component of ROCm that converts CUDA C++ code into portable HIP C++ code, allowing it to be compiled for AMD GPUs.

Performance: While it enables functionality, performance may vary compared to native NVIDIA hardware, and it works best on high-end cards like the RX 7900 XTX.

Limitations: This approach is not officially supported by NVIDIA and works best on Linux, although ZLUDA has enabled Windows support.

[PopinFRESH]

Comparably priced GPUs running a CUDA workload (e.g. local AI models, etc.) the AMD GPUs are roughly 15% - 30% slower; with that becoming a significantly wider gap going up the product stack.

There are still some significant bottlenecks that using a translation layer doesn't solve where the NVIDIA GPU will accelerate some things that are much slower on an AMD GPU because it lacks the hardware acceleration.

[Haruspex]

I use ROCm to run AI stuff on my 7900 XTX. It gives me pretty good performance over Vulkan, but not as good as if I were using an equivalent Nvidia card.

I am using Linux, which has much better ROCm support than Windows does currently. I managed to get a ComfyUI workspace up and running with ROCm once in Windows 11 with ZLUDA, but it was an enormous pain in the butt. On Linux it's been pretty nice and seamless.

[Caerimonia]

ROCm is an open specification, and Cuda is locked to Nvidia. That alone should have been an incentive to avoid Cuda altogether, but people just love every ensh*ttifying monopoly out there that will make your life worse later on, when you actually depend on them to do your stuff.

There's little that AMD can't do that Nvidia can. Even the supposed performance gap is bogus, since magazines and websites always compare cheaper AMD products with more expensive Nvidia gear, giving the later an advantage. Measured on equal costs, Nvidia actually performs worse in many cases.

[Adrift Illusion]

Vulkan, actually

[BurakZG]

https://youtu.be/BzLaALDc1bw?is=g6yjCrdUPNIa8eCh

Somebody did some comparison on running AI models on 9060XT 16GB and 5060Ti 16GB. Watch if you are interested.

Nvidia is faster but... not always.
It also caused troubles on linux (the guy spent couple of hours making it work).

Is it worth 120EUR more?

Coming to CUDA and ROCm, the thing is, every proprietary technology is with time replaced by open source. Always. It's only a matter of time. Sooner or later, the company that owns technology looses interest in it.
Open source stays open forever.

[Caerimonia]

Originally posted by BurakZG:

Is it worth 120EUR more?

That's what I said above. They insist in comparing products from different price tags, when they should be comparing product lines of roughly equal prices. The Nvidia equivalent to AMD is actually a lower grade card than the one that got into the test. That's how Nvidia has been giving the illusion of performance for so many years. It's very much the same story as game reviews, completely cooked.