DEEP LEARNING with AMD? Maybe we can....

When it comes to deep learning with GPUs, the narrative is always the same: the crowd applauds 英伟达 ?? for employing its CUDA cores with TensorFlow ??. At some point, every deep learning enthusiast has depended on an NVIDIA GPU to complete tasks. NVIDIA GPUs are necessary for the computations of even well-known online notebooks, such as 谷歌 Colab and Kaggle Notebooks ??.

Even with AMD as a competitor in the GPU industry, NVIDIA is still vastly preferred by users. When it comes to integrated chips, gaming GPUs ??, and other applications, AMD fights shoulder-to-shoulder. But AMD is well behind NVIDIA in the deep learning business ??♂?. Some reasons for this are:

• Nvidia gives CUDA toolkit, easy to setup and use.
• CUDA directly integrates with almost every other frameworks (like TF, PyTorch, etc.)
• Better support from the community

AMD is making every effort (maybe ??) to overtake NVIDIA. It's worth noting that MI300X chips, designed to compete with NVIDIA's H100, were recently released ??. These processors, however, are more suited for data centers than for consumers ??.

AMD's Radeon RX 5xxx, 6xxx, and 7xxx series are aimed for consumers and directly rival NVIDIA's RTX 3xxx and 4xxx series. NVIDIA continues to be preferred by users while offering comparable performance.

Welcome AMD ROCm

With ROCm, an open-source package created for GPU compute, AMD appears to be making progress ???. AMD refers to the Heterogeneous-Computing Interface for Portability (HIP) as something that ROCm makes use of. HIP enables ROCm developers to build portable apps for a variety of platforms, including exascale HPC clusters and specialized gaming GPUs ??.

ROCm is now at version 6.0.2 ??. Although Windows is supported by the HIP SDK ????, Linux appears to be ROCm's native environment. ROCm supports both Linux and Windows platforms. This implies that Windows users may encounter certain difficulties, but Linux users benefit from extensive assistance.

It only supports an extensive (although limited) list of AMD Radeon, Radeon PRO and Instinct GPUs. Read the full list from here: System requirements (Windows) — HIP SDK installation Windows (amd.com).

It's also worth mentioning that developers are attempting to incorporate several libraries into ROCm, such as PyTorch, TensorFlow, JAX, and Magma ???. This suggests that big developments could be on the horizon soon ??.

But you might think you have to wait years to use your AMD GPU for deep learning if your GPU isn't on the supported list ?.

Or perhaps not? ??

DirectML - The Workaround

We appreciate 微软 for launching DirectML! ?? With DirectML, ML tasks can be accelerated on any GPU, which means AMD GPUs can now finally join the party ??. Microsoft recently announced support for PyTorch 2.2 with DirectML ?? in its blog post.

Still, that's not all! Additionally, Microsoft offered examples of how to use DirectML for a variety of machine learning tasks, including transformer training, computer vision tasks, and LLM training (with models like Llama 2, Llama 3, Mistral, Phi 2, and Phi 3 Mini) ?????. All of these examples are available here: PyTorch + DirectML.

PyTorch support is therefore taken care of ?. But what about the user favorite, TensorFlow? ?? Two years have passed since the previous attempt to integrate TensorFlow with DirectML. I will leave this for you to think about... ??

But on a happier side, we have PyTorch working on AMD GPUs with the help of DirectML backend.

Conclusion

I've successfully finished the installation steps and benchmarks are next, but it appears that only the sample projects are operational. ?? It's time to start using PyTorch and create some machine learning applications for a side-by-side comparison. ??????

What comes next? ?? Perhaps we should hold off till some lower-end GPUs are supported by the AMD HIP SDK. Alternatively, AMD might develop a native ROCm application. Alternatively—though it seems unlikely—TensorFlow might provide a DirectML backend. ??

??I'll walk through how to install DirectML with PyTorch in my upcoming blog so you can utilize any GPU you own. ????

Till then, cross your fingers that AMD develops something revolutionary. ??

? Keep checking back! ????