AMD recently allowed me some time with their AMD Accelerator Cloud (AAC) leveraging multiple Instinct MI300X accelerators. During this brief opportunity to try out their latest software advancements with the Instinct MI300X and the ROCm compute stack, one of the most striking takeaways was their documentation improvements compared to previous forays into ROCm+Instinct compute. In addition, AMD is now offering more robust container options for easier Instinct compute deployments with more software options available and being more regularly updated.

[1]

Coincidentally it was about a year ago I [2]last had the chance to try out the AMD Accelerator Cloud as AMD's private cloud environment to enable customers, researchers, and others to try out EPYC CPUs and Instinct GPUs in an established environment and with the software stack ready to go. The AMD Accelerator Cloud allows users to try out their own codes with AMD hardware, kick the tires around ROCm, and more while enjoying AMD's streamlined process for launching accelerated compute instances without needing to make any initial time investment setting up the software stack and/or any capital expenditures on hardware. It's also convenient for those simply wanting to benchmark the latest AMD Instinct hardware/software.

[3]

For the first time in a year having access to MI300X accelerators, the ROCm software stack was much improved compared to the beginning of 2024. But that wasn't exactly a surprise given my routine ROCm coverage on Phoronix and tending to cover each and every interesting software change at AMD. Rather my biggest takeaway from this recent testing bout was the much more quality AMD documentation available and their diverse set of containers that are now available for quickly and easily deploying different ROCm-accelerated offerings. It was a night and day difference in these areas compared to a year ago or longer. Another tertiary area of improvement lately has been [4]AMD investing in their own open-source language models .

[5]

With AMD's ROCm container offerings they are also committing to updating them on a bi-weekly basis moving forward, which is a welcome change compared to the all too often world of containers being seldom updated and trailing in their capabilities compared to those that are willing to build their stack from source. Some of the areas they have been focusing on the most outside of ROCm proper has been PyTorch, Megatron-LM, vLLM, and others. They are also emphasizing and focusing both on the training and inference potential for Instinct accelerators. These containers also aren't restricted to just running in the AMD Accelerator Cloud or to special customer environments but intended to work on other public cloud service providers or even on-premise deployments.

[6]

As a sign of the times, just while writing this article AMD announced same-day support for Google's Gemma 3 models as a delight to see it happen same-day as something not expected to see out of AMD years ago. The timely DeepSeek support has been another great example of AMD being more committed to timely supporting new models and embracing the broader software ecosystem.

AMD also [7]announced this past week that they are now supporting the Open Platform for Enterprise AI (OPEA) as a Linux Foundation project that was [8]started in part by Intel .

For those wanting to get up to speed on AMD's documentation improvements and other software enhancements, the gateway to doing so is via their [9]rocm.blogs.amd.com and [10]ROCm AI Developer Hub entry-points. A great resource and hopefully they continue improving this documentation and their web assets at large as it's a big time-saver for those getting started with Instinct accelerators and helping to see how far the ROCm ecosystem has evolved in recent times. Via [11]hub.docker.com and [12]AMD Infinity Hub are also all of the ROCm containers for those wanting to try out these frequently-updated ROCm container images.

Thanks to AMD for providing the gratis access to the AMD Accelerator Cloud. Due to the limited time access on short notice and not being able to arrange remote access to another NVIDIA GH200/GB200 server on that timeline in a similar configuration to the AMD AAC deployment, this round of testing was not geared for performance benchmarking comparisons and just focused on trying out the latest ROCm stack and exploring the documentation and container improvements made by AMD over the past year.



[1] https://www.phoronix.com/image-viewer.php?id=amd-rocm-docs-containers-2025&image=amd_aac_1_lrg

[2] https://www.phoronix.com/review/amd-instinct-mi300x-rocm6

[3] https://www.phoronix.com/image-viewer.php?id=amd-rocm-docs-containers-2025&image=amd_aac_2_lrg

[4] https://www.phoronix.com/news/AMD-Intella-Open-Source-LM

[5] https://www.phoronix.com/image-viewer.php?id=amd-rocm-docs-containers-2025&image=amd_aac_3_lrg

[6] https://www.phoronix.com/image-viewer.php?id=amd-rocm-docs-containers-2025&image=amd_aac_4_lrg

[7] https://rocm.blogs.amd.com/artificial-intelligence/-opea-blog/README.html

[8] https://www.phoronix.com/news/Open-Platform-for-Enterprise-Ai

[9] https://rocm.blogs.amd.com/

[10] https://www.amd.com/en/developer/resources/rocm-hub/dev-ai.html

[11] https://hub.docker.com/u/rocm

[12] https://www.amd.com/en/developer/resources/infinity-hub.html