ROCm 10 Released: Inference Performance Triples by 3.3x, AMD Finally Starts Competing on Developer Experience

On August 28, AMD released ROCm 10. This version coincides perfectly with ROCm's tenth anniversary—the first iteration, ROCm 1.0, was released in April 2016.

Over the past decade, ROCm has evolved from a barely usable CUDA alternative to a platform with a clear direction.

💡 What You Will Learn

ROCm 10 Released: Inference Performance Triples by 3.3x, AMD Finally Starts Competing on Developer Experience On August 28, AMD released ROCm 10. This version coincides perfectly with ROCm's tenth an

📜 Table of Contents

• What is ROCm.AI
• How to Read the Performance Numbers
• What It Means for Regular Users
• But Don't Get Too Excited
• My Take

ROCm 10 Released: Inference Performance Up 3.3x, AMD Finally Starts Focusing on Developer Experience

On August 28, AMD released ROCm 10. This release just happens to coincide with ROCm's tenth anniversary—the first version, ROCm 1.0, was released in April 2016.

Over the past decade, ROCm has evolved from a barely usable CUDA alternative to a project at a clearly defined turning point. This time, ROCm 10's focus is no longer "we've added a few more low-level libraries," but rather a complete development toolchain called ROCm.AI, with a straightforward goal: making it less of a hassle to run AI on AMD GPUs.

What is ROCm.AI

ROCm.AI consists of three components, each handling a different stage.

ROCm CLI is a unified command-line tool. Previously, installing ROCm meant running a bunch of scripts, and one wrong environment variable would bring the whole thing crashing down. Now, a single rocm serve <model-name> command can pull up a model for inference, and rocm examine automatically diagnoses driver and environment issues. It also supports offline environments, allowing dependencies to be packaged into self-contained bundles. However, this is still a technical preview, and the interface will change in the future—don't rely on it as a stable tool just yet.

AMD Skills packages AMD's officially validated ROCm knowledge into skill packs that plug directly into AI coding assistants like Claude, Cursor, and Codex. Previously, if you asked an AI "how to optimize inference on AMD GPUs," it would likely fabricate a CUDA-based solution. Now, these assistants can tap into AMD's official skill catalog, making their answers far more reliable. The skill packs are hosted in the amd/skills repository on GitHub and can be installed with a single command.

Hyperloom is the most ambitious of the three. It's an open-source automated optimization system that runs performance analysis, identifies bottlenecks, modifies code, tweaks GPU kernels, validates results, and then repeats the cycle. AMD claims that optimization work that used to take engineers weeks can now be completed by Hyperloom in hours, and it can explore far more approaches than a human under time pressure would ever dare to try. In short, it's using AI to tune AI workloads.

How to Read the Performance Numbers

The numbers AMD provided: compared to ROCm 7, inference throughput is up 3.3x, and training is up 2.4x.

These numbers need to be broken down.

The test platform was an 8-GPU Instinct MI355X setup, running GLM-5, Kimi-K2.5, and DeepSeek-R1-0528. The tests were conducted on July 7, 2026, by AMD Performance Labs. The data comes from AMD itself, not third-party benchmarks.

More critically, the comparison baseline is ROCm 7, not the previous generation ROCm 9. ROCm 7 dates back to early 2024, with ROCm 8 and 9 in between. The 3.3x figure represents cumulative gains across more than three years, not a single-generation triple from ROCm 9 to 10.

One more thing: the tests used the data-center-grade MI355X. AMD hasn't provided data on how much improvement consumer Radeon GPUs will see. If you're running local LLMs on an RX series card, don't directly apply that 3.3x figure to your own hardware.

What It Means for Regular Users

ROCm 10 supports a broader range of hardware than before: Instinct accelerators, Radeon consumer GPUs, and Ryzen integrated graphics—full product line coverage, with both Windows and Linux support.

This means two things.

First, the barrier to running AI on AMD GPUs is lowering. ROCm CLI simplifies installation and startup, AMD Skills lets AI assistants provide reliable AMD platform advice, and Hyperloom automates the most headache-inducing part—performance optimization. Stacked together, these three represent a tangible improvement for users who "want to run LLMs on AMD GPUs but don't want to deal with environment setup."

Second, AMD's strategy has shifted. In past years, ROCm's approach was "whatever CUDA has, we'll add it too," chasing low-level libraries one by one, but developer experience remained a weak spot—installation failures, things not running, poor performance; these three issues drove many people away. ROCm 10 is no longer just competing on library count; it's competing on the smoothness of the entire workflow. This direction is correct.

But Don't Get Too Excited

A few practical issues need to be addressed.

ROCm CLI is still a technical preview, Hyperloom was just open-sourced, and AMD Skills depends on the maturity of third-party AI assistants' skill ecosystems. All three are still in early stages, far from "plug-and-play."

Consumer GPU support has always been ROCm's Achilles' heel. While AMD officially claims full Radeon series support, in practice, many users still encounter driver compatibility issues, unsupported models, and performance that falls short of expectations. Whether ROCm 10 improves this remains to be seen through community testing.

There's also a fundamental issue: CUDA's ecosystem advantage can't be closed through toolchains alone. Nearly all AI frameworks and models default to CUDA support first, with ROCm always playing catch-up. ROCm.AI can lower the barrier to entry, but it can't change the reality that "CUDA versions come first, ROCm versions follow."

My Take

ROCm 10 represents a strategic pivot in AMD's AI software stack.

From ROCm 1 to 9, AMD was filling in the low-level gaps, chasing "what CUDA can do, I can do too." ROCm 10 is now chasing "easier to use than CUDA." This pivot itself matters more than the 3.3x performance figure—performance can be stacked through hardware iteration, but the developer experience gap is the fundamental reason users don't stick around.

For those already running AI on AMD GPUs, ROCm 10 is worth upgrading to try, especially ROCm CLI and Hyperloom. For those still on the fence, debating between an Nvidia card and an AMD card, ROCm 10 is a positive signal, but it's not yet a compelling enough reason to switch camps. It's worth waiting for real-world consumer GPU benchmarks before making a final call.