AMD's Ten-Year Bet to Beat CUDA

AMD's open-source AI software stack turns ten this year, and the anniversary release is the company's clearest admission yet that chips alone were never going to be enough.

Ten years ago, AMD open-sourced a GPU compute stack that almost nobody outside a handful of research labs had heard of. On August 27, 2026, the company shipped its tenth major version, AMD ROCm 10, and used the milestone to make a broader claim: that the software gap which has dogged its AI accelerators for a decade is finally closing. The headline feature, ROCm.AI, is now generally available, and AMD says a system running it delivers roughly 3.3 times the inference throughput and 2.4 times the training throughput of ROCm 7 on the same hardware.

Those numbers are AMD's own, produced by AMD Performance Labs, and they should be read that way. But the release itself — a new build system, a faster ship cadence, and a set of tools explicitly built for AI coding agents — is a genuine attempt to fix the part of AMD's AI business that has mattered more than its silicon specs: whether developers can actually get Instinct GPUs to run production workloads without a fight.

What actually shipped

ROCm is AMD's answer to Nvidia's CUDA — an open-source stack of drivers, compilers, libraries and tools that lets developers run AI and high-performance-computing workloads on AMD GPUs. Like CUDA, it is free to use; AMD makes money selling the hardware underneath it, not the software itself.

The version number is the first tell that something changed. AMD's newsroom confirms that ROCm 10 succeeds a 7.x line that had become a versioning mess: ROCm 7.2 was the last stable release while 7.9 through 7.14 shipped as technology previews, according to Phoronix, which has tracked the ROCm release train closely. Rather than continue that confusing sequence, AMD jumped straight from 7.14 to 10.0, using the number to mark the stack's tenth anniversary and to draw a line under the old build process.

That old process is also gone. ROCm 10 is built end to end on TheRock, AMD's newer open-source build and packaging system that reached production status with ROCm 7.14 in July. Phoronix reports that TheRock replaces a more fragmented set of build and packaging workflows that had made ROCm notoriously harder to install and reproduce than CUDA. AMD is also moving to a six-week release cadence going forward, shipping features, performance work and hardware support in smaller, faster increments rather than in occasional large jumps — a change Phoronix reported AMD committed to earlier this year.

The centerpiece, though, is ROCm.AI, which AMD first previewed at its Advancing AI 2026 conference in July and made generally available with this release. According to AMD's own announcement, ROCm.AI combines three pieces:

• ROCm Hyperloom is an autonomous agentic system that optimizes inference workloads across both host code and GPU kernels. It profiles a workload, finds bottlenecks, tries optimizations, benchmarks the results and validates that outputs are still correct — the kind of tuning that previously required a specialist AMD or customer engineer. With ROCm 10, Hyperloom's support expands across AMD Instinct GPUs with vLLM and SGLang, and it can target HIP, Triton and FlyDSL code paths.
• AMD Skills packages curated AMD expertise into skill files that plug into AI coding agents developers already use — Claude Code, Cursor and Codex, plus an open catalog on GitHub. The idea is that a developer working in one of those tools gets AMD-specific guidance (how to install ROCm correctly, how to diagnose a broken kernel, how to serve a model on an Instinct GPU) without leaving their normal workflow.
• ROCm CLI, shipping as a technology preview, is a single command-line interface for installing, managing and running ROCm environments — including side-by-side runtime versions with rollback, integrated model serving through adapters for Lemonade (client devices) and vLLM (Instinct GPUs), and a real-time monitoring console, formerly called "dash," that reports GPU utilization, power draw and tokens per watt.

AMD frames the payoff as measurable: "through AI-driven optimization of kernels, memory management and scheduling, a system configured with ROCm.AI delivers an average 3.3x inference improvement and 2.4x training improvement over ROCm 7 on the same hardware," per the company's announcement. The fine print matters here. The inference figure comes from AMD Performance Labs testing an 8-GPU Instinct MI355X system running three models — GLM-5, Kimi-K2.5 and DeepSeek-R1 — comparing ROCm 7.0 against a preview build of ROCm.AI on ROCm 7.2.2. The training figure comes from the same hardware running Megatron-LM on DeepSeek-V2-Lite, DeepSeek-V3-16B and Qwen3-30B-A3B. Both are AMD's own benchmarks, on AMD's chosen models and configuration, and have not been independently reproduced as of this writing.

On Windows, AMD retired its standalone HIP SDK in favor of the same ROCm Core SDK used on Linux, syncing both platforms to the same release cadence. A native Windows installer is planned for later in 2026; for now, Windows ships as a static package.

None of this costs anything. ROCm, like CUDA, is free software; the business model is selling the Instinct GPUs and EPYC processors it runs on. The people who should care are AI infrastructure engineers already running or evaluating AMD Instinct hardware — a market surveyed in Edgewisely's rundown of the leading AI chip makers — cloud providers weighing AMD capacity against Nvidia, and any developer whose framework — PyTorch, vLLM, SGLang — has a ROCm backend that was previously a second-class citizen next to CUDA.

The gap ROCm 10 is trying to close

For most of the last decade, AMD's AI accelerators have had a consistent problem: the hardware specs looked competitive, but the software didn't hold up under real workloads. VideoCardz and other outlets covering the ROCm 10 launch have framed it explicitly as AMD's attempt to answer that history — kernels that fell back to slow generic code paths, frameworks that ran but underperformed their CUDA equivalents, and installation processes that were a project in themselves. Nvidia's CUDA, by contrast, has a fifteen-year head start, a deeper library ecosystem, and defaults in nearly every AI framework and tool.

That gap has had real commercial consequences. SemiAnalysis, which has tracked AMD's software progress closely and critically since 2024, reported that Microsoft dropped AMD's MI300X in 2023 over unreliable memory and software issues, and subsequently skipped both the MI325X and MI355X generations — only returning with a commitment to deploy AMD's newer MI455X "Helios" racks, reportedly aimed largely at serving OpenAI's compute needs. The same analysis noted Anthropic has committed to deploying roughly 2 gigawatts of AMD chips, a sign that at least some large AI labs see the software stack as usable enough to bet infrastructure on, even as the analysts flagged unstable internal test clusters as a continuing risk to AMD's pace of improvement. None of that is AMD's own claim — it is outside reporting on how customers have actually behaved, and it is the more reliable signal of whether ROCm's problems are actually fixed.

HPCwire, covering the ROCm 10 release itself, corroborated the core technical details AMD announced — the ROCm.AI components, the TheRock build system, and the performance claims — without independently verifying the benchmark numbers, which is the appropriate level of skepticism for a vendor's own testing.

Who should care, and why

For AI infrastructure teams already running Instinct hardware, ROCm 10 is a straightforward upgrade worth testing: a faster, more reliable build system and a unified CLI address real operational pain points, independent of whether the 3.3x number holds up in a given workload. For teams evaluating AMD for the first time, the calculus is different — AMD Skills and Hyperloom lower the expertise bar to get started, but the SemiAnalysis reporting on Microsoft's earlier retreat is a reminder that "generally available" software has burned enterprise customers before, and due diligence on your specific model and framework combination still matters more than AMD's benchmark slide.

For AMD itself, ROCm.AI and the six-week cadence are a bet that developer experience, not raw FLOPs, is the remaining constraint on Instinct adoption. Packaging AMD's own engineering knowledge into Claude Code, Cursor and Codex skills is a tacit acknowledgment that most developers will never read a ROCm troubleshooting guide — they will ask an AI agent, and AMD wants that agent to already know the right answer.

For Nvidia, the competitive read is more mixed than AMD's framing suggests. CUDA's advantage was never just raw performance; it was the compounding effect of being the default for a decade of tooling, documentation, and institutional knowledge. A faster release cadence and better agent integration narrow that gap incrementally, but they do not erase fifteen years of ecosystem lock-in, and Nvidia continues to ship its own optimizations and invest in its own open-source tooling on a comparable or faster timeline. For cloud providers weighing Instinct capacity against Nvidia GPUs — including governments and regions actively trying to diversify away from a single chip supplier — better developer tooling reduces the operational cost of running AMD alongside Nvidia, which matters most for buyers already committed to that diversification, not for anyone starting from zero.

Takeaways

ROCm 10 is a real, substantive release: a cleaner build system, a faster and more predictable cadence, and a genuine attempt to meet developers where they already work — inside AI coding agents rather than in standalone documentation. The 3.3x and 2.4x performance figures are worth watching for independent verification, but they are not the most important part of this release. The build system change and the release cadence shift are more consequential for whether AMD can sustain the pace of improvement that closing a software gap actually requires. And the customer behavior reported independently — Microsoft's return after walking away, Anthropic's capacity commitment — is a better signal of software maturity than any AMD-run benchmark, because it reflects money changing hands under real production risk.

Zoom-out

Software moats do not fall because a challenger ships a good release; they erode because a challenger ships good releases, repeatedly, for long enough that customers stop asking whether the alternative works and start asking whether it's better. AMD has spent ten years learning that lesson the hard way. ROCm 10 is the clearest evidence yet that the company understands the assignment — closing the gap now depends less on any single benchmark and more on whether AMD can keep shipping every six weeks without the wheels coming off.

Frequently Asked Questions

Is ROCm free to use? Yes. ROCm, including ROCm.AI, is open-source software and free to download and use, similar to Nvidia's CUDA. AMD's revenue comes from selling the Instinct GPUs and EPYC processors the software runs on, not from licensing ROCm itself.

What is ROCm.AI, exactly? ROCm.AI is AMD's AI-native developer experience layer, combining three tools: ROCm Hyperloom (an agentic system that automatically optimizes inference workloads), AMD Skills (curated AMD expertise built into AI coding agents like Claude Code and Cursor), and ROCm CLI (a unified command-line tool for installing and running AI workloads on AMD hardware).

Are AMD's 3.3x and 2.4x performance claims independently verified? No, not as of this writing. The figures come from AMD Performance Labs' own testing, comparing a preview build of ROCm.AI against ROCm 7 on the same AMD Instinct MI355X hardware, using specific model configurations AMD chose. Independent benchmarks from third parties have not yet reproduced these results.

Does ROCm 10 mean AMD has closed the gap with Nvidia's CUDA? Not fully. ROCm 10 addresses real, longstanding pain points — build reliability, installation friction, and developer tooling — but CUDA's advantage rests on roughly fifteen years of ecosystem depth. Independent reporting suggests some large customers are cautiously increasing their bets on AMD, but the gap is narrowing, not closed.