Somebody Fixed Decade-Old Radeons and It Wasn’t an AI Agent

At XDC 2026, Valve’s Timur Kristóf essentially closed the book on a year of work: Southern Islands and Sea Islands Radeons now default to the modern amdgpu kernel driver, with Vulkan, working display support, and a GPU reset path that actually does something useful. That’s the summary. No launch event, no waitlist, no “introducing” blog post with a gradient background.

I review AI tools for a living, which means I spend most of my week reading announcements that promise to change how software gets built. So I want to sit with what Kristóf did, because it’s the exact kind of work that every autonomous coding agent demo pretends to be doing and almost never is.

What actually happened

Old AMD GPUs — the GCN 1.0 and 1.1 generation, cards and APUs that are roughly a decade old — were stuck on the legacy radeon driver. Legacy means maintained in the sense that a vacant lot is landscaped. The modern amdgpu driver is where the performance work, the display features, and the Vulkan support live.

Moving that hardware over wasn’t a flag flip. Kristóf spent the past year on it, and the list of what he had to repair reads like a tour of every unglamorous corner of a graphics stack:

• Display code defects that had to be tracked down and fixed
• Power management problems on hardware nobody’s been profiling in years
• Soft reset support, so a hung GPU doesn’t take the session down with it

The payoff is about a 30% performance boost for those GCN 1.0/1.1 era cards. Thirty percent, on silicon that was already written off. And he’s not done — there’s an ongoing patch targeting initialization of the UVD block on older hardware, which is the sort of detail you only find if you’re genuinely still looking.

Why this matters to anyone shopping for AI tools

I keep a mental list of tasks I use to evaluate coding agents, and this category sits at the top: long-horizon maintenance in a large, old, poorly documented codebase where the failure modes are hardware-specific and the feedback loop requires physical machines.

Nothing I’ve tested comes close. Agents are decent at bounded, legible problems. Write the test, rename the thing, scaffold the endpoint, translate the function. They fall apart when the problem requires holding a year of context, knowing which of three plausible explanations is the real one because you remember a mailing list thread from 2015, and having the judgment to decide that a decade-old video decode block is worth one more patch.

That last part is the piece the tooling conversation keeps skipping. Kristóf’s work isn’t impressive mainly because it was technically difficult, though it clearly was. It’s impressive because somebody decided it was worth doing. No metric rewarded it. The install base is small and shrinking. There’s no roadmap item that says “make 2013 graphics cards 30% faster.” A planning agent handed this backlog would deprioritize it instantly and be defensible in doing so.

The honest version of the agent pitch

When a vendor tells me their agent can modernize legacy code, this is the benchmark I hold it against. Not a Django upgrade. Not a React class component rewrite. A kernel driver transition where “it compiles” tells you nothing and the only real test is a pile of aging hardware under your desk.

I’m not arguing agents are useless here. Used well, they’re fine at the mechanical slices of a job like this — chasing call sites, summarizing commit history, drafting a first pass at a patch someone competent then rewrites. That’s real value and I’d use it. What I object to is the framing where that assistance gets sold as the whole job, because it flattens the distinction between typing code and understanding a system.

Kristóf’s year is a useful calibration tool. Next time a demo shows an agent “fixing a bug in a legacy codebase” in ninety seconds, ask what the bug was, how it was found, and who decided it mattered. Those three questions sort the genuinely helpful tools from the ones optimized for the video.

Credit where it’s due

Somewhere right now there’s a person running a GCN 1.0 card in a machine they can’t afford to replace, and that machine just got faster, got Vulkan, and stopped dropping display modes. They probably don’t know why. One engineer at Valve spent a year making it happen and presented the results at a conference most people have never heard of.

That’s the standard. Everything else is a product page.

🕒 Published: October 4, 2026