Thirty percent. That’s the rough performance gain Valve’s Timur Kristóf squeezed out of AMD graphics cards that are roughly a decade old, running on an operating system most of their original owners never installed. No new silicon. No new model year. Just a year of patient kernel work moving GCN 1.0 and 1.1 era Radeons off the aging radeon driver and onto the modern amdgpu one.
I review toolkits for a living. Most of what crosses my desk is the opposite of this story: a shiny new framework, a fresh wrapper around someone else’s API, a repo with 400 stars and three open issues that nobody will answer. So when I see someone spend twelve months fixing display code defects on Southern Islands cards, I pay attention. Not because I own one. Because it’s a rare example of the thing I keep asking vendors for and almost never get.
What actually got fixed
The short version, per Kristóf’s XDC 2026 presentation and the coverage around it: Southern Islands and Sea Islands Radeons now default to the modern kernel driver. That transition brings Vulkan support, better display features, and better performance. Along the way he fixed display code defects, sorted out power management problems, and added soft reset support so a hung GPU doesn’t take the whole session with it.
The work hasn’t stopped either. There’s a further patch in flight targeting initialization of the UVD block on older hardware. That’s the unglamorous tail end of any migration project, the part where you’re chasing a specific block on a specific generation of chip that maybe a few thousand people still run.
Why a reviewer of AI tools cares
Because the failure pattern is identical across both spaces, and this is the counterexample.
In AI tooling, the lifecycle I see over and over goes like this. A project ships fast, gets traction, accumulates users with real workflows built on top of it. Then the maintainers’ attention moves to the next thing. The old code sits there with a README that still says “actively maintained” and a changelog that stopped eighteen months ago. Nothing is technically broken on day one. But the model API changes, a dependency bumps a major version, and suddenly your working pipeline is a weekend of debugging someone else’s abandoned abstractions.
What Kristóf did is the inverse. He took hardware that every reasonable product manager would have written off, declared it worth supporting properly, and then did the actual work instead of announcing an intention. The payoff isn’t a feature announcement. It’s that a user with a 2012 card gets Vulkan, a working display stack, and a GPU that recovers from a reset.
The maintenance tax nobody budgets for
Here’s what I’d want every toolkit author to take from this. Moving old hardware to a new driver is a migration, and migrations are where quality either shows up or doesn’t. The easy path is a compatibility shim: keep the legacy driver limping along, tell users it’s fine, ship nothing. The harder path is consolidating onto one well-maintained code path so every future improvement lands for everyone.
That second path is what happened here, and it’s why the performance number exists at all. The 30% didn’t come from a clever optimization. It came from old cards finally sitting on the code path that gets continuous attention. Consolidation is the optimization.
I see the same choice in AI toolkits constantly. Two provider adapters that drifted apart. A v1 API kept alive “for compatibility” while all the real work happens in v2. A CLI and an SDK that disagree about config format. Every one of those is a legacy driver nobody wants to retire, and every one of them taxes users who picked the wrong side.
What this doesn’t mean
I’ll be honest about the limits. This is one engineer at one company, and Valve has a specific commercial reason to care about Linux graphics working well across a wide range of machines. It isn’t a general sign that the industry has rediscovered long-term support. Nobody should read a 30% bump on GCN 1.0 and conclude that their abandoned Python package is about to get the same treatment.
And the number itself is a ballpark figure for an era of cards, not a benchmark you can quote for your specific GPU. Treat it as the direction of travel.
Still, as a reviewer I grade tools on whether they’ll still work in two years. Most of what I test fails that bar. This work passes it in a way that’s almost old-fashioned. Somebody looked at hardware everyone else had given up on, found the defects, and fixed them one at a time until the thing ran properly. That’s the standard. The fact that it feels remarkable says more about the rest of us than it does about Kristóf.
🕒 Published: