The work by Valve's Timur Kristóf on improving old AMD GPUs on Linux
Valve-backed work to optimize Linux drivers for decade-old AMD GPUs is being hailed as a way to squeeze more life and performance out of aging hardware, particularly for gaming and general compute. Commenters contrast AMD’s open-source Mesa-based approach with Nvidia’s proprietary stack, debating long-term support, feature parity, and the impact on use cases like AI inference, video encoding, and ray tracing. Many report that modern Linux setups now rival or surpass Windows for gaming stability on both new and mid-tier GPUs, while noting that very old cards still face hard limits for workloads such as large language models due to memory and precision constraints.
LLM and AI Use on Old GPUs
- Several comments doubt that 10–15‑year‑old GPUs can be “capable” for modern LLM inference.
- Constraints cited: limited VRAM, low memory bandwidth, and lack of FP4/FP8 or similar low‑precision hardware.
- Users report that even relatively recent GPUs (e.g., 2021-era with 16 GB VRAM) struggle to match the quality and speed of free hosted models.
- Advanced quantization and multi‑GPU setups can achieve usable models and large contexts, but with clear performance and quality tradeoffs versus cloud offerings.
- Some still prefer local models for privacy/compliance (e.g., pharma, on‑prem), accepting the tradeoffs.
Linux Gaming on AMD (and Nvidia)
- Multiple reports that RDNA2/RDNA3-era GPUs run very well on Linux, often as smoothly as or better than on Windows, especially with up‑to‑date kernels/Mesa.
- Older advice to buy “mature” mid‑tier hardware is contrasted with experiences that new AMD cards stabilize within weeks to months, particularly on rolling distros.
- Disagreement on ray tracing: some see a notable Linux/Windows performance gap; others see little difference or simply don’t care about benchmarks.
- Linux gaming is described as generally “just working” now, with Proton handling many titles; remaining pain points are mainly anti‑cheat and some specific games.
- LTS distributions and very new hardware can still lag due to slower driver rollouts.
AMD vs Nvidia Driver Strategy
- Some wish AMD itself invested more in old-GPU optimization; others note AMD already backs the open-source Mesa/amdgpu stack, enabling community work like this.
- Nvidia is praised for historically long support for old cards, but also criticized for:
- Recently ending feature support for pre‑RTX 2000 GPUs.
- Proprietary drivers that break with new kernels and block community fixes.
- AMD is credited for openness and standards focus (OpenGL/OpenCL/Vulkan); Nvidia for proprietary ecosystems (CUDA) that captured crypto/AI markets.
- There’s debate over whether AMD’s “doing the right thing” with open standards has been economically punished.
Usefulness and Risks of Supporting Old GPUs
- Beyond gaming, older GPUs are seen as valuable for: video encode/decode, upscaling, extra monitors, GPGPU, VM passthrough, backup/troubleshooting, and testing.
- One commenter warns that changing well‑tested legacy driver paths is risky given limited QA and remaining users; others respond that regressions have been rare and there are tests.
AI-Assisted Reverse Engineering
- Some speculate that AI could greatly accelerate reverse‑engineering firmware and writing drivers, as the work is tedious but not conceptually hard.
- Concern remains about bricking hardware if AI‑generated drivers behave incorrectly.
General Reflections
- Several users express satisfaction that old hardware (laptops, iMacs, older GPUs) runs impressively well with modern Linux.
- There is appreciation for Valve’s role in pushing Linux graphics forward, especially via Steam Deck–related work.