Documentation for the AMD 7900XTX

Original Article ↗ Hacker News Discussion ↗

Scope of the Documentation & Reverse Engineering Effort

  • The repo documents low‑level behavior of the 7900XTX, focusing on bypassing layers of user‑space drivers, kernel drivers, and on‑device firmware to talk more directly to RDNA compute engines.
  • AMD’s official ISA docs (GPUOpen) are described as only “the tip of the iceberg”; substantial functionality is hidden behind multiple firmware components using several undocumented ISAs.
  • This work is seen as a way to enable independent tooling (e.g., CUDA‑like stacks, HIP/ZLUDA, or tinygrad) and to debug crashes that currently hard‑lock systems.

Open Source vs Closed Firmware

  • Several comments argue AMD’s “open source” claim is misleading: drivers and APIs are open, but critical firmware/microcode blobs remain closed and buggy.
  • Because firmware isn’t open, outsiders can’t fix stability issues; instead, AMD allegedly “mitigates per application,” which works for a few AAA games but fails for general compute and ML.
  • Others argue firmware is only part of the problem and that the rest of the software stack and development process are at least as important.

AMD vs NVIDIA: Ecosystem and Reliability

  • NVIDIA is credited with a 16‑year head start with CUDA and a robust ecosystem; this long investment is seen as the main reason for its dominance and high margins.
  • Multiple users report that NVIDIA consumer GPUs are reliable for CUDA and ML, while AMD GPUs (including enterprise and officially “supported” compute models) frequently crash under OpenCL/ROCm and ML workloads.
  • Some note AMD CPUs are now highly competitive or superior to Intel, but AMD’s GPU software quality has lagged for decades.

Competition, Business Strategy, and Antitrust Concerns

  • Many express strong desire for real competition in AI hardware to reduce NVIDIA dependence and prices, mentioning AMD, Intel, Google TPUs, and future ASICs.
  • There is frustration that AMD underinvests in software despite obvious demand and potential profit, with some speculating about structural/organizational issues or legal risk around copying CUDA.
  • A few raise concerns about market functioning and antitrust (e.g., extreme NVIDIA profit margins, limited ROCm support, and narrow AMD focus on supercomputing/MI300), while others counter that NVIDIA earned its position through long‑term investment.