An interview with AMD CEO Lisa Su about solving hard problems

Original Article ↗ Hacker News Discussion ↗

AMD’s recent interview-focused publicity around CEO Lisa Su has reignited debate over why the company lags Nvidia in AI and machine‑learning despite strong hardware. Commenters trace AMD’s position to years of underinvestment in software, a focus on HPC and FP64 workloads rather than deep learning, and a leadership culture seen as prioritizing chips over tooling and developers, while CUDA gained a decade‑long head start and broad ecosystem support. Some note that ROCm and AMD’s overall fortunes have improved markedly since the company’s near‑bankruptcy years, but many doubt AMD can close the AI gap without a much more aggressive, long‑term software strategy, especially as ARM and custom silicon threaten x86 margins.

AMD GPU Software Stack vs. NVIDIA CUDA

  • Many commenters argue AMD’s core unsolved “hard problem” is its GPU software stack (ROCm, drivers) for ML, which is viewed as fragile, poorly supported, and far behind CUDA.
  • Users report ROCm crashes, awkward installation (especially on Debian/Ubuntu), limited “official” GPU support, and long‑standing instability in past APIs (OpenCL, Vulkan compute, OpenGL).
  • Others note ROCm has improved noticeably since the Frontier supercomputer work; on supported hardware with the recommended kernel it can be usable, and some consumer cards work via environment overrides.
  • Consensus: NVIDIA’s stack is “least bad” and generally “just works,” from low‑end cards to data‑center parts, whereas AMD often requires hacks and has inconsistent support matrices.

HPC Focus vs. AI/ML Market

  • Several posts stress AMD historically targeted HPC and FP64 workloads (national labs, Frontier, El Capitan), not deep learning.
  • ROCm was initially optimized for DOE supercomputers and MI‑series accelerators, not consumer GPUs; this explains narrow “official” support.
  • Critics counter that AMD underinvested in GPGPU software for more than a decade, even while spending heavily on acquisitions and buybacks, and missed visible “dense compute” and AI trends.

Lisa Su, Software Reticence, and Interview Takeaways

  • Some readers see the interview as revealing a deep hardware bias: Su repeatedly frames herself and AMD as semiconductor‑focused and downplays software shortcomings.
  • Her denial that AMD ever had a software problem, and lack of explicit contrition or strong “software-first” messaging, is read as bearish by skeptics.
  • Others defend her track record, noting AMD was near bankruptcy pre‑Zen, and turning it into a profitable multi‑line semiconductor company is itself a major “hard problem” solved.

x86, ARM, and Long‑Term Prospects

  • Debate on whether AMD’s x86 strength is a “Titanic” in a world moving to ARM (Apple M‑series, Graviton, Snapdragon X, in‑house cloud CPUs).
  • Some argue ISA matters less than execution; AMD already has ARM experience and sees itself as a “compute company.”
  • Others fear commoditization: as ARM spreads and hyperscalers design their own chips, AMD’s CPU margins and moat may erode, making missed software bets more damaging.