I was at AMD in the mid-late 2000s helping design CPU/APU/GPUs

Original Article ↗ Hacker News Discussion ↗

Accessing the Article / Twitter Issues

  • Several readers couldn’t view the original Twitter thread due to login walls and browser blocking.
  • Threadreader links were shared as a workaround, with criticism of X/Twitter’s UX and privacy-hostile behavior (e.g., blocking Firefox private mode while blaming the browser).

AMD vs Intel CPU History

  • Strong disagreement with the claim that AMD’s products were simply “superior” while Intel won on marketing.
  • Many recall AMD leading around Athlon 64 / Opteron (2003–2005), especially in servers and multicore workloads.
  • Consensus that Intel’s Core 2 era decisively flipped performance back to Intel, helped by better process and AMD’s delayed, buggy Barcelona and later Bulldozer designs.
  • Debate over “true” vs “glued” multicore; commenters note that today everyone glues chiplets and that AMD’s early multi-die products had real drawbacks.

AMD vs Nvidia: Hardware vs Software

  • Recurrent theme: AMD has strong engineering but chronically weak software, drivers, and ecosystem, especially for GPUs and AI.
  • AMD does well in consoles and has promising parts like MI300X, but reviews still flag ROCm as immature compared to CUDA.
  • Some argue AMD only really pivoted to AI software in late 2023 and needs time; others think this is AMD’s third failed attempt at a usable compute stack.

CUDA Ecosystem and Alternatives

  • CUDA praised as a major moat: stable, high-level, single-source programming model, rich tooling (profilers, debuggers), broad language support, and extensive libraries.
  • OpenCL and vendor alternatives are described as lower-level, buggier, slower to track hardware features, and poorly tooled.
  • Commenters stress that CUDA “just works” on cheap consumer cards, which built the ecosystem; AMD is seen as neglecting this entry-level path.
  • Attempts to run CUDA on AMD (e.g., ZLUDA) are mentioned, along with legal and strategic concerns.

Nvidia’s Lock-In, Business Practices, and Platform Strategy

  • Some are uncomfortable with Nvidia’s proprietary stacks, lock-in, and “monopoly-like” position; others argue Nvidia simply played within the rules and invested heavily in software.
  • Discussion of Nvidia’s broader platform: GPUs plus high-speed NICs, InfiniBand switches, NCCL, DGX systems, and cloud offerings, making them more than “just a GPU vendor.”
  • Skepticism from a minority that Nvidia is in a bubble and will be displaced once hyperscalers’ own accelerators are “good enough,” but many counter that software and ecosystem will keep Nvidia ahead for a while.

Linux Drivers and Desktop Experience

  • Mixed views on Nvidia’s Linux drivers: described as painful or neglectful for general desktop/Wayland users, but solid for paying workstation/VFX customers.
  • AMD is credited with eventually better open drivers, but also blamed for repeatedly shipping incomplete compute stacks that major apps (Blender, Octane) abandoned.
  • Recent community efforts (e.g., new open drivers) are mentioned as promising but not yet fully changing the landscape.

Valuations, Profits, and Investing vs Betting

  • Commenters note the huge gap: Nvidia’s market cap and quarterly net profit dwarf AMD’s entire revenue, and easily exceed Intel’s market value.
  • Debate over whether stock buying is “investing” or “betting”; some emphasize risk and speculation, others distinguish fundamentals-based investing from price gambling.
  • Anecdotes about small AMD positions bought near its historical lows that could have yielded life-changing gains if sized larger or held longer.

Reactions to the Author’s Framing

  • Some enjoy the insider history and culture anecdotes (e.g., early resistance to GPUs, CPU+GPU integration struggles, internal jokes about the ATI merger).
  • Others criticize the thread as self-promotional, historically fuzzy (e.g., confusing OpenGL/OpenCL), and biased toward AMD’s narrative of “superior but unlucky” products.
  • Claims that mid-2000s designs directly shaped “what we see today” are viewed as overstated given the Bulldozer detour and major later architectural shifts.

Miscellaneous Technical and Industry Notes

  • Discussion of hypothetical AMD–Nvidia mergers: some think cross-licensing could cut costs; more expect a merged entity would simply exploit monopoly power.
  • Mention of Nvidia’s past ARM-based Tegra chips and current Grace Hopper and upcoming Blackwell systems; questions about real-world experience with NVL-scale racks.
  • Comments on other accelerators (Google TPUs, Meta’s in-house chips) and Intel/TSMC/Samsung fabs versus Nvidia’s fabless, software-heavy model.
  • Minor tangents: historical jokes like “AMD+ATI=DAAMIT,” the “engg” abbreviation origin, and Steam GPU usage stats showing Nvidia’s overwhelming share in PC gaming.