Cray-1 vs Raspberry Pi

Original Article ↗ Hacker News Discussion ↗

Benchmarks comparing a 1978 Cray‑1 supercomputer to modern Raspberry Pis and phones highlight how today’s cheap, low‑power devices can outperform what was once the world’s fastest machine—often by tens to nearly a hundred times in floating‑point workloads. Commenters explore why that’s possible (vector vs SIMD architectures, software efficiency, Moore’s law, and changing workload types), reflect on how supercomputing tasks like weather forecasting, nuclear simulations, and CFD map onto current hardware, and debate when it still makes sense to use tiny boards like the Pi versus microcontrollers or small PCs for real projects.

Cray‑1 vs Raspberry Pi performance

  • Original claim: a 2012 Raspberry Pi Model 1 outperforms the Cray‑1 by ~4.5× on certain floating‑point benchmarks; later Pi 400 is tens of times faster (up to ~95× in some tests).
  • Commenters note this is remarkable both as progress and as a testament to how fast the Cray‑1 already was in the 1970s.
  • Several people ask how newer Pis (Pi 4, Pi 5) would compare; rough guess offered: Pi 5 ≈ 3× Pi 4 performance.

Vector architectures, SIMD, and RISC‑V

  • Some argue a fairer comparison to the Cray‑1 (a vector machine) would be RISC‑V with Vector 1.0, GPUs, or TPUs.
  • Thread dives into differences between classic vector machines and fixed‑width SIMD:
    • RISC‑V vectors have runtime‑varying length (portable across widths); x86/ARM SIMD is fixed‑width and often needs multi‑binary dispatch.
    • Concerns raised about cache/page crossings and specialized SIMD tricks (e.g., byte shuffles) in variable‑length designs.
    • Others respond with concrete RISC‑V instructions, examples (like vectorized strlen), and note ARM SVE has similar concepts.

Historical context and applications

  • Cray‑1 OS (COS) described as batch‑oriented but with time‑sharing support; debate over how “interactive” Crays really were and lack of full virtual memory vs. later systems.
  • Early use cases: nuclear weapons simulations, weather and climate modeling, CFD for ship hulls, fluid simulations, sparse linear algebra, and 3D rendering.
  • Anecdotes: supercomputing clubs offloading render jobs; national labs pairing Crays with front‑end minicomputers; stories about sitting on Cray couches and even pranks inside them.

Modern usage patterns and alternatives

  • Raspberry Pi praised as the cheapest general‑purpose Linux box, making many projects (e.g., sprinklers, home servers, robotics) easy to program and integrate.
  • Counterpoint: for simple control tasks (watering plants, sensors) microcontrollers like ESP32 or RP2040 are more appropriate and vastly cheaper, though harder to program than Linux.
  • Pi’s desktop role: Pi 4 is limited, especially for graphics and I/O; Pi 5 is reported as “snappier” yet still core‑count‑constrained.
  • Some recommend used “one‑liter” mini PCs (small x86 desktops) over Pi clusters for cost, performance, and storage reliability.

Moore’s law, performance scaling, and nostalgia

  • Multiple personal back‑of‑the‑envelope comparisons: modern desktops or phones would have topped early TOP500 lists or rivaled all 1980s compute combined.
  • Discussion of famous clusters built from commodity hardware and consoles (e.g., Mac G5 cluster “Big Mac”, PS3 clusters).
  • Mixed views on Moore’s law “ending”: some argue it keeps being extended (e.g., moves toward 2nm, 3D stacking), others wonder about physical and economic limits.
  • Reflection that many workloads which once took hours on supercomputers (rendering, physical simulations) can now be interactive or real‑time.

Form factor, design, and culture

  • Repeated jokes and nostalgia about the Cray‑1’s iconic circular form and built‑in seating vs. the Pi’s lack of “couch”.
  • Desire expressed for Cray‑style cases for Pis/PCs and better Pi enclosures (cooling, breadboards, even “seating”).
  • Side discussions on plastic use, biodegradability, and environmental trade‑offs.