Cray versus Raspberry Pi
Comparing a 1970s Cray‑1 supercomputer to today’s Raspberry Pi highlights how orders‑of‑magnitude gains in compute, memory, and cost have pushed “supercomputer‑class” power into cheap consumer boards and even peripherals. Commenters explore what early scientists might have done with such devices, how modern GPUs now exceed past “one Cray per pixel” ray‑tracing estimates, and why many historically “big iron” workloads like weather, CFD, and nuclear simulations have deepened rather than disappeared. Alongside this, they note that software bloat, shifting AI paradigms, and plateauing hardware scaling complicate simple Moore’s‑law narratives about inevitable progress toward general superintelligence.
Design, Nostalgia, and Retro Builds
- Cray-1’s iconic cylindrical look is compared to Apple’s “trash can” Mac; people speculate about subconscious design influence.
- Several commenters fantasize about Pi or Pico clusters housed in Cray-style cases and reference existing Cray-shaped DIY builds and 3D-printed Y-MP cases.
- There’s broader nostalgia for 70s–80s sci-fi props (Knight Rider, Blake’s 7, Space: 1999) that could now be almost trivially replicated with modern SBCs.
Sci-Fi Expectations vs Today’s Reality
- Commenters note that a 1970s person shown an RPi5 or modern phone would find it “impossible,” echoing how sci-fi imagined talking computers and cars.
- Early text-to-speech (C64, Atari, car voice warnings) is contrasted with current LLM-based conversational systems; consensus is that KITT-level dialogue is only now becoming plausible.
- There’s sharp disagreement over whether self-driving is “already mundane”: some argue tech is effectively ready but blocked by law; others say current systems are still “sparkling lane assist” and nowhere near safe, unattended autonomy.
Could Old Supercomputers Have Run LLMs?
- One line of discussion claims Cray-era machines could have run small neural models useful for autocomplete, linting, or summarization; the blockage was concepts and datasets, not hardware.
- Others push back, arguing that even tiny models require far more parameters, data, and training compute than those systems could feasibly support.
- The debate dives into parameter counts, FLOP estimates, historical systems like LeNet-5, and whether a 300K-parameter toy model proves anything beyond “it technically runs.”
What Happened to Cray-Class Workloads?
- Original Cray workloads (weather forecasting, CFD, nuclear simulations, fusion coil design, CGI like “2010”) are still done, but at higher resolution, in 3D, or inside optimization loops.
- Several note that many scientific and engineering problems remain compute-bound; better hardware mostly buys finer meshes, more physics, and higher accuracy, not “instant solutions.”
Hardware Progress, Moore’s Law, and Cost
- Multiple comparisons: Cray-1 vs Pi, Pico 2 / RP2350, Pi Zero 2, and consumer GPUs (e.g., ray tracing “1 Cray per pixel” vs a single RTX 4080).
- Discussion highlights that Moore’s law is about transistor counts, not FLOPS, and that real systems (including TOP500 supercomputers) don’t track the idealized curve.
- Some stress that the miracle isn’t just performance but economics and infrastructure: supercomputer-class capability in sub-$20 boards or essentially free microcontrollers.
Software Bloat and Use of Compute
- Several lament that vast gains in hardware are “spent” on bloated web stacks, JavaScript-heavy sites, Electron-style apps, and tracking/ads instead of pure computation.
- Others counter with examples where massive compute has quietly enabled whole fields (modern CAE, improved forecasts, stealth design, etc.).
Time-Travel Thought Experiments
- People speculate what 70s–80s scientists might have done if each had an RPi-class machine instead of queuing for shared Crays; ideas focus on higher-dimensional simulations and more ambitious experiments.