Asianometry Podcast Summary

Episode: VLIW: The “Impossible” Computer
Host: Jon Y
Date: April 5, 2026

Episode Overview

This episode explores the origins, theory, and legacy of Very Long Instruction Word (VLIW) computer architectures—a bold approach to making computers radically faster, not by building faster hardware, but by offloading unparalleled complexity onto the compiler. Through the story of Josh Fisher and the rise and fall of his startup Multiflow, host Jon Y unpacks how VLIW attempted to overturn accepted limits on computing speed and why, despite technical brilliance and initial success, it ultimately failed to break through commercially. The narrative covers technical concepts, business drama, and the influential afterlife of what was once considered an "impossible" computer.

Key Discussion Points & Insights

1. Setting the Stage: The “Impossible” Computer

• Metaphor Introduction ([00:03])
  Jon opens with a vivid Overcooked video game analogy, comparing CPUs to kitchens and instructions to recipe steps:

  "The world of the CPU is a little like Overcooked. A CPU is a kitchen that takes in raw data, inputs, ingredients and transforms it to get finished outputs.”

• Instruction Lifecycle
  The episode breaks down the life of a CPU: fetching, decoding, and executing instructions, and how human-written code is translated by compilers into machine instructions.
• Quest for Speed
  • Clock speed (Moore’s Law) is one route, but is there another?
  • Introduction of parallelism as a way to speed up execution, with a focus on instruction-level parallelism (ILP).

2. The Limits of Parallelism… and a Radical Breakthrough

• Code Dependencies & Branches ([08:20])
  Discussion of why exploiting parallelism is hard—dependencies between instructions and unpredictable branching paths.
  • Reference: 1970 paper asserts average parallelism limited to about two simultaneous operations.
• Josh Fisher’s Leap:
  • Working on the CDC 6600 emulator at NYU, Fisher observes a conceptual similarity between chip layout and code scheduling ([12:55]):

    “Both ingest one dimensional lists and put out two dimensional maps or grids.”

  • Develops trace scheduling—treats a program as a single block, predicts the most likely execution path, and schedules instructions for maximum parallelism.
• Compiler as Time Traveler ([16:00])
  The radical idea: shift the complexity from hardware to a very aggressive, predictive compiler.
• Problem of “Compensating Code”
  • If the trace prediction is wrong, additional code handles backtracking or correcting errors, risking code bloat.

3. From Paper to Product: ELI512 and Multiflow

• First VLIW Proposal ([19:23])
  • Fisher proposes the ELI512 computer and introduces the term “very long instruction word,” or VLIW.
  • Paper’s audacity: A simple RISC-based device able to execute 10–30 instructions per cycle, orchestrated by a trace-scheduling compiler.
• Industry Skepticism
  • Carnegie Mellon grad Bob Colwell’s reaction:

    “He wants to do what with a compiler? This guy is nuts.” ([21:17])

• The Entrepreneurial Leap
  • With encouragement from investors and the then-booming VC climate, Fisher leaves Yale to found Multiflow in 1984, naming the company for the “flow” of instructions.
  • Early struggles: Collapsed deals, frantic work schedules, and hardware–software clashes during TRACE 7200’s development ([28:00]).

4. Multiflow TRACE Computers and Compiler Magic

• Hardware Simplicity, Software Complexity
  • Emphasis on hardware simplicity: Multiple execution units, heavy interconnects, but almost all “intelligence” and parallelism handled by the compiler.
• TRACE Compiler Phases ([39:50])
  • Phase 1: Source code to intermediate representation (IL1)
  • Phase 2: Further reinterpretation and optimization (IL2), especially loop unrolling
  • Phase 3: Actual trace scheduling, prediction of likely execution paths, and insertion of compensation code
• Results and Real-World Use ([46:20])
  • Up to 10x speedups in best-case benchmarks, though varying by use case (“your mileage will vary”).
  • Compilations could take days due to complexity.
• Customer Reception
  • Launch event at the World Trade Center (1987); high-profile beta clients like NSA’s Supercomputer Research Center:
    • Grumman Data Systems:

      “The computer was running their software two hours after being uncrated.” ([47:55])

    • Sikorsky Aircraft’s CAD chief:

      “What the Multiflow people told us it could do seemed like black magic. But now…you don’t have to rewrite software significantly.” ([49:30])

  • PR push contributed to early momentum.

5. Market Dynamics, Competition, and Downfall

• Mini-Supercomputer Boom & Bust ([51:40])
  • Fierce competition: By 1987, up to 20 vendors for a $350 million market (e.g., Convex, Alliant, Cydrome).
  • Cray’s moves: Pushed out more powerful and cheaper models, adopted UNIX for compatibility, undermining mini-super value.
  • Rise of “killer micros” (RISC Unix workstations):

    “Ever faster cycle times let the killer micros make up for any architectural disadvantage. And at $100,000 these workstations price points cannot be beaten.” ([56:20])

  • Multiflow machines, reliant on expensive, discrete modules, couldn’t keep up with cost and size scaling from Moore’s Law.
• Final Years and Closure
  • Failed expansion, especially to international markets.
  • DEC considers acquiring Multiflow but backs out in 1990; company liquidates soon after.

6. Lasting Impact & Legacy

• Talent Diffusion

  • Despite commercial failure, Multiflow alumni go on to influence the industry:
    • Bob Colwell becomes chief architect of Intel’s Pentium Pro/II/III/4 CPUs.
    • Others evangelized VLIW concepts at HP, Intel, DEC.
  • Josh Fisher receives the Eckhart Mauchly Award for Computer Architecture.

• VLIW Lives On

  • The ideas behind VLIW architecture and trace scheduling influenced later architectures, like HP’s and Intel’s work.
  • Episode ends with a teaser:

    “But that is a story for another day.”

Notable Quotes & Memorable Moments

• On the radical rethinking of software vs. hardware

  “To Fisher, this made little sense, because to him, the hardware was the easy part. The compiler does all the hard work…Get the compiler right, and the rest falls into place.” ([23:55])

• On industry skepticism

  “You can get more people, a lot more people to come to your talk if you promise them bizarre sounding hardware instead of a compiler technique.” ([25:22])

• On the nature of computer progress

  “Big iron mainframes just fundamentally could not keep up with the greatest cost scaling items in human history.” ([57:10])

• On Multiflow’s legacy

  “Judging by what they were able to do and their influence on the computing world, they succeeded beyond their wildest dreams.” ([01:00:25])

Important Segment Timestamps

• 00:03 — Overcooked analogy and introduction to CPU & compilers
• 08:20 — Challenges of instruction-level parallelism
• 12:55 — Josh Fisher’s insight: chip layout vs. code scheduling
• 19:23 — Submission of ELI512/VLIW architecture paper
• 21:17 — Colwell’s skepticism
• 28:00 — Multiflow hardware/software development conflicts
• 39:50 — Breakdown of TRACE compiler phases
• 46:20 — Real-world performance, tales from early users
• 49:30 — Customer astonishment at Multiflow performance
• 51:40 — Market analysis & competition
• 57:10 — Rise of “killer micros,” market shifts
• 01:00:25 — Reflection on Multiflow’s real legacy

Conclusion

This episode weaves together computer architecture theory, business drama, and the unpredictable course of innovation. Through the story of VLIW and Multiflow, Jon Y illustrates how radical ideas often face skepticism, how technological brilliance struggles against business realities, and how bold vision can change the world—even when the company at the center fades away.

Further Reading Recommended in Podcast:

• Multiflow: A Startup Odyssey by Elizabeth Fisher

For more on the further evolution of VLIW, stay tuned for future Asianometry episodes.