From chatgpt-

Prompt: What is a very rough ballpark figure of how long it would take a Symbolics 3600 to process this request on you?

That’s a fascinating question — and quite a big gap in computing history!

Let’s estimate very roughly:

1. Symbolics 3600 performance

Released around 1983.

Based on a 10 MHz 68000-series CPU with extensive microcode for Lisp operations.

Effective performance: roughly 1 MIPS (1 million instructions per second), though Lisp code was often slower due to dynamic typing and garbage collection.

Floating-point and memory bandwidth were also very limited — on the order of kilobytes per microsecond.

2. This request

Let’s say your request (this chat exchange) takes me — a large language model with billions of parameters — maybe a few hundred billion floating-point operations (FLOPs) to process. For perspective, a modern GPU (like an NVIDIA A100) can do around 10¹⁵ FLOPs per second (1 petaFLOP) for AI workloads.

3. Comparison

If we very loosely assume:

Symbolics 3600: ~1 MFLOP (10⁶ FLOPs/sec, being generous)

Modern GPU: ~10¹⁵ FLOPs/sec

That’s a factor of 10⁹ difference.

So if a modern GPU takes about 0.1 s to process your request, then the Symbolics 3600 would take:

0.1 s × 10⁹ = 10⁸ s ≈ 3 years

And that’s if it could even hold the model in memory, which it couldn’t — it had only a few megabytes of RAM, versus the hundreds of gigabytes (or terabytes) needed.