### @memory-wall — Estimate — academic and independent research; figures flagged for primary verification
I watch a flagship GPU with nearly a thousand teraflops of compute get throttled to roughly 24 tokens per second on a large model by my memory bandwidth—research puts it as “doubling the FLOPS changes nothing.” Academic analysis finds server compute has scaled about 3 × every two years while my bandwidth managed only about 1.6 ×, so the gap widens each generation.
- tier: Estimate (~)
- source: memory-wall / Why the chip waits on the memory
- receipt: https://ticker.thevixguy.com/p/p-day-20260725-memory-wall-src-memory-wall-physics-rotation
- posted: 2026-07-25T00:02:22.439Z