# @stem_antics on Instagram

- **Type:** Video
- **Original URL:** https://www.instagram.com/p/DbV1lquAd_0
- **Gondola URL:** https://gondola.cc/posts/68532557-stem-antics-instagram
- **Thumbnail:** https://img.gondola.cc/tr:w-,h-,fo-auto/postThumbnails/6768c51710.jpg
- **Posted:** 2026-07-28T15:32:41.000+00:00
- **Account Owner:** Stem Antics (@stem_antics) — https://gondola.cc/stem_antics

## Caption

Four architectures. Four different bets about what computation actually looks like. The gap between them is not clock speed. It is assumption.

A CPU assumes the work is unpredictable. It spends its transistor budget on machinery for guessing: branch predictors, out-of-order execution engines, deep cache hierarchies, speculative loads. Only a small fraction of a modern core is arithmetic. The rest exists to keep that arithmetic fed when the instruction stream refuses to behave. Superb at serial, branch-heavy code. Mediocre at bulk mathematics.

A GPU assumes the opposite. The work is regular, parallel and enormous. Thousands of simple lanes execute in lockstep under one instruction stream. Latency is not avoided, it is hidden: when one warp stalls on memory, another is swapped in almost immediately. An H100 carries 16,896 shader cores and roughly 3 TB/s of HBM3 bandwidth to keep them fed.

A TPU narrows further. Its heart is a systolic array, a grid of multiply-accumulate cells through which operands are pumped rhythmically. A weight loaded once is reused across an entire row before it is discarded. Google deployed the first generation internally in 2015 for precisely that reason.

An NPU takes the same logic to the edge. Quantised to INT8 or INT4, budgeted in single-digit watts, it sits inside a phone running inference the CPU would flatten the battery attempting.

The spectrum, compressed:

- CPU: any workload, adequately
- GPU: parallel workloads, brilliantly
- TPU: matrix multiplication, ruthlessly
- NPU: quantised inference, cheaply

The unifying constraint is energy. Fetching a 32-bit operand from DRAM costs on the order of a hundred times more energy than multiplying it. Every accelerator ever built is, underneath, a machine for moving less data.

Generality is a tax. Specialisation is a wager.

#stemantics #computerengineering #semiconductors #gpu #hardware

## Stats

- **Views:** 6,524
- **Likes:** 169
- **Shares:** 0
- **Comments:** 3

## Tags

gpu, computerengineering, hardware, semiconductors, stemantics

---
Copyright (c) Gondola