# Moore's perspective

So I've stumbled upon an [article mentioning](https://en.wikipedia.org/wiki/PlayStation_3_cluster) that the most powerful supercomputer in 2010 was [Tianhe-1 with a peak performance of 2.5 petaFLOPS](https://en.wikipedia.org/wiki/Tianhe-1).

![China Restarts One of World's Fastest Supercomputers | Data Center  Knowledge | News and analysis for the data center industry](https://www.datacenterknowledge.com/sites/datacenterknowledge.com/files/styles/article_featured_retina/public/wp-content/uploads/2015/08/Tianhe-1-Tianjin-China.jpg?itok=aZw03NrH align="center")

Fast forward, it's 2023. We've got [Graphcore IPUs available for free](https://console.paperspace.com/signup?R=SP8I2F) (ref link so I'll use it for paid plans...) at [Paperspace](https://docs.paperspace.com/gradient/machines/#free-machines-tier-list).

And they've got... **1 petaFLOPS with FP16** ([half-precision floating point format](https://en.wikipedia.org/wiki/Half-precision_floating-point_format)):

![](https://cdn.hashnode.com/res/hashnode/image/upload/v1673302332791/cea8e9ef-b8d9-4d7f-a81c-8f0c5a402de2.png align="center")

Okay, okay, Tianhe-1 has 2.56 on FP32, with the free IPU having just 0.25... But just think about it: **13 years** had passed, and we've got <mark>10% of the world's most powerful computing unit power</mark>, costing $20M/year in operating expenses and serviced by 200 people worth of staff, <mark>for free</mark>. And it weighs [about 16.5 kgs](https://www.graphcore.ai/products/bow-pod16#product-spec) instead of a datacenter-level size.

![Underrated Characters Imagine — Game of Thrones Preference "How they react  to you...](https://64.media.tumblr.com/c97106abca3067dec6bbe366a55778fa/13e3b36567959d40-a7/s540x810/d3472980ad44170f67f02ea5277662a6a6eb89c1.gifv align="center")