What Puts The 'Super' In Supercomputer? [1:27:10]

YouTube ^ | September 25, 2026 | Scott Manley

[SunkenCiv]

[snip] In July I paid a visit to Lawrence Livermore National Labs which hosts the fastest supercomputer in the US, a couple of months earlier it was the fastest supercomputer in the world, but Moore's law means computer supremacy is fleeting. 

El Capitan is an exaflop scale system, meaning is can run 10^18 calculations per second (a billion, billion), and this is measured by benchmarks that require tight cooperation between it's thousands of corse. There are datacenters with many more processors, but those may not perform at the same level because of the interconnectedness required for this kind of task. 

So, I wanted to look at the whole installation from bottom to top to make it clear that a supercomputer is more than having huge numbers of processors in a datacenter. [/snip] 

What Puts The 'Super' In Supercomputer? | 1:27:10 
Scott Manley | 1.87M subscribers | 1,311 views | September 25, 2026

---

TOPICS: Computers/Internet
KEYWORDS: linux; lll; scottmanley; supercomputer

[SunkenCiv]

YouTube transcript reformatted at textformatter.ai is probably not going to follow due to length (over an hour, lot of talk time).

[SunkenCiv]

💾💿⌨️🖥️

[algore]

I have not seen one of these myself, but

https://www.amd.com/en/products/rackscale-solutions/helios.html

AMD Helios is a purpose-built, rack-scale AI infrastructure platform designed to compete directly with Nvidia’s high-end rack systems.

Key Specifications

GPU Count: 72 AMD Instinct™ MI455X GPUs per rack

CPU Count: 18 AMD EPYC™ “Venice” server CPUs (configured with 4 GPUs per tray)

Memory: 31 TB of total HBM4 memory (432 GB per GPU)Networking: AMD Pensando™ Vulcano AI NICs delivering up to 2.4 Tbps scale-out bandwidth per GPU

Interconnect: Open standards-based Ultra Accelerator Link (UALink) and Ultra Ethernet Performance & Adoption

Performance: Delivers up to 1.4 exaFLOPS of FP8 and 2.9 exaFLOPS of FP4 performance per rack.

Clients: Major tech companies including Meta, Microsoft, OpenAI, Oracle, and Anthropic have plans to deploy the platform.

[yuleeyahoo]

Blast processing!

[Harpotoo]

I did, going to tech school in the basement of the empire state building (Empire Tech School) in 1979 and learning on an IBM 360/30 and paving the way for all the little ai programmmers!-)

[Carry_Okie]

How long will it be before that kind of capability is on the desktop?

[Nateman]

...Moore's law means computer supremacy is fleeting...

Semi-conductors are close to the bottom . Only a handful of Silicon atoms are used for a transistor now .

[dfwgator]

I figure in about 10 years, it will be on an iPhone.

[Honorary Serb]

New!

[cymbeline]

I’ll sum it up. This computer is a mob of thousands of

It does useful things but it’s by brute force.

[Carry_Okie]

I figure in about 10 years, it will be on an iPhone.

I want to know about the desktop. The question was not rhetorical or figurative, but was asking people who know more than I do about this realm for planning my next major upgrade. I have work to do an AI that understood me well could greatly facilitate.

[Paal Gulli]

'What Puts The 'Super' In Supercomputer?"

That changes, almost from day to day.

Rear Admiral Grace Hopper (AKA "Grandma Cobol") used to carry a nonosecond of string in her pocket. To illustrate how the distance between computer components was limiting the speed of processing, she carried as piece of string to show how far electrical signal traveled in one nanosecond. Her string was a foot long but it should only have been 11.8" because that's how far light travels (in a vacuum) in one one-billionth of a second.

Fast forward to 1985 and Seymour Cray built his second supercomputer, the Cray-2. In recognition of the distance problem Hopper was championing, Cray crammed the components too close together to be air-cooled, then put the entire computer in a tank filled with Fluroinert, a dielectric liquid. The fluid was pumped through and around tightly packed circuit boards, then cooled through heat exchangers.

The Cray-2 was 45" tall, 53" across, and weighed 5500 lbs, drew as much as 200 kiloWatts, and its processing maxed out at 1.9 GFLOPs. Some state of the art smartphone chipsets are rated at 2-3 GFLOPs in overall workload (liquid cooling not included).

What's "super" today might not be so super tomorrow.

https://en.wikipedia.org/wiki/Cray-2

https://en.wikipedia.org/wiki/Fluorinert

[algore]

“How long will it be before that kind of capability is on the desktop?”

15-20 years.

I remember in 1999 I had a 4tb raid array that was connected to a compaq 5000 the raid array took 6 standard racks full of 9.1gb drives.

Today I have Dual Proc AMD server with 512 cores and 1024 threads, 3tb of memory and 24x 128gb ram with 12x24tb hard drives all in a nice 4u package

although I have mi350 card I have never seen anything like what is in the Helios system

[Carry_Okie]

15-20 years.

That sounds like more of an economic than physical barrier.

[spintreebob]

GIGO. Most AI data is dirty. Much of it is inaccurate.
Processing dirty data faster still gives unreliable results.
For advertising pizza, not much of a problem.

For life and death issues, police actions, military decisions dirty data is a problem.