a gross CFD computational requirement

[aerohead]

1) In 1989, the Department of Aerospace Engineering, at Princeton University published a simple metric for the computational requirement for solving the Navier-Stokes equations for all scales of turbulence.
2) The requirement = Reynolds number raised to the 9/4 power ( or an exponent = 2.25 ).
3) A swept wing example was assigned Rn = 1.0 X 10 ( to the eleventh power )
, or Rn-10,000,000.
4) 10,000,000 to the 2.25-power = 5-quadrillion, 623-trillion, 413-billion, 252-million calculations.
--------------------------------------------------------------------------------------
* At the time, the most powerful supercomputer, the Cray XMP, was rated at 100-floating-point operations per second ( 100 Mflops ).
** If the Cray were used, a single iteration of the N-S would require 1-year, 9-months, and 11-days to compute.
*** The authors reported that, in order to solve the Navier-Stokes equations for all manner of turbulence, would require a computer 1,000-times as powerful as the Cray XMP.
**** If the Moore's Law relationship of a doubling of a computer's power each 18-months held, then it would be the year 2002-1/2 before a supercomputer would exist, which was capable of calculating an iteration within 48-hours ( which is presently a welcome attribute ).
***** In 2011, Chalmers University of Technology published that no computer existed which could solve the N-S equations for an automobile, due to it's enlarged flow domain, compared to an aircraft.
****** In 2014, Mercedes-Benz reported that they WERE solving 99% accurate CFD solutions within 48-hours of run time.
******* Presently, I've not run across any reporting of commercially-available CFD packages which can solve automotive designs within 99% accuracy of a full-scale wind tunnel test other than, the Dassault Exa Powerflow CFD, and SIEMEN's STAR CCM+ CFD software.
------------------------------------------------------------------------------------
Presently, the University of North Texas System ( across the street ) has gone in on the Lonestar6 supercomputer at UT Austin, which will be shared by five entities.
It's a $ 8.4-million system, by DELL / AMD ( both of Austin, Texas ), with a capacity of 3-petaFlops, 7th-fastest in the USA.
With 41-petabytes of usable storage.
It would solve the N-S equations in under 2-seconds.

[freebeard]

I think it's an architectural problem. The evolution from 1989 has followed calssic Von Neumann architecture (wiki/Von_Neumann_architecture)

You are local to another entity, Tesla, who have the Dojo architecture

https://electrek.co › 2021 › 08 › 20 › tesla-dojo-supercomputer-worlds-new-most-powerful-ai-training-machine
Tesla unveils Dojo supercomputer: world's new most powerful ... - Electrek
20 Aug 2021At its AI Day, Tesla unveiled its Dojo supercomputer technology while flexing its growing in-house chip design talent. The automaker claims to have developed the fastest AI training machine in the...

https://nextbridge.com › teslas-dojo-supercomputer-is-unveiled
Tesla's Dojo Supercomputer is Unveiled and it's a BEAST!
17 Sept 2021Today's computer processing speed reaches 500 PetaFlops but, Tesla's Dojo is expected to reach up to 1 ExaFlop which would be one quintillion (1018) floating-point operations per second. This is how it could become the world's fastest AI-incorporated computer. For your reference: 1 Flop = 1 Floating Point 1 PetaFlop = 1,000 TeraFlops

technewsinsight.com/wp-content/uploads/2021/08/1629534447_80_Tesla-Introduces-New-AI-Supercomputer.jpg

The basic unit or cell is a tile. These are grouped in a way that solves the power management problem in a novel way.

Although Tesla use it to train their self-driving system, being an auto manufacturer they have an inherent interest is the aero problem.

[Talos Woten]

Originally Posted by freebeard

You are local to another entity, Tesla, who have the Dojo architecture

When I saw the Dojo architecture, I thought it was the most brilliant piece of engineering I had seen in awhile.

Aerohead, when I did classified research for the government years ago, we had a supercomputer in the basement. It was sadly only the 12th most powerful one in the country at the time. One of the highlights of my secret sojourn was when I was on a team working on a time critical project... and we got to boot everyone off every government sponsored supercomputer in the country for four hours.

[aerohead]

Originally Posted by freebeard

I think it's an architectural problem. The evolution from 1989 has followed calssic Von Neumann architecture (wiki/Von_Neumann_architecture)

You are local to another entity, Tesla, who have the Dojo architecture



technewsinsight.com/wp-content/uploads/2021/08/1629534447_80_Tesla-Introduces-New-AI-Supercomputer.jpg

The basic unit or cell is a tile. These are grouped in a way that solves the power management problem in a novel way.

Although Tesla use it to train their self-driving system, being an auto manufacturer they have an inherent interest is the aero problem.

Whatever the 'problem' actually was, the solution has been multi-core parallel processing.
The mathematics has, and remains the same.

[aerohead]

Originally Posted by Talos Woten

When I saw the Dojo architecture, I thought it was the most brilliant piece of engineering I had seen in awhile.

Aerohead, when I did classified research for the government years ago, we had a supercomputer in the basement. It was sadly only the 12th most powerful one in the country at the time. One of the highlights of my secret sojourn was when I was on a team working on a time critical project... and we got to boot everyone off every government sponsored supercomputer in the country for four hours.

That's a powerful adjective, and likely to alter 'priorities.'

[freebeard]

Whatever the 'problem' actually was...

I picked that illustration because it shows the 'problem' -- DC in, heat out. Here's a more indepth description: semianalysis.substack.com/p/tesla-dojo-unique-packaging-and-chip

The description around that illustration:

The power delivery is unique, custom, and extremely impressive as well. With so much bandwidth and over 10KW of power consumption on the package, Tesla innovated on power delivery and feeds it vertically. The custom voltage regulator modulator is reflowed directly onto the fan out wafer. Power, thermal, and mechanical are all interfacing directly with the tile.

It appears the total tile is 15KW of power even if the chips themselves are only 10KW total. Power delivery, IO, and wafer wires are drawing a ton of power as well. Power comes in from the bottom while the heat comes out the top. Chips are not the unit of scale for Tesla, the 25 chip tiles are. This tile far surpasses anything from Nvidia, Graphcore, Cerebras, Groq, Tenstorrent, SambaNova, or any other AI training geared start up in per unit performance and scale up capabilities.

[aerohead]

The mathematics and software for aerodynamics isn't a thermodynamics of integrated circuits issue.
AI isn't either.
It only has to do with the computational rigor.

[freebeard]

The mathematics and software for aerodynamics isn't a thermodynamics of integrated circuits issue.
AI isn't either.

I suggest it's an issue of latency and scaleability.

Originally Posted by TFA

Tesla’s architectural solution was a distributed compute architecture.... The biggest problem with these types of networks is scaling up bandwidth and retaining low latencies. In order to scale to larger networks, Tesla focused on these latter two especially. This influenced every part of their design from the chip fabrics to packaging.

It only has to do with the computational rigor.

It's turtles, all the way down.

[aerohead]

Originally Posted by freebeard

I suggest it's an issue of latency and scaleability.



It's turtles, all the way down.

'latency' has never made it into any formal discussion of CFD.
My mention of parallel processing, by default, deals specifically with 'scale.'
I don't know what to make of 'turtles.' That's not part of my lexicon.

[freebeard]

You are conflating hardware and software.

Turtles all the way down
"Turtles all the way down" is an expression of the problem of infinite regress. The saying alludes to the mythological idea of a World Turtle that supports a flat Earth on its back. It suggests that this turtle rests on the back of an even larger turtle, which itself is part of a column of increasingly larger turtles that continues indefinitely. URL="https://en.wikipedia.org/wiki/Turtles_all_the_way_down"]Wikipedia[/URL]

Sturgill Simpson - Turtles All The Way Down (Official Video)