Someone just started folding using an Nvidia H100 SXM5!!

[Demmers]

https://imgur.com/a/QxKP5zs

[muziqaz]

Image is broken, and I will use my magic ball and say that image was showing LAR systems broken entry into their database. The entry is broken, and does not show real FAH performance on that GPU.
H100 has 67TFLOPS of FP32 performance
5090 has 105 TFLOPS of FP32 performance
4090 has 82.6 TFLOPS of FP32 performance

So other than a novelty impression that someone is using that overpriced piece of hardware for folding, there is no significance in that
What I am saying, it is cool people test their magically expensive toys with FAH, but 5090 is still the fastest GPU for FAH

[Demmers]

Image is broken, and I will use my magic ball and say that image was showing LAR systems broken entry into their database. The entry is broken, and does not show real FAH performance on that GPU.
H100 has 67TFLOPS of FP32 performance
5090 has 105 TFLOPS of FP32 performance
4090 has 82.6 TFLOPS of FP32 performance

So other than a novelty impression that someone is using that overpriced piece of hardware for folding, there is no significance in that
What I am saying, it is cool people test their magically expensive toys with FAH, but 5090 is still the fastest GPU for FAH

Fixed the link.

Yes I see what you mean about the numbers. I'm curious how data sent to lar.systems gets broken?
And lets be honest, I think a lot of people would call the 5090 a magically expensive toy too

[muziqaz]

If people are folding on older fahclient, data taken by LAR can be very unreliable. v8 is more friendly with info it provides to LAR extension

[toTOW]

Yes, someone on Discord asked me to add such GPUs to stress test their brand new systems before moving into production.

[arisu]

A benchmark: https://openmm.org/benchmarks

The H100 folds a little slower and sometimes as well as an RTX 4080. On occasion it is even beaten by an RTX 3090.

[toTOW]

I'm seeing between 25 and 40 million PPD on the H100 I'm running for a couple of days ... the 120M on Lars is probably another example of the level of trust we can have on this site ...

[arisu]

I'm seeing between 25 and 40 million PPD on the H100 I'm running for a couple of days ... the 120M on Lars is probably another example of the level of trust we can have on this site ...

It thinks you have been running for 18 days and thinks you have returned 264 WUs, most of them on COVID projects from 2021.

I notice that Lars is often very broken. It often has incorrect project information (somehow it gets some project mixed up with others). And when I started folding on a GPU on Linux which was the first time Lars saw that combination, it suddenly said it saw more than 30 WUs and that each WU was completed in just a few minutes. The data on TDP for various GPUs also seems to be broken so the kWh-per-points stats are mostly bogus.

Plus a lot of links to GPUs are broken because the GPU has a slash in the name or something else that screws up the (unescaped!) URL.

[muziqaz]

Lar Dev frequents LTT forum fah section

[calxalot]

Here?
https://linustechtips.com/profile/668492-lar_systems/

[muziqaz]

They reply from time to time and they did do some changes to their system to fix few things upon my request.

[arisu]

I'm seeing between 25 and 40 million PPD on the H100 I'm running for a couple of days ... the 120M on Lars is probably another example of the level of trust we can have on this site ...

Would it be helpful to improve the site's reliability if I wrote a script that could run in the background and connect to fah-client's websocket, and upload more reliable data to Lars? I've been thinking of doing that for myself so that I can get better statistics without keeping a browser open 24/7, but I could post the script in the 3rd party apps section.

[calxalot]

They have a fork of web control that collects data.

[arisu]

They have a fork of web control that collects data.

Well yes, but like I said that requires keeping the browser open 24/7 and has some accuracy issues.

[calxalot]

Yup. I would say it would be good for the client itself to be more accurate.