Over 2 exaflops!

[MeeLee]

Well, if there isn't any usable stats, I suppose that there isnt' any world record, so maybe the wikipedia article should be corrected :

Folding@home is one of the world's fastest computing systems. With heightened interest in the project as a result of the COVID-19 pandemic,[7] the system achieved a speed of approximately 1.22 exaflops by late March 2020 and reached 2.43 exaflops by April 12, 2020,[8] making it the world's first exaflop computing system.

[8] Pande lab. "Client Statistics by OS". Archive.is. Archived from the original on April 12, 2020. Retrieved April 12, 2020.

I used these high scores to try to convince friends to join F@H.
Are you sure that F@H is the world's first exaflop computing system ??
I feel a little stupid now.

It might be close.
The fastest super computer in 2018 was the IBM Summit, which at 200Petaflops held the crown for 2 years.
Only in 2020, when Japan turned on their Fujitsu Fugaku super computer, did they reach 1.4 Exaflops.
They did come into service in June, a couple of months after FAH hit their peak.

GPU supercomputing is a relatively new field, with the fastest GPU super computer being from Nvidia in the top 500 list of fastest super computers.
And while FAH isn't a super computer, but distributed computing, it did topple the Fujitsu Fugaku in peak performance by just over 0.6 Exaflops.
Unlike FAH though, government super computers have this power available at nearly all the time (unless nodes go offline, which usually are few).

I would estimate FAH to have had access to 2 exaflops at some time, but things like server overloads, and end user PC systems coming on and offline, made that this power was available only for a short amount of time.
Not to mention, that a lot of super computers are tested with FAH batches.
I've seen a couple of companies or PCs reach 1 Billion points in a matter of a few days; probably helping to reach that 2 exaflops result.

Aside from CPU and GPU supercomputers helping out for a moment,
There also were a lot of bitcoin folding farms that 'entered the fold'.
A lot of server farms switched over to FAH, when the bitcoin currency went down, some of them reaching well over 200M PPD (that's 100x 2060 GPUS)

And it's by far not the end of things.
A lot of users will be upgrading their GPUs in the coming months, resulting in a MUCH higher performance (in part thanks to the RTX3000 series GPUs, and progress).

[bruce]

When FAH decided t develop official containers, it seemed like a good idea. ...but even good ideas sometimes produce unexpected results. The problem of replacing an old installation with a new one without notifying the server that the old one is permenantly out-of-service has been with us to a lesser degree but the magnitude prudently grew and was identified. Solutions are being explored, but setting the timeout to a smaller value is at least a temporary solution.

This may be helped once the benchmarking and species realignment is completed. If we can predict when you'll return your next result, we can deduct that projected completion when it doesn't arrive on time. That's not a perfect solution, either.

With a unified super computer, the hypervisor knows which nodes are/are-not producing. With a distributed super computer all we really know is what work has/has-not been assigned.

[oreggin]

If every WU has a measureable FLOP value then the FLOPS of F@H distributed computer is clearly measured by returned successfully folded WUs per time. I think this mentioned by bruce in other thread.
The phantom clients is a little bit harder issue. In boinc system a client periodically (~in every hour) contacts to the project server and exchange infos and it has the same client authentication issue even periodic exchange. Without strict client authentication there is no way to measure clients but strict auth is maybe harm science. BTW I think the FLOPS value of F@H is more important then the number of virtualized clients.

[bruce]

Various projects are made up of different proteins and simulate a variety of time spans. Most projects (except Moonshot) consist of a number of WUs which are then assembled into a number of trajectories through the simulated space. WUs in one project rarely contain the same number of FLOPs except within that project ... and even that isn't necessarily constant.

The project is assigned a value for baseline points which is a measure of scientific value. Those points, as well as your daily production are sort of a measure of FLOP count, so watching the total FAH points each day is a rough measure of FAH's productivity ... but not in meaningful units. Capturing the FLOP count from every wu and reporting a total will put it in a more meaningful set of units.