News Scientists used a retired supercomputer to prep for NASA’s Roman mission — Argonne Theta supercomputer created nearly four million simulated images

[Admin]

NASA is preparing to launch a new space telescope and is calibrating expectations with a supercomputer and the OpenUniverse project.

Scientists used a retired supercomputer to prep for NASA’s Roman mission — Argonne Theta supercomputer created nearly four million simulated images : Read more

 

[bit_user]

Considering the operating costs of running it for those 9 days, it probably would've been cheaper to just rent the equivalent amount of compute power on AWS.

 

[derekullo]

Considering the operating costs of running it for those 9 days, it probably would've been cheaper to just rent the equivalent amount of compute power on AWS.

Was literally doing the math of how many laptops I'd need when I read that lol.
Hard to tell if the laptop is an i3 with 4 threads or an i7 with 20+ threads.
300 years = 109,500 days
109,500 days / 9 days = 12,166 laptops
At an easy math cost of $1000 a laptop that's $12.1M in laptops!

AWS probably easier since you don't need to figure out how to cluster 12166 laptops and you don't have to find a way to get rid of 12166 laptops when the computation is done!

 

[bit_user]

Was literally doing the math of how many laptops I'd need when I read that lol.

For an old supercomputer to be that much faster, I figured they must be using an old laptop as a reference point.

109,500 days / 9 days = 12,166 laptops

Good thinking! The next step would be to go from laptops to cores. So, if we assume the figure dated back to the supercomputer's heyday, then they would've probably been talking about dual-core laptops, which means 25k cores. If we assume each AWS core is now twice as fast, then maybe we're back to 12k cores. If they use 192-vCPU instances (c6a.48xlarge?), they'd need 64 days worth of them. I have no idea about AWS pricing... what would that work out to?

So, I naturally wanted to check our estimates and pulled up its specifications. It has mostly CPU instances and a handful of GPU ones.

https://www.alcf.anl.gov/alcf-resources/theta
​

They claim the CPU instances have a total of 281k cores, so I was only off by an order of magnitude! However, if I'm reading the entry correctly, those cores are assembled from 64-core, 1.3-GHz Intel Xeon Phi 7230. Those are quad-thread, dual-issue Atom-class CPUs. Aside from each having 2x AVX-512 pipelines, those cores are basically trash. That would put the right answer probably at some small multiple off of what I said.