The model was run (and I think trained?) on very modest hardware:
The computer used for this paper contains an NVIDIA Quadro RTX 6000 with 22 GB of VRAM, 200 GB of RAM, and a 32-core Xeon CPU, courtesy of Caltech.
That’s a double VRAM Nvidia RTX 2080 TI + a Skylake Intel CPU, an aging circa-2018 setup. With room for a batch size of 4096, nonetheless! Though they did run into some preprocessing bottleneck in CPU/RAM.
The primary concern is the clustering step. Given the sheer magnitude of data present in the catalog, without question the task will need to be spatially divided in some way, and parallelized over potentially several machines
I mean, “modest” may be too strong a word, but a 2080 TI-ish workstation is not particularly exorbitant in the research space. Especially considering the insane dataset size (years of noisy, raw space telescope data) they’re processing here.
Also that’s not always true. Some “AI” models, especially oldschool ones, function fine on old CPUs. There are also efforts (like bitnet) to get larger ones fast cheaply.
That’s even overkill. A 3090 is pretty standard in the sanely priced ML research space. It’s the same architecture as the A100, so very widely supported.
5090 is actually a mixed bag because it’s too new, and support for it is hit and miss. And also because it’s ridiculously priced for a 32G card.
And most CPUs with tons of RAM are fine, depending on the workload, but the constraint is usually “does my dataset fit in RAM” more than core speed (since just waiting 2X or 4X longer is not that big a deal).
I’ve managed to run AI on hardware even older than that. The issue is it’s just painfully slow. I have no idea if it has any impact on the actual results though. I have a very high spec AI machine on order, so it’ll be interesting to run the same tests again and see if they’re any better, or if they’re simply quicker.
The model was run (and I think trained?) on very modest hardware:
That’s a double VRAM Nvidia RTX 2080 TI + a Skylake Intel CPU, an aging circa-2018 setup. With room for a batch size of 4096, nonetheless! Though they did run into some preprocessing bottleneck in CPU/RAM.
That’s not modest. AI hardware requirements are just crazy.
For an individual yes. But for an institution? No.
I mean, “modest” may be too strong a word, but a 2080 TI-ish workstation is not particularly exorbitant in the research space. Especially considering the insane dataset size (years of noisy, raw space telescope data) they’re processing here.
Also that’s not always true. Some “AI” models, especially oldschool ones, function fine on old CPUs. There are also efforts (like bitnet) to get larger ones fast cheaply.
So a 5090, 5950x3d & 192gb of RAM would run it on “consumer” hardware?
That’s even overkill. A 3090 is pretty standard in the sanely priced ML research space. It’s the same architecture as the A100, so very widely supported.
5090 is actually a mixed bag because it’s too new, and support for it is hit and miss. And also because it’s ridiculously priced for a 32G card.
And most CPUs with tons of RAM are fine, depending on the workload, but the constraint is usually “does my dataset fit in RAM” more than core speed (since just waiting 2X or 4X longer is not that big a deal).
I’ve managed to run AI on hardware even older than that. The issue is it’s just painfully slow. I have no idea if it has any impact on the actual results though. I have a very high spec AI machine on order, so it’ll be interesting to run the same tests again and see if they’re any better, or if they’re simply quicker.
Is it a PyTorch experiment? Other than maybe different default data types on CPU, the results should be the same.