4 ms·

A stockpile of rapidly depreciating assets bought at eye-watering margins is an unusual brag for any company, no?

nomel 3y ago
From a naive perspective, it seems that true research/advances in AI (methods of training, etc) aren't necessarily related to model size. It seems that the goal of "building a big model that everyone else converges to because the training data is the same" doesn't have all that much value, especially since you could wait a couple years, do it all for a fraction of the price, and catch up immediately. Meta doesn't have an AI product yet, so it's not like they would be loosing money. I suspect this is more about talent attraction/retention.
dist-epoch 3y ago
CPUs are rapidly depreciating, hard drives are rapidly depreciating, SSDs are doubly rapidly depreciating, with this logic no hardware buy would make sense.
- formerly_proven 3y ago
  Computer hardware has always been rapidly depreciating. You'd always get much more (performance/capability) for the same money just a few months to 1-2 years down the road. GPUs have been a complete outlier in this area for around ~8 years, and even they depreciate relatively quickly still.
crotchfire 3y ago
I am going to be buying these by the truckload on ebay three years from now. The current generation of GPUs is definitely going to have a long usable lifetime. Manufacturers finally have HBM totally figured out and yielding/aging well. Today's GPUs are the analog of the 28nm logic node -- it was the sweet spot for an absurdly long time. There will be something faster/lower-power 2-3 years from now that will cause the BigCos to cycle their fleet, but it won't be anywhere near the VRAM/MemBandwidth boost we have today compared to 2-3 years ago. Their data centers are power-constrained, so they'll upgrade regardless, and the rest of us will feast on the windfall. Don't ninja-bid on my auctions.

From a naive perspective, it seems that true research/advances in AI (methods of training, etc) aren't necessarily related to model size. It seems that the goal of "building a big model that everyone else converges to because the training data is the same" doesn't have all that much value, especially since you could wait a couple years, do it all for a fraction of the price, and catch up immediately. Meta doesn't have an AI product yet, so it's not like they would be loosing money. I suspect this is more about talent attraction/retention.

dist-epoch 3y ago

CPUs are rapidly depreciating, hard drives are rapidly depreciating, SSDs are doubly rapidly depreciating, with this logic no hardware buy would make sense.

formerly_proven 3y ago
Computer hardware has always been rapidly depreciating. You'd always get much more (performance/capability) for the same money just a few months to 1-2 years down the road. GPUs have been a complete outlier in this area for around ~8 years, and even they depreciate relatively quickly still.

crotchfire 3y ago

I am going to be buying these by the truckload on ebay three years from now. The current generation of GPUs is definitely going to have a long usable lifetime. Manufacturers finally have HBM totally figured out and yielding/aging well. Today's GPUs are the analog of the 28nm logic node -- it was the sweet spot for an absurdly long time. There will be something faster/lower-power 2-3 years from now that will cause the BigCos to cycle their fleet, but it won't be anywhere near the VRAM/MemBandwidth boost we have today compared to 2-3 years ago. Their data centers are power-constrained, so they'll upgrade regardless, and the rest of us will feast on the windfall. Don't ninja-bid on my auctions.