The trouble with the railways comparison is that after investing tons of cash the railways were built. With AI the GPUs have no value after 6 years (if that). So the investment must continue forever. It’s madness.
It’s not that they don’t technically work. It’s just they’re no longer efficient compared to newer versions that can do more with less power. So to remain competitive you need to upgrade otherwise your cost to execute a model is too high.
Hyperscalers used to write GPU’s down to zero value after three years, over the last couple of years they’ve all increased this to six.
But transistors break after what? 100’000 cycles?
GPUs can get “used up”. And if your computing center has twice as much running costs due to old, less efficient hardware, it isn’t competitive.
Edit: looks like transistors can partially recover with sleep cycles.
Editedit: that with the cycles was in flash storage. Looks like it’s higher in computing?
It’s not actually the transistors that break down in flash memory. Flash memory works by storing charges in what is effectively a grid of capacitors, and in order for the data to remain stored, the insulating oxide layers in the cells need to be preserved. Every time a cell gets written, a charge is forced through the insulation with high voltage, and this degrades the insulation. A single flash cell might only have a few 1000 writes before this insulation goes bad and it no longer holds data. Modern SSDs have wear levelling techniques to make the drive as a whole last longer.
Transistors on the other hand don’t have any inherent degradation that I’m aware of other than external factors like corrosion. The first thing that’s likely to die on a GPU is the electrolytic capacitors in the power filtering electronics, which have fluid in them that dries out over many years.
I’m not an expert but I was under the assumption that electronic components (including silicon chips and their internals) will age and give out on the decade timescale
The trouble with the railways comparison is that after investing tons of cash the railways were built. With AI the GPUs have no value after 6 years (if that). So the investment must continue forever. It’s madness.
The other trouble with the railways comparison is that trains actually work and can generate a profit for their owners.
What? GPUs don’t age. They might get old technologically wise, but they don’t just… die. The silicone chip itself doesn’t care about age.
It’s not that they don’t technically work. It’s just they’re no longer efficient compared to newer versions that can do more with less power. So to remain competitive you need to upgrade otherwise your cost to execute a model is too high.
Hyperscalers used to write GPU’s down to zero value after three years, over the last couple of years they’ve all increased this to six.
But transistors break after what? 100’000 cycles?GPUs can get “used up”. And if your computing center has twice as much running costs due to old, less efficient hardware, it isn’t competitive.Edit: looks like transistors can partially recover with sleep cycles.
Editedit: that with the cycles was in flash storage. Looks like it’s higher in computing?
It’s not actually the transistors that break down in flash memory. Flash memory works by storing charges in what is effectively a grid of capacitors, and in order for the data to remain stored, the insulating oxide layers in the cells need to be preserved. Every time a cell gets written, a charge is forced through the insulation with high voltage, and this degrades the insulation. A single flash cell might only have a few 1000 writes before this insulation goes bad and it no longer holds data. Modern SSDs have wear levelling techniques to make the drive as a whole last longer.
Transistors on the other hand don’t have any inherent degradation that I’m aware of other than external factors like corrosion. The first thing that’s likely to die on a GPU is the electrolytic capacitors in the power filtering electronics, which have fluid in them that dries out over many years.
I doubt the transistor on a GPU wafer break after 100k cycles, as they run at gigahertz frequencies, some cycle billions of times a second.
SURE, BUDDY.
I’m not an expert but I was under the assumption that electronic components (including silicon chips and their internals) will age and give out on the decade timescale
Check out “References” part here