Does flash, like solid state drives, have the same lifespan in terms of write? If so, it feels like this would most certainly not be useful for AI, as that use case would involve doing billions/trillions of writes in a very short span of time.
Manufacturers say to expect flash drives to last about 10 years based on average use. But life expectancy can be cut short by defects in the manufacturing process, the quality of the materials used, and how the drive connects to the device, leading to wide variations. Depending on the manufacturing quality, flash memory can withstand between 10,000 and a million [program/erase] cycles.
For AI processing, I don’t think it would make much difference if it lasted longer. I could be wrong, but afaik, running the actual transformer for AI is done in VRAM, and staging and preprocessing is done in RAM. Anything else wouldn’t really make sense speed and bandwidth wise.
Oh I agree, but the speeds in the article are much faster than any current volatile memory. So it could theoretically be used to vastly expand memory availability for accelerators/TPUs/etc for their onboard memory.
I guess if they can replicate these speeds in volatile memory and increase the buses to handle it, then they’d be really onto something here for numerous use cases.
Does flash, like solid state drives, have the same lifespan in terms of write? If so, it feels like this would most certainly not be useful for AI, as that use case would involve doing billions/trillions of writes in a very short span of time.
Edit: It looks like they do: https://www.enterprisestorageforum.com/hardware/life-expectancy-of-a-drive/
For AI processing, I don’t think it would make much difference if it lasted longer. I could be wrong, but afaik, running the actual transformer for AI is done in VRAM, and staging and preprocessing is done in RAM. Anything else wouldn’t really make sense speed and bandwidth wise.
Oh I agree, but the speeds in the article are much faster than any current volatile memory. So it could theoretically be used to vastly expand memory availability for accelerators/TPUs/etc for their onboard memory.
I guess if they can replicate these speeds in volatile memory and increase the buses to handle it, then they’d be really onto something here for numerous use cases.