Comment by trenchpilgrim
2 days ago
> hardware that has commercially useful lifetimes measured in months
What hardware is this? Most hardware including GPUs are cycled between 5 and 8 years.
2 days ago
> hardware that has commercially useful lifetimes measured in months
What hardware is this? Most hardware including GPUs are cycled between 5 and 8 years.
A gpu from 8 years ago is cost competitive, efficient and "worth using" for modern tasks?
I don't want to be picky, but there is still a lot of value left in "not modern" tasks, like video encoding/transcoding. If somewhere the trickle-down effect is real, then it is computing hardware. Take Hetzner's server auction. If the hardware is physically deployed and running, you just need to find appropriate payloads/customers. https://www.hetzner.com/sb/
We have a box at work for employees bring hardware in they’re getting rid of, along with hardware we’re throwing out that we don’t need anymore.
It has a pile of GPUs that are completely obsolete for any task: they use way too much power, have a large form factor that burns up a PCIe x16 slot, are loud, some need extra power cables, lack driver support on modern operating systems, and in return for all that don’t have as much power as something much better you could get for $100.
Value on eBay seems to be about $10-$15, mostly for people with a retro computing hobby or people removing semiconductor components for other purposes.
An obsolete data centre isn’t worth much either. (We have a small one made from equipment being liquidated from local data centres that have been upgraded.) The power consumption is too high and it is not set up for efficient HVAC for modern ultra high power draw workloads.
1 reply →
The V100 is ~8 years old and AFAIK mostly not that common anymore, but the A100 is ~5.5 years old now and is still very commonly used, it's maybe the most common HPC cluster GPU. On the consumer side, 3090s are still very popular, representing a good balance between cost, performance and efficiency (this is mostly due to 4090s and 5090s being much more expensive).
The GPUs have a much shorter lifecycle, on the order of ~3 years.
Exactly, I'm a mechanical engineer and I still have tools given to me by my machinist great uncle from WWII that are not only functional, they're identical to a new tool I'd buy today for that purpose, from the same manufacturer. This is the difference the OP was highlighting
9 replies →
No they don't. The 3 year number came from some random person on the internet who claimed to be a Google employee and was denied by Google, as you can see on any of the articles about this claim:
> Recent purported comments about Nvidia GPU hardware utilization and service life expressed by an “unnamed source” were inaccurate, do not represent how we utilize Nvidia’s technology, and do not represent our experience.
Data centre hardware is more like 3 years.
No they don't. 5-8 years is common. The source for the 3 year number is an unnamed random person claiming to be a Google engineer, and Google specifically reached out to all the journalists publishing that claim with this response.
> Recent purported comments about Nvidia GPU hardware utilization and service life expressed by an “unnamed source” were inaccurate, do not represent how we utilize Nvidia’s technology, and do not represent our experience.
Data centers might be, GPUs not really. No one needs GPUs from 8 years, and hardly even 5.
2 replies →
It dropped down, but is now being extended to 5 years.
The 3 year number was reported by the FT this year from a bunch of companies, where they were saying their accountants are extending to 5 years.
Another example, MS just moved from 4 years to 5.
7 replies →
So like, up to 100 months.