I’m pretty sure the thing about datatypes is wrong. From experience programming shaders the most typical float values were 4 bytes. The physics simulations are run on cpus typically, not gpus, but for graphics processing of all kinds, smaller floats are used. The conclusion is right though.
That doesn’t really matter. Both AI hardware and consumer hardware use the same underlying materials and components. When supply is low due to AI companies hoarding compute, the prices for all the components rise.
If you are being slightly serious, I hate to break it to you, but the HPC equipment that is taking up all manufacturing capacity has limited consumer use. It’s just so power hungry, specialized, and heavy.
People replying to this with “muh enterprise” don’t understand that I can still plug in a Supermicro server with the correct power input, spin up Linux, and do whatever the hell I want with my giga ultra 8x NVlink GPU 2X Xeon CPU morbillion dollar server.
Now even though it will be sold for a measly couple thousand dollars after the collapse, it doesn’t mean I can’t burn through my unemployment savings to have fun lol.
That’s quite the tall order for a lot of the now ‘bread and butter’ of this bubble. Some of these systems are needing over 25 kilowatts now, being fed by 8 C19 connections.
And that’s ignoring the ones that just dump a crapton over a 48VDC bus bar and don’t have anything even vaguely approaching decent residential electrical hookup friendly.
While most will bemoan that the systems will be too loud and the gpus not really interesting for gaming, the sheer power supply issue is beyond what we’ve seen in datacenters historically. Usually the most egregious was still a few kilowatts over C19, which awkward but doable in residential in certain circumstances. Now we are talking dedicating half your residential service to breakers and outlets to feed the servers.
But the computer stuff will be so cheap on the second hand market!
LAN party in the alley when we lose everything!
Sadly no. AI hardware does not relies on the same fundamentals as consumer hardware.
Long story short AI stuff use Float 4 or 8 because accuracy is not a factor. Games or physics simulation use Float 32 or 64.
I’m pretty sure the thing about datatypes is wrong. From experience programming shaders the most typical float values were 4 bytes. The physics simulations are run on cpus typically, not gpus, but for graphics processing of all kinds, smaller floats are used. The conclusion is right though.
4 bytes not 4 bits, so 32 bits. You can’t do graphics on 4 bits, that’s way too small.
Good catch, thanks
Yeah, even the Game&Watch was 8-bit
But you can use 4bits neural networks.
Some people already replied but I would like to add that even if float type where the right one.
AI dedicated hardware doesn’t bundle stuff you usually espect from a gpu like dedicated encoder and decoders.
The “GPUs” used in these AI datacentres can’t even do graphics anymore. They’re now sloppy approximate matrix math machines.
SAS hard drives and server ECC ram incoming!
That doesn’t really matter. Both AI hardware and consumer hardware use the same underlying materials and components. When supply is low due to AI companies hoarding compute, the prices for all the components rise.
If you are being slightly serious, I hate to break it to you, but the HPC equipment that is taking up all manufacturing capacity has limited consumer use. It’s just so power hungry, specialized, and heavy.
Nvidia and rest of manufacturers will set fire to abandoned data centers before allowing a big second hand market nuking their profits
This is where I hope China swoops in with cheap consumer hardware. Might take a few years though.
They’ll want to, but they don’t own the data centers and the ones that do have no concept of sacrifice for loyalty. Those parts will be sold.
People replying to this with “muh enterprise” don’t understand that I can still plug in a Supermicro server with the correct power input, spin up Linux, and do whatever the hell I want with my giga ultra 8x NVlink GPU 2X Xeon CPU morbillion dollar server.
Now even though it will be sold for a measly couple thousand dollars after the collapse, it doesn’t mean I can’t burn through my unemployment savings to have fun lol.
That’s quite the tall order for a lot of the now ‘bread and butter’ of this bubble. Some of these systems are needing over 25 kilowatts now, being fed by 8 C19 connections.
And that’s ignoring the ones that just dump a crapton over a 48VDC bus bar and don’t have anything even vaguely approaching decent residential electrical hookup friendly.
While most will bemoan that the systems will be too loud and the gpus not really interesting for gaming, the sheer power supply issue is beyond what we’ve seen in datacenters historically. Usually the most egregious was still a few kilowatts over C19, which awkward but doable in residential in certain circumstances. Now we are talking dedicating half your residential service to breakers and outlets to feed the servers.