When are we getting dedicated AI cards?
Why would we want them when GPUs are faster?
>>107791941Discrete NPUs are already available. Dell lets you add one (the Qualcomm AI-100) to the Pro Max 16 Plus. It increases the price by a couple thousand dollars, of course.
>>107791941last year? various dedicated NPU's have been out for so long that it's been a few years since google stopped supporting their Coral NPU's that run object detection on my self-hosted NVR
>>107791941>When are we getting dedicated AI cards?Never. By the time neuromorphic hardware becomes practical, you'll either be dead in some ditch (likely blown apart by AI drones, probably) or too poor to afford food, let alone more consoomer slop.
>>107791941In 2-4 years
>>107791958GPU are faster but not energy efficient, if you are going to use AI one the go with a laptop its much better an NPU.
>>107792070>>107794523I will make it happen. Screencap this.
>>107791941Why does it say "Analog Devices" when it's clearly digital?
>>107794846"Digital" doesn't exist physically.It's an interpretation of an analog signal.
>>107791941inference/datacenterl GPUs were around for a while
>>107794883"Analog" doesn't exist physically.It's an interpretation of discrete neuronal firing events. "Physical" doesn't exist physically, either, by the way.
>>107794926>>107794883This morning my body ejected a physical anal-log and since my printer was all out of paper, I had to wipe digitally.
>>107791941If local AI is ever going to become actually useful for something then I do expect dedicated accelerators to pop up. I think GPUs are just being used right now because they're programmable to some degree and are already here and ready to be put into action, but specialized hardware for the required tasks is bound to pop up eventually.
>>107791941Theres plenty of startups producing them, they are annoying since they all have their own way of converting models to their proprietary format. I hope RISC-V comes up with some Intel AMX equivalent so we can have generic tensor accelerators that can be used with nothing but a C compiler or Rust if you are a troon.
>>107791941im quite excited for them theyre going to become more popular, nvidia has said theyre going to be making them im pretty sure
>>107791941>>107796005they already exist but they're all produced by small-ish companies that can't keep up with development, both software and hardware so they get massively eclipsed by nvidia. Also nvidia just bought the most promising company called groq (not grok). Any real competitive accelerator will probably be snapped up by nvidia/broadcom/google cus they're flush with cash.
