Meteor Lake was taped out in May 2021 and launched in December 2023. Still much slower than the pace of LLM development, to be fair. It seems more like an “if you build it, they will come” approach. But that’s also how we got stuck with (for most consumer purposes) useless tensor cores on our GPUs. Does anyone even give a shit about raytracing/DLSS anymore?
That middle paragraph is very misleading. It’s Generative AI as a service that is actively harmful to the environment. Having a 15 W chip to do tasks like erasing objects from a photo is not any more harmful to the environment than a GPU that uses 15W. In fact, NPUs can be more efficient at some tasks than GPUs.
The problem is opening your phone/browser, and being able to call on demand GPT-4 to wake up a cluster of 128 Nvidia A100s operating at around 300-400W each. That’s 51.2 kW.
Now you can draw some positives and negatives from that figure, such as
Given that an iPhone 15 Pro’s A17 has a thermal design power of 8 W, GPT-4 on the server is about 6400 more energy intensive than anything you can do on an iPhone. 10 seconds of GPT need a similar amount of energy to an iPhone 15 Pro operating flat out at maximum power for 18 hours. Now in those 10 seconds, OpenAI says they “handle multiple user queries simultaneously”, but still - we’re feeding the machine.
51.2 kW is also roughly how much power a large SUV needs to roll at constant speed on a motorway. Each of those large clusters uses a similar amount of energy to a single 7-seater SUV, but serving many users at the same time. Plus unlike cars, a large portion of their energy usage comes from renewables. So yes, I agree that it’s a significant impact but largely overrepresented and we have bigger fish to fry; personal transport is a way bigger issue.
Because the NPUs were designed and built and included long before Windows 11’s AI features were announced?
If I recall correctly, it typically takes about 4 years for a CPU to go from design to distribution.
Meteor Lake was taped out in May 2021 and launched in December 2023. Still much slower than the pace of LLM development, to be fair. It seems more like an “if you build it, they will come” approach. But that’s also how we got stuck with (for most consumer purposes) useless tensor cores on our GPUs. Does anyone even give a shit about raytracing/DLSS anymore?
It actually sounds like Microsoft is betraying Intel for Qualcomm, since their upcoming processor in the new Surface tablet is the only one that actually meets the requirements. So it looks like Microsoft doesn’t give two shits about supporting existing hardware either way.
Tensor cores can be used to play chess, generate images, do realistic text to speech, do noise cancellation, content-aware fill, etc.
They are only useless to you and other people with no imagination
Chess engines have outplayed humans for thirty years, and they didn’t need teraflops of computing power to do it.
Generative AI is actively harmful to the environment, slowing the phase-out of coal in the US and guzzling billions of gallons of water. It’s likely going to kill jobs and it’s already filling the internet and the academic world with garbage. It’s also likely a bubble that will burst before long, potentially bringing the economy down with it.
I’ll give you noise cancellation and text-to-speech, that’s pretty cool.
But personally, I’d rather have more CUDA cores.
That middle paragraph is very misleading. It’s Generative AI as a service that is actively harmful to the environment. Having a 15 W chip to do tasks like erasing objects from a photo is not any more harmful to the environment than a GPU that uses 15W. In fact, NPUs can be more efficient at some tasks than GPUs.
The problem is opening your phone/browser, and being able to call on demand GPT-4 to wake up a cluster of 128 Nvidia A100s operating at around 300-400W each. That’s 51.2 kW.
Now you can draw some positives and negatives from that figure, such as