3 ms·

Commuting for work with a car is mostly a US problem. And here we are, talking about the US. Great, that's settled. OTOH, considering the proliferation of th

Commuting for work with a car is mostly a US problem. And here we are, talking about the US. Great, that's settled. OTOH, considering the proliferation of the EVs, greenhouse gas emissions from cars are going down In the US, most electricity comes from carbon-based sources. EVs help, certainly, but it would still be best from an ecological standpoint if people didn't have to use them every day to get to work. OTOH, many datacenters are installing their own greenhouse emitting power gensets. Exactly, which is why we should work for grid renewal and expansion at the same time we build these things out. If the hyperscalers insist on building their own power plants, those power plants should be required to be (a) carbon-neutral; and (b) grid-tied. These GPUs won't be scrapped soon, because besides AI, they are useful for accelerating not-so-precise floating point calculations en-masse. Really. What are some other applications for not-so-precise floating point calculations at the exaflops level?

bayindirh 2mo ago
> And here we are, talking about the US. Great, that's settled. Proliferation of law skirting AI datacenters is a worldwide problem. Many companies are extending this to Europe as well, so AI datacenters is not a U.S. problem only. From what I see in Electricity Maps[0], it's not clear cut. Some of the states have well over 50% of their generation is from carbon-free sources. As a general trend, west is greener than eastern US. So, no, EVs will have an impact there. However, you need electric semis a lot as a country since rail is not used a lot in the US as well. > Really. What are some other applications for not-so-precise floating point calculations at the exaflops level? In short, HPC. As a more detailed, by short list: - Weather and climate modeling. - Drug discovery. - Computational Fluid Dynamics. - Material science, behavioral and process analysis. - All kinds of large scale simulation which can be modeled with matrices. - Any kind of problem requiring linear algebra. While I generally work at full 64bit precision, from what I see, even though these modern GPUs do only support up to FP16 (or FP32 in some cases), they can be used for higher resolution simulations as well. We support some astronomy and climate researchers which use classical algorithms and in-house ML models (not LLMs and such) to accelerate their simulations and predictions on multiple GPUs. I manage HPC clusters, and sometimes develop code on them, so yeah. [0]: https://app.electricitymaps.com/map/zone/US-NW-PACE/live/fifteen_minutes?signal=electricity-price https://app.electricitymaps.com/map/zone/US-NW-PACE/live/fif...
- CamperBob2 2mo ago
  OK, I see I should have used "trillions of dollars worth of hardware capable of performing not-so-precise floating point calculations at the exaflops level" as the basis for my question. The hyperscalers are making a singularity-or-nothing bet on this stuff; there are no other uses for this much hardware of this type. If there were, such hardware would already be in place. (IMHO, the fact that AI models tolerate absurdly-low precision means that we're doing something wrong and wasteful in our entire computational approach to the field, but that's neither here nor there.)

bayindirh 2mo ago

> And here we are, talking about the US. Great, that's settled. Proliferation of law skirting AI datacenters is a worldwide problem. Many companies are extending this to Europe as well, so AI datacenters is not a U.S. problem only. From what I see in Electricity Maps[0], it's not clear cut. Some of the states have well over 50% of their generation is from carbon-free sources. As a general trend, west is greener than eastern US. So, no, EVs will have an impact there. However, you need electric semis a lot as a country since rail is not used a lot in the US as well. > Really. What are some other applications for not-so-precise floating point calculations at the exaflops level? In short, HPC. As a more detailed, by short list: - Weather and climate modeling. - Drug discovery. - Computational Fluid Dynamics. - Material science, behavioral and process analysis. - All kinds of large scale simulation which can be modeled with matrices. - Any kind of problem requiring linear algebra. While I generally work at full 64bit precision, from what I see, even though these modern GPUs do only support up to FP16 (or FP32 in some cases), they can be used for higher resolution simulations as well. We support some astronomy and climate researchers which use classical algorithms and in-house ML models (not LLMs and such) to accelerate their simulations and predictions on multiple GPUs. I manage HPC clusters, and sometimes develop code on them, so yeah. [0]: https://app.electricitymaps.com/map/zone/US-NW-PACE/live/fifteen_minutes?signal=electricity-price https://app.electricitymaps.com/map/zone/US-NW-PACE/live/fif...

CamperBob2 2mo ago
OK, I see I should have used "trillions of dollars worth of hardware capable of performing not-so-precise floating point calculations at the exaflops level" as the basis for my question. The hyperscalers are making a singularity-or-nothing bet on this stuff; there are no other uses for this much hardware of this type. If there were, such hardware would already be in place. (IMHO, the fact that AI models tolerate absurdly-low precision means that we're doing something wrong and wasteful in our entire computational approach to the field, but that's neither here nor there.)