Direct Current Data Centers

[jabl]

To comment on using direct current rather than alternating current, at least NVIDIA is working on 800VDC:

https://developer.nvidia.com/blog/nvidia-800-v-hvdc-architec...

https://developer.nvidia.com/blog/building-the-800-vdc-ecosy...

More generally, a bunch of the electrical industry heavyweights are involved in a standard for DC microgrids called Current/OS: https://currentos.org/

There are certainly advantages to DC vs AC, but there's of course a huge amount of sunken cost in contemporary AC networks.

[boulos]

IIRC, Google contributed +-400 racks last year to Open Compute: https://cloud.google.com/blog/topics/systems/enabling-1-mw-i...

[bob1029]

From a purely engineering perspective I think it becomes difficult to argue with the gas turbine once you get into the gigawatt class of data center. The amount of land required for this much solar is not to be understated. In many practical scenarios the solar array would need to be located a distance away from the actual data center. This implies transmission infrastructure which is often the hardest part of any electrical engineering project. You can put a gigawatt of N+1 generation on a 50 acre site with gas. It's dispatchable 24/7/365 and you can store energy for pennies on the dollar at incredible scale.

Having both forms of generation available at the same time is the best solution. Once you put a data center on the grid you can mix the fuel however you want upstream. This should be the ultimate goal and I believe it is for all current AI projects. I am not aware of any data center builds that intend to operate on parking lot generators indefinitely.

[hjoutfbkfd]

they are talking about covering the desert with solar panels. why would you not put the data center in the middle of it?

[dan-robertson]

For inference you don’t need gpus to be clustered together as much (generally training has lots of synchronisation steps so you can be bottlenecked on that instead of ‘real’ work) as they can handle separate tasks in parallel. But maybe other economies of scale still make you want to put them together (and therefore on average further from the power).

I guess there was a bit of thought about transmission with the reference to high voltages. Another interesting thing: batteries allow you to reduce the needed capacity for transmission lines – if you have batteries near generation and then transmit power at a lower maximum, same average rate than if you only have batteries near use, you can more efficiently use the available transmission.

I guess the main reason for gas to be a problem is if you can’t get new generation (eg lack of turbines).

[matt-p]

Sadly, I agree until we get SMRs (I think we are few years off). Obviously it would be more ideal to use grid+solar with curtailment but not super realistic.

[cinntaile]

If you have predictable demand at that scale, nuclear might make more sense than the combination of gas and solar.

[Havoc]

Slightly OT, but I see the Chinese are talking about space DCs now too which would suggest they reckon it could work too. (Unlike me and others here)

[hhh]

datacenters in space are a great way to claim vast amount of viable orbit space for a stupid project to eventually sell the slot for something else when it’s rarer.

[numpad0]

It can't work if you're launching from Earth. Datacenters are too heavy with or without the solar and radiator panels.

If you could make those panels and chips on the Moon, Deimos, Mars, high Jupiter, wherever, then space datacenters can totally work.

[bgnn]

Except none of that data center grade chips can work in the space. No GPUs, no memory, no SSD. They are not radiation-hardened (rad-hard). Rad-hard chips generally cost an oder of magnitude or more compared to normal commercial chips, and they are in general an order of magnitude less complex, plus they operate much lower frequencies. Data centers in space is straight up stupid.

[Galanwe]

Not a physician, but wouldn't space be terrible for heat dissipation?

[alansaber]

I think it's more of a classic mirror move where IF they do work, they're at danger of falling behind.

[phtrivier]

I'm curious about Handmeier's opinion on location of data centers.

Should they be close to the solar arrays (that is, in the desert, with data networks connecting them to were the tokens are used)

Or close to their customers (which mean far from the solar arrays, with electricity networks)

He's talking a lot about removing movable parts, but aren't the wires going to be an limiting factor ?

[bgnn]

Fiber is much much less of a cost and technical challenge compared to transfering GWs of power. Unless the customer cannot handle up to 100ms latency, it's totally logical to place the data centers close to the power source, or vice versa (power source close to the data center).

[hambes]

it is difficult to comprehend for me that soneone spends all this time thinking through and calculating how to harness as much energy as possible and then wants to use it for large language models instead of something useful, like food production, communication, transport or any other way of satisfying actual human material needs. what weird priorities.

[Hendrikto]

Whether you like it or not, we are burning a lot of electricity on datacenters. That is a fact. And energy consumption is likely going to significantly increase in the near future. If we can reduce that energy usage, that is a good thing and a big improvement.

I do not think I even understand your complaint. Different people can work on different problems. We do not have to pick only one.

> My improvement is more important than yours.

We can just do both.

[samus]

LLMs and other IT applications have the distinct advantage that they require no other raw materials as input, aside from initial setup, extension, and maintenance. Under these conditions the requirements essentially boil down to real estate and high bandwidth internet connections. Also, demand for AI is currently so high that the solution can be scaled up far enough to be viable.

All the other concerns require more subtle approaches because human requirements are much more messy.

[sandworm101]

Tell that to the 1000-watt space heater in the corner that i tasked with upscaling some old home movies! Four GPUs worked very hard all night to get footage of my first dog up to 1080p. My living room is a little warm this morning.

[compass_copium]

Well, I've never seen anything written by AI evangelists that doesn't sound like it was written in day three of an adderall binge. This essay is no different.

[stingraycharles]

Sometimes (often) solving the problem is the most fun part, regardless of how it’s used.

The scale of AI energy consumption is quite unique from what I heard, and there’s a lot of money flowing into that direction. So that seems to me a decent reason to think about that.

I haven’t heard yet that food production is constrained by these kind of things.

It appears to make that you’re just taking a cheap jab at AI.

[dan-robertson]

I don’t think that’s a great description of what’s going on here. I think there are two things:

1. The actual thing the authors spend a lot of time thinking about seems to be more generally how to make good use of solar power for things that people find valuable – synthetic fuels desalination, etc – and the implications of the sun only shining some of the time – maybe you don’t want to pay more for more efficient systems as then you want steady power which is more expensive.

2. I think the blog post is a bit of a response to lots of public discussion about AI data centres. IMO seems better to see what someone who thinks a lot about energy has to say than eg, a government suggestion that you delete old pictures to reduce water consumption.

[627467]

I share the reaction but I'm also aware how easy is it to inventivize (aka subsidize) ineffective old processes in the name of "productive" priorities. The problem is not LLM/DC, the problem is food production, transport and communications are not sexy in a "post-scarcity" (entitled/distracted) societies. People take too many things for granted

[fnord77]

the saying goes something like: the brightest minds in the world are getting together to figure out how to deliver more ads

[gruez]

>instead of something useful, like food production, communication, transport or any other way of satisfying actual human material needs. what weird priorities.

You realize that even pre-AI, that this complaint would still hold for most of tech? Adtech, enterprise SaaS, and B2C apps are hardly "actual human material needs". Even excluding tech, the next lucrative sector would be banking, and same complaint would be applicable. In other words, this is a decades (centuries?) old complaint, repackaged for the current thing.

[hjoutfbkfd]

if anything we are producing too much food

and what communications you find lacking?

[ErroneousBosh]

Why are we wasting resources on toy chatbots?

[dang]

"Eschew flamebait. Avoid generic tangents."

https://news.ycombinator.com/newsguidelines.html

[boxed]

If you think this is what LLMs are, then you are a bit behind the times. Opus 4.5 is a huge step up. The previous generation was good for starting basic hobby projects, now we can do pretty big time-consuming changes with it.

I have been extremely skeptical and dismissive of LLMs for a long time, but after a certain level of improvement you have to realize that at least for programming the advantages are substantial.

[Joel_Mckay]

Borrowing state money that ultimately indentures a country with over-engineered massive boondoggle projects.

That regulatory capture con strangled more emerging economies than most like to admit. =3

"The Dictator's Handbook: Why Bad Behavior is Almost Always Good Politics" (Bruce Bueno de Mesquita, Alastair Smith)

[alansaber]

Because fusion energy isn't cool anymore.

[adamsb6]

Why are we wasting resources hosting countless replicas of alt.tv.simpsons?