493
Virginia county asks all employees, including schools, to conserve power due to AI-driven electricity price hikes — state's 400-plus data centers steadily increasing demand, grid expansion, and prici… (www.tomshardware.com)
submitted 2 months ago by to c/technology@lemmy.world
61 comments fedilink hide all child comments
you are viewing a single comment's thread
view the rest of the comments
[–] 2 points 2 months ago (7 children)

You have that backwards. The only thing you gain from running local models is privacy. It is not cheaper, it is not more efficient. You are actively hurting the environment MORE by using a local model on your own. LLM efficiency sky rockets the more users there are on a single loaded model.

IMO the only way we get to efficient LLM usage would be by having very efficient non frontier models running only for its local community to use, where you can have assurances on whether its power source is clean or not. That doesn't help with the plagiarism aspect though

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 7 child comments
    •
  • [–] 0 points 2 months ago (6 children)

    Are you serious?

    • Local model: Spends most of its time turned off. Only active when I want it to be active, and only for a little while. Dedicated solely to generating the small amounts of code I use it for. Does nothing else. Costs $0 per token, and electricity costs are negligible.
    • Frontier model: Always on, running on millions of GPUs. Would be burning down the planet even if hardly anyone was using it. Incredibly wasteful, being used for trivial tasks and convincing people that their horrible ideas are visionary every day. Misspelling "strawberry" for the masses. Trained specifically to be addictive. Can easily cost a software developer who is addicted to AI thousands of dollars a month, with the recent price increases.

    I'd love to see some data to back up the assertion that frontier models are somehow cheaper and more efficient than running a model locally.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 6 child comments
    •
  • [–] 3 points 2 months ago (2 children)

    You're probably burning more energy turning it off and on again. It doesn't really use any noticeable power sitting idle.

    Anyway, a direct comparison would be pretty difficult because your model is probably tens of billions of parameters, not over a trillion. Energy consumption per output token will probably be a bit higher for the frontier models but something that people have found is that higher quality models often need fewer tokens to achieve the same goal. Plus how many times do you re-prompt your local model vs Claude Fable or Opus for example to get the desired result?

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 2 child comments
    •
  • [–] 0 points 2 months ago (1 child)

    You’re probably burning more energy turning it off and on again. It doesn’t really use any noticeable power sitting idle.

    I am absolutely not burning more energy than a frontier model by doing things like putting my laptop to sleep or shutting down unused services when I want to conserve battery power.

    Anyway, a direct comparison would be pretty difficult because your model is probably tens of billions of parameters, not over a trillion.

    True.

    Energy consumption per output token will probably be a bit higher for the frontier models but something that people have found is that higher quality models often need fewer tokens to achieve the same goal.

    That's actually not true. In fact it's much the opposite. Frontier models churn through tokens at a much higher rate, because of their higher complexity and higher number of parameters. Research is still new on this, but having a frontier model analyze your code files versus a small, local model for the same task seems to be enormously wasteful. If you must use a frontier model for something, have it do that work after receiving the output from an agent using a small model to read and summarize your code.

    Plus how many times do you re-prompt your local model vs Claude Fable or Opus for example to get the desired result?

    ...Almost never? I'm not a fan of letting AI do much of ANY of my coding, because it will inevitably bloat my codebase with garbage regardless of which model I use. So I severely restrict my model usage to simple, clearly-defined, narrow-scoped tasks that can save me a bit of time, and that's it. With guardrails and discipline like that, I barely ever have the need to re-prompt.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [–] 1 point 2 months ago

    I am absolutely not burning more energy than a frontier model by doing things like putting my laptop to sleep or shutting down unused services when I want to conserve battery power.

    I was under the impression you keep loading the model into VRAM and unloading it when finished using it, I meant it's less power efficient than just keeping it in VRAM.

    That’s actually not true. In fact it’s much the opposite. Frontier models churn through tokens at a much higher rate, because of their higher complexity and higher number of parameters.

    Thing is, the input/reading part of it is cheap and wastefully generating extra tokens as output costs you more in energy (or money if using an external service). Put it this way: Claude has historically had 3 models: Haiku (small), Sonnet (medium), Opus (big). Sonnet 5 came out recently and people using Claude Code have reported that it's so verbose, it's now more expensive to use for the same task than Opus, which has much bigger costs per Mtok. That would mean it probably also uses more energy than the bigger model.

    …Almost never? I’m not a fan of letting AI do much of ANY of my coding, because it will inevitably bloat my codebase with garbage regardless of which model I use. So I severely restrict my model usage to simple, clearly-defined, narrow-scoped tasks that can save me a bit of time, and that’s it. With guardrails and discipline like that, I barely ever have the need to re-prompt.

    At that point, why bother with a local model, you could use Deepseek V4 flash and probably spend less than a tenner a month on it. It's surprisingly capable (I mean sometimes you can barely tell it's not a frontier model) and costs next to nothing to use.

    If you must use a frontier model for something, have it do that work after receiving the output from an agent using a small model to read and summarize your code.

    It's sort of what my workflow does when I use OpenCode. Bigger model (GLM-5.2 or GPT-5.5 depending on which one hasn't run into its usage limit) reads my prompt, the .md files describing the repo and the overall file structure of the repo, then fires off parallel DeepSeek V4 Flash scouts on usage credits to read and summarize the files as needed. The big model then does the planning and again DeepSeek V4 Flash is the one to execute it via subagents. The subagents running DeepSeek usually come back with 1-2 cents in cost.

    I did try a Qwen-3.6 distillation locally and it was pretty capable in terms of output, but it's more expensive for me than the DeepSeek Flash on API usage costs, since electricity isn't free here and my GPU is 2 generations old. And it's slow as hell, since it has to offload a lot to CPU/RAM over GPU/RAM.

    The big models I only use as subscriptions that I'm prepared to end at any moment if they reduce the usage I get. Let the AI companies eat the cost, I'll never pay them API pricing if they want 20 or 30 dollars for a million output tokens.

  • permalink
  • fedilink
  • source
  • parent
  • [–] 1 point 2 months ago (2 children)

    Very serious. Your personal amount of usage means nothing at all in this conversation. It is entirely about tokens per watt. The amount of energy the memory operations involve scale incredibly well when people are accessing the same object in memory simultaneously. Last I looked it was around a 10x difference for the same models efficiency.

    If you want me to be your personal search engine you’ll need to wait a bit, im making dinner right now and would rather look for the articles on my desktop.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 2 child comments
    •
  • [–] 2 points 2 months ago (1 child)

    Very serious. Your personal amount of usage means nothing at all in this conversation. It is entirely about tokens per watt. The amount of energy the memory operations involve scale incredibly well when people are accessing the same object in memory simultaneously. Last I looked it was around a 10x difference for the same models efficiency.

    Hold up. Are you talking about caching? Because if you are... yeah. That has nothing to do with the model and everything to do with the service layer around the model. The same service layers can be - and have been - implemented in tools like Lemonade Server, llama.cpp, Ollama, etc.

    And I really do want to know your sources.

    Mine say GPT 5.5 is probably using quite a lot more than 0.34 Wh per query (0.34 Wh is what Sam Altman claimed for the then-current version of GPT in June of 2025, but he hasn't released numbers since then and no one has done an independent analysis). With Claude, an independent estimate from last year pegged Sonnet at 0.8 Wh for a short prompt, 2.8 Wh for a medium one, and 5.5 Wh for a long one. Current numbers are, again, almost certainly much higher. And just for fun, there's DeepSeek (which I've never used and never would use), with the reasoning-tuned DeepSeek-R1 hitting a whopping 29 Wh for a complex query.

    Meanwhile, small, open models are probably in the 0.07 - 0.2 range, depending on the model, the hardware it's running on, and the nature of the query. Of course, there are much weightier open models too, with ones like Llama 3.1 405B using about 9 Wh for a medium-length prompt. On the other hand... who is going to run that on their local machine?

    Look... If I'm wrong, and using local models the way I do - sparingly and infrequently - really does consume more electricity than using Claude Code, I want to know. I have no problem whatsoever with eschewing AI models entirely, since I despise all of them. But given how tight-lipped OpenAI and Anthropic are about energy consumption per average prompt, and what independent analyses have estimated, I am highly skeptical that they are acting as some sort of paragons of environmental stewardship.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [–] 1 point 2 months ago

    Not talking about caching (though there would be some decent memory savings due to that on general platforms like ChatGPT and tools like Codex). I am talking about large batch sizes, which are concurrent requests all accessing the same memory at the same time. The model is loaded once onto the GPU(s) and then many simultaneous requests can read that memory at the same time. When those requests are all processing their responses simultaneously, the energy per token drops off a cliff.

    And yes, running a smaller model would generally take less power, but thats not really a fair comparison. Small models just wont give you the same results as larger ones. You need to compare it apples to apples. If you want to compare your local Qwen model running on your laptop, you compare those numbers to larger systems supplying that same qwen model to thousands of people. Just because we are comparing cloud services to local doesn't automatically mean GPT 5.6 vs Qwen 3.6 27B. There are plenty of cloud AI providers running all sorts of models and sizes.

    As for one of the articles I learned alot of this from originally, this is one I recommend going through. It really goes deep into the whole topic: https://arxiv.org/html/2601.22076v1

  • permalink
  • fedilink
  • source
  • parent