Incredibly important research. We've reached the point where local LLMs are good enough! It takes less time for local model to take the first action on your task than it does for Claude to validate your login, put you into queue and start issuing the commands. Local models are persistent and 100% predictable unlike any cloud offering. It's better for the power system for the demand to be distributed. During the winter time the GPU also doubles as a 300W in-house heater. Not to mention avoiding personal data collection and re-selling.
> We've reached the point where local LLMs are good enough!
For some tasks, yes. For most of my deeper work they're not even close to my subscriptions.
> It takes less time for local model to take the first action on your task than it does for Claude to validate your login, put you into queue and start issuing the commands.
I have some decent LLM hardware here and I strongly disagree with this. Claude responds quickly. Using Fable or Opus it will deliver a working result faster than my local models because it gets there in fewer tokens. That's just how it is.
> During the winter time the GPU also doubles as a 300W in-house heater.
This is a curse in the summer. I'm feeling it right now.
Of course, but the power dissipated in a data center provide zero heat for your house.
(Tangent: I wonder why we haven't seen deployment of organic Rankin cycle generators in AI data centers, the exhaust temperature should be compatible and that could yield a 10-20% energy bill saving).
> I wonder why we haven't seen deployment of organic Rankin cycle generators in AI data centers
Because there was never a long term plan for AI data centers. It's an AI market capture and cash grab scheme that ends when local models eat their lunch.
That makes no sense. Do you think their plan is to spend billions of dollars and build infrastructure mega projects just to let it sit idle?
Did the rise personal computing make data centers and super computers obsolete?
Any advance in inference that allows local models to do the job will also benefit hyperscalers. Imagine the sheer amount of compute they could throw at problems if each 32GB of VRAM was enough for frontier reasoning.
Easier said than done, but if you water-cool the GPU just upstream of your domestic water heater, it wouldn't be a bad thing. Perhaps using coolant and a counterflow heat exchanger, rather than the potable water, but the point stands. Would just need a secondary tank as a buffer (able to soak up heat from the GPU at all times, even when there's no demand for hot water) which then flushes in when demand exists.
They use it in the city heating system, which is obviously the optimal way of reusing the heat but most data centers aren't put next to such a heating system.
I'm talking about making electricity back from the heat (using a low-temp thermodynamic cycle). It has a low yield (due to the low input temperature) but it's usually economically viable when using heat that would end up in the heavens anyway.
> During the winter time the GPU also doubles as a 300W in-house heater
There could be a service that works in reverse where if someone needs a heater for a few months, they could rent a portable server (e.g. using older repurposed GPUs) with a built-in 5G modem that would run inference on LLM queries. As an incentive perhaps renting itself could be free (or you could earn money?), but you'd still have to pay your electricity bill.
I saw a company pitching almost exactly this on linkedin earlier this month. Except not as a portable heater but as your house's central heating system.
For sure but aren’t there more regulations and building codes around heaters? It’s been close to a decade since I took heat and mass transfer courses but this seems like saying you can use your oven as a heater for your house in the wintertime.
> It takes less time for local model to take the first action on your task than it does for Claude to validate your login, put you into queue and start issuing the commands.
This is only true if your local model is already resident in RAM / VRAM.
it really does feel like 'intelligence' is becoming so much more of a marketing term moreso than it is any actual definitive measure. I was acquaintances with a cognitive science post-doc many years ago who trained under the same program that Douglas Hofstadter was part of (and who he saw as a great popular science communicator but perhaps a bit over-rated when it came to his own research)
my very naive question to him back then was how close we were to understanding human cognition. we were both fans of grand strategy games (though the few hundred hours of Stellaris I played vs his thousands in EU4 paled in comparison) and I was asking if it was possible to map human cognition to the same array of interdependent logical chains-of-reasoning that games like that could be boiled down to
his answer, in short, was 'we are so, so, so far away and no, that's, at best, a reductive mental model of intelligence'
I keep that conversation in mind whenever I hear about all this talk of AGI - that realistically we're so far away from actual AGI in the same way that the inventor(s) of the wheel were from a gas-powered car, and there's many paradigm shifts to go in how we even understand what the nature of intelligence is before we get there
I tried calculating historical "intelligence per cost" recently but stopped when I realized intelligence is not linear. For any meaningful "x per y" you can just double "y" if you have a half as efficient system to get the same result but so-called intelligence doesn't work like that.
Very cool paper and some interesting things for local serving. Though its a little apples-to-oranges we do publish live energy stats for all models on our service here https://portal.neuralwatt.com/energy-pricing in case you are interested in what this looks like on the cloud side. FWIW DSV4.1 flash is really getting popular due to its IPW.
Some of the items like model routing, if you do it per request instead of per session, can break down on the cloud from an energy and cost POV since one of the best things you can do for both is to maintain the KV cache which both reduces time component of energy and the quite expensive prefill energy.
I am keen on the future where we have local/cloud hybrid serving which is cache aware. I do think that could be the best use of energy resources for AI.
ok now let's do some math to calculate the current levels of daily intelligence to calculate how much power we need to replace 10% of it with AI. Never mind new use cases or approaching human replacement. My guess is it is a few orders of magnitude more power than we produce today.
Saw some measurements on SBC NPUs (3588) and that did seem to have a decent win on power over CPU...but also a perplexity loss relative to CPU so think this will prove quite hard to reliably quantify in practice.
> Unless I misread it, are they saying local GPUs use less energy?
You misread it. From the abstract:
> local accelerators achieve at least 1.4× lower IPW than cloud accelerators running identical models
That's "intelligence per watt". They also have IPJ, per Joule.
So, they find local is 40% "dumber" than cloud for the same power or 40% more power for the same "intelligence".
Tables 13 and 14 summarize their IPW and IPJ metrics.
But, to your actual point, I think the "local is 40% dumber per watt than cloud" message is still an understatement. And maybe this is something I failed to find in the paper but they seem to ignore the "idle baseline" costs and talks about explicitly focusing on the power consumption of just the accelerator under load.
There is a large baseline power consumption just to support the accelerator. CPUs, memory, PS losses, network, fans, general environment cooling. This "cost floor" is different for data centers and a "random local computer" and I think must be in favor of data centers which are designed and built with efficiency in mind.
Idleness should also be considered. My local GPUs at $WORK and home are idle more than they are used. Idle time energy in real world scenarios should be somehow attributed to those brief, punctuated times when LLM functions are actually active on the accelerator. Actual, local LLM usage of a GPU is brief (assuming one user per PC). Even with my heavy usage developing s/w I'd guess I heat up a GPU about one hour per day total, sometimes much less. If that is local then one must pay 23 hours of idleness for that 1 hour of "intelligence". Of course a local PC is used for other things and the idleness penalty must somehow account for that. OTOH, data centers try to maximize utilization so their idle time penalty would be much less, perhaps close to zero, by construction.
You can most definitely batch local models and do unattended inference on a 24/7 basis to maximize utilization on local hardware too. The limits are usually set by some combination of memory utilization for KV cache (particularly on small dGPUs) and overall thermals/power limits (particularly on iGPUs with unified RAM/VRAM). (If you're not near thermal limits, the main alternative to batching is to use MTP or speculative decoding in order to raise arithmetic intensity and speed with the same memory utilization. But batching requests is generally viewed as preferable.) Newer models, especially from DeepSeek, do a nice job of reducing KV cache memory impact for any given context length and/or amount of parallel sessions, so batching on local hw really ought to be quite feasible.
Small models are much smaller than frontier models though, which is how they end up consuming less energy despite low batch count. (Though with local models growing strong agentic capabilities, batching becomes a reality with local models as well).
this is the metric i've been waiting for. we run everything local (ollama + neo4j) for compliance reasons, so 'quality per watt' is literally our budget line. one data point from our setup: qwen2.5:3b on an m2 macbook handles nl-to-cypher for simple graph schemas at ~3-5s per answer, and the energy cost is a rounding error compared to shipping the same queries to a frontier api. the hard part was never the model though, it was parsing pdfs locally without a vision model. would love to see parsing/ocr covered in future benchmarks.
I don't think that's what they are saying. In fact if they did the metric would be pointless. Rather they are saying by estimating that value on different architectures, one can find more efficient ones. They use open model to be able to remove unknowns. They aren't advocating for one model or another, only more efficient architectures.
Intelligence per Joule would be more appropriate in many cases. If a model can do the same work but takes 10 times as long as a bigger one that can still be useful (e.g. due to memory constraints), but at the same wattage it burns 10 times the energy.
Even more so on mobile devices.
They also define and measure an "IPJ" as well as "IPW"
> the NVIDIA B200 achieves 1.6× to 2.3× higher intelligence per joule
than the APPLE M4 MAX across QWEN 3 and GPT-OSS model variants
The B200 = "cloud", M4 = "local".
So "cloud" does even better in energy than it does in power compared to "local". Or, to flip it, "local" is both slower and more expensive than "cloud".
Yes, there is a massive difference, and you cannot invert them to get the other.
Watt per Intelligence means that you have a fixed, deterministic, measure of intelligence, and you calculate how many watts it takes to get there.
If your goal is to measure which model can reach a specific outcome with the least energy possible (which is what GP says the goal is for this metric), then you cannot have a variable outcome, which is what intelligence per watt describes. As opposed to watt per intelligence, where the outcome is fixed and the numerator defines how much energy expenditure is needed to reach this fixed outcome.
For some tasks, yes. For most of my deeper work they're not even close to my subscriptions.
> It takes less time for local model to take the first action on your task than it does for Claude to validate your login, put you into queue and start issuing the commands.
I have some decent LLM hardware here and I strongly disagree with this. Claude responds quickly. Using Fable or Opus it will deliver a working result faster than my local models because it gets there in fewer tokens. That's just how it is.
> During the winter time the GPU also doubles as a 300W in-house heater.
This is a curse in the summer. I'm feeling it right now.
Just a reminder that heat pumps can consume 300W of electricity to provide 1200W of heat.
(Tangent: I wonder why we haven't seen deployment of organic Rankin cycle generators in AI data centers, the exhaust temperature should be compatible and that could yield a 10-20% energy bill saving).
Because there was never a long term plan for AI data centers. It's an AI market capture and cash grab scheme that ends when local models eat their lunch.
Did the rise personal computing make data centers and super computers obsolete?
Any advance in inference that allows local models to do the job will also benefit hyperscalers. Imagine the sheer amount of compute they could throw at problems if each 32GB of VRAM was enough for frontier reasoning.
https://eu-mayors.ec.europa.eu/en/news/stockholm-sweden-heat...
I'm talking about making electricity back from the heat (using a low-temp thermodynamic cycle). It has a low yield (due to the low input temperature) but it's usually economically viable when using heat that would end up in the heavens anyway.
But yes
There could be a service that works in reverse where if someone needs a heater for a few months, they could rent a portable server (e.g. using older repurposed GPUs) with a built-in 5G modem that would run inference on LLM queries. As an incentive perhaps renting itself could be free (or you could earn money?), but you'd still have to pay your electricity bill.
Please correct me if I am wrong.
A bit slow for agentic coding of course but fine for any chatbot use-case.
This is only true if your local model is already resident in RAM / VRAM.
[0] https://arxiv.org/pdf/1911.01547
We have no dang clue what intelligence is, nor how to measure it.
[1] https://taylor.town/crowpower
- Sam Altman[0]
"Can you define intelligence?"
"Yes, it is this many moneys."
[0] https://www.businessinsider.com/sam-altman-ai-utility-electr...
my very naive question to him back then was how close we were to understanding human cognition. we were both fans of grand strategy games (though the few hundred hours of Stellaris I played vs his thousands in EU4 paled in comparison) and I was asking if it was possible to map human cognition to the same array of interdependent logical chains-of-reasoning that games like that could be boiled down to
his answer, in short, was 'we are so, so, so far away and no, that's, at best, a reductive mental model of intelligence'
I keep that conversation in mind whenever I hear about all this talk of AGI - that realistically we're so far away from actual AGI in the same way that the inventor(s) of the wheel were from a gas-powered car, and there's many paradigm shifts to go in how we even understand what the nature of intelligence is before we get there
Some of the items like model routing, if you do it per request instead of per session, can break down on the cloud from an energy and cost POV since one of the best things you can do for both is to maintain the KV cache which both reduces time component of energy and the quite expensive prefill energy.
I am keen on the future where we have local/cloud hybrid serving which is cache aware. I do think that could be the best use of energy resources for AI.
That’s surprising, almost unbelievable, due to batching. Local is usually not batched.
You misread it. From the abstract:
> local accelerators achieve at least 1.4× lower IPW than cloud accelerators running identical models
That's "intelligence per watt". They also have IPJ, per Joule.
So, they find local is 40% "dumber" than cloud for the same power or 40% more power for the same "intelligence".
Tables 13 and 14 summarize their IPW and IPJ metrics.
But, to your actual point, I think the "local is 40% dumber per watt than cloud" message is still an understatement. And maybe this is something I failed to find in the paper but they seem to ignore the "idle baseline" costs and talks about explicitly focusing on the power consumption of just the accelerator under load.
There is a large baseline power consumption just to support the accelerator. CPUs, memory, PS losses, network, fans, general environment cooling. This "cost floor" is different for data centers and a "random local computer" and I think must be in favor of data centers which are designed and built with efficiency in mind.
Idleness should also be considered. My local GPUs at $WORK and home are idle more than they are used. Idle time energy in real world scenarios should be somehow attributed to those brief, punctuated times when LLM functions are actually active on the accelerator. Actual, local LLM usage of a GPU is brief (assuming one user per PC). Even with my heavy usage developing s/w I'd guess I heat up a GPU about one hour per day total, sometimes much less. If that is local then one must pay 23 hours of idleness for that 1 hour of "intelligence". Of course a local PC is used for other things and the idleness penalty must somehow account for that. OTOH, data centers try to maximize utilization so their idle time penalty would be much less, perhaps close to zero, by construction.
Stupid metric. It's not because a model is better performing that it necessarily requires more energy or compute.
> the NVIDIA B200 achieves 1.6× to 2.3× higher intelligence per joule than the APPLE M4 MAX across QWEN 3 and GPT-OSS model variants
The B200 = "cloud", M4 = "local".
So "cloud" does even better in energy than it does in power compared to "local". Or, to flip it, "local" is both slower and more expensive than "cloud".
Watt per Intelligence means that you have a fixed, deterministic, measure of intelligence, and you calculate how many watts it takes to get there.
If your goal is to measure which model can reach a specific outcome with the least energy possible (which is what GP says the goal is for this metric), then you cannot have a variable outcome, which is what intelligence per watt describes. As opposed to watt per intelligence, where the outcome is fixed and the numerator defines how much energy expenditure is needed to reach this fixed outcome.
Stupid metric. It‘s not because you spend more time that you travel farther.
s/