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".

Then the metric should be Watts per Intelligence, not the contrary.

We propose miles per hour (MPH), distance travelled per unit of time

Stupid metric. It‘s not because you spend more time that you travel farther.

s/

Not a well-thought-out take.

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).

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.

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.

In my case the bad internet also tips the scale.

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.

Under ideal conditions with outside temperatures*

But yes

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.

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.

I rather like this paper, but I think it is generous to say their benchmarks measure "intelligence".

[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