Every knob the benchmark has turned, on one plane.

Which modelCapability & speed vs energy

One dot per configuration, (model, quantization, engine, node, settings), on a single benchmark, filtered by GPU and board. The first four modes plot a metric against energy per correct answer (down is cheaper); TTFT vs accuracy and TTFT vs tok/s plot responsiveness against capability or throughput, with energy moved into the dot colour. Axis labels name which direction is better; the Pareto front (amber dots joined by a stepped hairline, or ink-ringed dots in the coloured modes) is nothing-measured-is-better-on-both-axes. A hollow dot is a quantized checkpoint.

What is measuredThe four benchmarks

Every task is scored through the same generative path: the model writes its answer as text, and the answer is extracted and checked. The standard alternative for multiple choice, comparing token likelihoods without generating, would leave joules-per-token undefined, so it is not used. Two tasks are prefill-shaped (long prompt, tiny answer) and two are decode-shaped (short prompt, long answer); a decode token costs 40–60× a prefill token, which is why the shape matters to energy.

What is testedThe models, by family

Size, architecture, and the three features that move energy the most: whether the model has a thinking mode and what its template does when nothing is set, whether it is a mixture of experts (only a fraction of the weights run per token), and whether it carries a vision tower (extra weights that these text-only benchmarks never exercise). "Built for" is the publisher's own positioning; "what we found" is from this corpus.

Quantization variants (AWQ, GPTQ-Int4/Int8, Q4_K_M, MXFP4, and PrismML's PQ2_0 / PTQ1_0 / Q1_0 below two bits) are the same weights re-packed at fewer bits; each is its own row in the index because two packings of "the same" model differ measurably. Ollama tags are Ollama's own GGUF conversions of the named model.

1 · RosterEvery model at its reference configuration, per node

Tier A, all four tasks pooled at the reference item counts (100 / 100 / 50 / 50), Wilson 95% intervals, one configuration per (model, quantization, engine, node, thinking mode), so a model that ran thinking on and off appears twice. The shaded band is every configuration whose interval overlaps the leader's: one statistical group, spread more than an order of magnitude on energy. This is the only figure that pools tasks.

2 · Model sizeWalking up a family's size ladder

Arrows step from the smallest to the largest build within one family on one node, tier A only, and within one thinking mode, so a Qwen3.5 ladder with thinking off is a different line from the same ladder with thinking on. Where an arrow goes up without going right, size bought tokens, not answers.

3 · ThinkingThinking off → on, ten models, three nodes

Every pair in the corpus where only the thinking setting changed: the same model, task, node, cap, graph mode and items. Arrow colour is the verdict on the Wilson intervals: green where thinking raised accuracy, red where it lowered it, grey where the intervals overlap. Pooled over the tier-A pairs, thinking costs 9× and helps in 6 of 48.

3b · Reasoning effortgpt-oss at low → medium → high

The only graded reasoning dial in the roster, run on all four tasks at tier A with gpqa at the full 198 items. Every arrow is short: accuracy intervals overlap on every task, token counts sit within 4% of each other, and energy moves by less than the run-to-run floor. The dial never reached the model. The low and high reasoning traces are byte-identical on 142 of 198 gpqa items, while each differs from the medium row on 185: vLLM 0.25.1 takes harmony reasoning effort from the top-level reasoning_effort request field, and the harness sent it inside chat_template_kwargs. All three rungs ran at medium, so what the arrows show is boot-to-boot drift on an MoE, not a dose-response. Kept here because it is the kind of mistake anyone measuring this dial can make.

3c · Item countgpqa at 50 → 198 items

The accuracy interval halves in width from 50 to 198 items, and the point estimates regress toward each other: Qwen3.5-4B fell from 0.74 to 0.68, gpt-oss rose from 0.59 to 0.61–0.65. They still overlap. What 198 items did resolve is the bottom of the frontier: the no-thinking coder's 0.42 is now clearly below both, at 886 J per correct answer against 14–64 kJ.

4 · Power limitEvery power ladder, every node

Nine ladders across three cards and three models. The knee sits near 225 W on both 3090s. Accuracy is flat on every ladder, so the plane view would be nine vertical lines; these keep the cap on the x-axis so the knee is visible.

5 · Clock lockEvery clock ladder

Six ladders: three tasks on node A, two on node B, one on node C. Node A's gsm8k minimum is at 1350 MHz. The second panel is the finding that a "smoothness" grade, as first defined, rewards a slower card: power CV rises with clock even with no cap in play.

6 · NodeThe same cell on two or three cards

139 cells ran the same configuration on more than one node. Accuracy lands in the same column; energy does not. Node A at 390 W sits above node B at 350 W on almost every cell, and node C's 2080 is cheaper still on the small models it can hold.

7 · CUDA graphsEager → graphs on

26 pairs where only the graph flag changed. On decode-heavy tasks the arrow points down (graphs save energy); on the two-token multiple-choice tasks it can point up. Accuracy never moves.

8 · EngineThe same weights on different engines

34 cells from both engine waves. Within a cell the weights are byte-identical (vLLM ↔ SGLang share AWQ; llama.cpp ↔ Ollama share Q4_K_M; gpt-oss is MXFP4 on both). Accuracy ties everywhere; the vertical spread is the engine.

9 · RepeatsThe same cell run again

22 clusters of three or more runs. Dense models land on the same point every boot; mixtures-of-experts scatter 2–4% in energy; and the earliest repeat studies, run before prefix caching was disabled, walk downward as the cache warms.

10 · Output capRaising the cap, up to uncapped

123 sequences: the deliberate cap ladders plus every capped/uncapped pair in the reference wave. On gsm8k the arrows stop moving after 800 tokens. On math500 they keep climbing. Uncapped points at the top are the runaway items measuring the context window.

11 · Quantizationfp16 → Int8 → Int4 → AWQ

Every precision sequence on a shared configuration, one panel per model and node, the benchmarks as the coloured series. The full four-step ladder exists on Qwen2.5-7B on node B; two-step fp16 → AWQ pairs exist for Qwen2.5-7B and Qwen3.5-9B on node A. Arrows go down, rarely left. Each line states its tier; only sequences with the same tier are comparable to each other.

CoverageWhere each model has been swept

Distinct levels per axis for every build with runs. ⊘ marks a does-not-fit verdict somewhere in the fleet.