Read on the NPU. Where to write depends on the reply.

• Prefill: the NPU is fastest on every model, 10.5 to 15.9 times the CPU and 5.1 to 8.2 times the GPU.
• Decode: the CPU is fastest on eight models and level with the NPU on Qwen3 4B (15.27 against 15.35 tok/s). On the transformers its lead over the NPU shrinks with size, from 1.27 to 1.37 times at 0.6B and 1B to 1.08 times on Llama 3.2 3B. The two LFM2.5 hybrids keep a wider lead: 1.48 times at 1.2B, 1.34 times at 2.6B.
• Long context: past 2,048 tokens of context the NPU decoded faster than the CPU on five transformers (Qwen3 4B: 12.6 against 10.6 tok/s). Four or five turns each, late in replay and so also the hottest: a tendency, not a rule.
• Fastest pairing at 64 tokens: NPU prefill with CPU decode on all nine models, at 41% to 52% of CPU-only time. On Qwen3 4B it ties with NPU only: 0.2% ahead in the first pass, 0.3% behind in the second. On Llama 3.2 3B, NPU only is 7.4% slower. At the conversations' own reply lengths this changes; see below.
• GPU: never the fastest at either phase. On the two LFM2.5 hybrids it decodes faster than the NPU (46.7 against 42.7, 21.6 against 20.7 tok/s); on the seven transformers it decodes slowest.
• Handoff: copying the conversation state from one processor to another took a median 0.6 to 11 ms per turn.
• Second pass: five models ran again with every cell in another order. Each kept its slowest pairing, and the four smaller ones their fastest; cells moved a median 1.2% to 2.9%, at most 6.4%. Qwen3 4B's top two swapped, NPU only taking 149.6 s and NPU + CPU 150.1 s: over both passes, a tie.

64-token replies, first pass. Ratios and decode rates are the mean of the three cells that use each processor. A pairing is named prefill processor first, so NPU + CPU reads the prompt on the NPU and writes the reply on the CPU.

• Why: every turn above writes 64 tokens. The datasets' replies to those turns are about six times longer: 401 tokens median by Qwen3's tokenizer, up to 1,141 by Gemma's. At that length writing the reply is 84% to 99% of each pairing's time.
• Run: each turn decodes its own reply's length, every step timed, in up to four pairings per model chosen from the 64-token ranking, on three units of the phone. The cooling gate before each cell waits, up to 120 s here (300 s for Qwen3 0.6B in r9), without cooling anything, and after the first cell the phone rarely got back under 55 °C in that time: 32 of the 40 cells started with the CPU's clock ceilings lowered, most of them to 1.90 and 2.23 GHz with the GPU capped.
• Throttled, the decoder decides, and it depends on the model. Where both NPU-prefill cells started throttled, NPU + CPU took 10% to 20% longer than NPU only on Qwen3 1.7B (in two orders), its Q8_0 and Llama 3.2 3B even without the CPU's stalls. The two tied on the two smallest transformers, and on Qwen3 4B, over its 33 turns, NPU only led only through the stalls. On the hybrid LFM2.5 1.2B the throttled CPU slowed but stayed ahead, and NPU + CPU won every turn; on LFM2.5 2.6B the two were level in one run, and in another NPU + CPU lost with the CPU's stalls and won without them. Gemma 3 1B's CPU decoded faster than its NPU even capped (51.5 against 46.5 tok/s on matched capped turns, one schedule).
• CPU stalls: on the four models with the largest weights, Qwen3 4B, Llama 3.2 3B, Qwen3 1.7B Q8_0 and, in one of its three runs, LFM2.5 2.6B, cells decoding on the CPU had single steps of 0.5 to 15.9 s, 54 to 291 s per cell. Each of those cells reached about 107 °C at some turn's end (the stalled turns themselves ended at 94.6 to 106.6 °C); nothing recorded shows the cause. No NPU-only cell stalled, nor any cell of the five smaller models.
• When NPU + CPU started after a passed gate (Qwen3 0.6B, Gemma 3 1B), it was fastest, at 75% and 76% of CPU-only time, as at 64 tokens. Qwen3 0.6B's same pairing, started throttled on another unit, took 87%.
• The NPU keeps its pace: its 64-token decode rate predicted its time at reply length within about 5% on seven models (Qwen3 4B over the 20 turns all its cells completed), and where it ran more than once its totals came within 3% of each other. Cells decoding on the CPU took 3% to 61% longer than predicted, and their totals differed by up to 25% between runs.

r8 to r11 are the report's labels for the runs, each one order on one unit. Seconds are summed over the 33 turns; where a pairing completed them in more than one run, they are the mean of the runs named under it, with how far apart those were. LFM2.5 2.6B's NPU only in r9, cut at 29 turns, is left out of its mean, and r10 has no NPU + CPU: the phone stopped answering before it finished. The last two columns give how much longer NPU + CPU took than NPU only (minus: less): with 64-token replies in the first pass, and at reply length within each run, on the turns both cells completed. * NPU + CPU started after a passed cooling gate and NPU only after a timed-out one. Matching gate results are not matched thermal histories either: a gate that timed out covers a range of temperatures and caps. “Without stalls” deletes the stalled steps' time: it measures their share, not what the cell would have taken without whatever caused them. GPU + CPU ran on Qwen3 0.6B and 1.7B only, at 94% to 100% of CPU-only time.