Synthetic prefill · 15 complete sweeps · QWEN-PLACEMENT-CROSSOVER-001

Qwen placement crossover

Session-level repeatability for synthetic prompt-prefill time across three distinct RTX 3090 configurations. These measurements are not answer-quality scores.

The narrow intervals are the result: repeated sessions on this host were highly repeatable for these synthetic prefill cells. They do not rank model quality or establish a causal PCIe effect.

Five-sweep results

Fifteen complete fresh-daemon sweeps (five per lane), 24 measured requests per sweep. Within each cell, two request repetitions were reduced to a sweep median; intervals are 95% Student-t intervals across five sweep medians. Tight intervals indicate within-host repeatability for synthetic prefill, not answer quality or cross-day performance. GPU0/GPU1 does not isolate slot causality; Qwen3.6 dual is a separate model/card-count cohort.

Mean prefill seconds; shared y-axis within each workload row; whiskers show 95% t interval (n=5 sweeps) qwen38-gpu0 · low workload* 0 21 42 63 83 Prefill duration (s) 32k 65k 98k 131k 196k 262k qwen38-gpu1 · low workload* 0 21 42 63 83 Prefill duration (s) 32k 65k 98k 131k 196k 262k qwen36-dual · low workload* 0 21 42 63 83 Prefill duration (s) 32k 65k 98k 131k 196k 262k qwen38-gpu0 · high workload* 0 546 1093 1639 2186 Prefill duration (s) 32k 65k 98k 131k 196k 262k qwen38-gpu1 · high workload* 0 546 1093 1639 2186 Prefill duration (s) 32k 65k 98k 131k 196k 262k qwen36-dual · high workload* 0 546 1093 1639 2186 Prefill duration (s) 32k 65k 98k 131k 196k 262k
* Different y-axis scales by row: low workload uses 0–83 s ticks; high workload uses 0–2,186 s ticks. Each row shares its linear scale across all three configurations, so compare vertical heights only within a row. The broad upward trend is descriptive, not one shared scaling law: Qwen3.6 dual low workload stays nearly flat while its high workload rises.

qwen38-gpu0

CONFIGURED CONTEXTWORKLOADMEAN PREFILL (s)95% t INTERVAL (s)ACTUAL PROMPT TOKENS
32,768low8.124[8.050, 8.198]10,013
32,768high21.415[21.320, 21.509]24,587
65,536low8.113[8.085, 8.140]10,013
65,536high47.805[47.593, 48.018]49,165
98,304low13.030[12.995, 13.065]10,013
98,304high118.231[117.916, 118.546]73,742
131,072low16.085[15.736, 16.435]10,014
131,072high202.496[199.202, 205.790]98,321
196,608low22.331[22.203, 22.459]10,014
196,608high452.963[451.170, 454.756]147,470
262,144low26.049[25.766, 26.332]10,014
262,144high746.614[742.113, 751.116]196,620

qwen38-gpu1

CONFIGURED CONTEXTWORKLOADMEAN PREFILL (s)95% t INTERVAL (s)ACTUAL PROMPT TOKENS
32,768low10.120[9.122, 11.119]10,013
32,768high24.741[24.291, 25.191]24,587
65,536low9.975[9.823, 10.128]10,013
65,536high55.558[54.567, 56.548]49,165
98,304low25.275[25.216, 25.334]10,013
98,304high210.131[209.750, 210.513]73,742
131,072low37.649[37.624, 37.675]10,014
131,072high439.977[439.585, 440.369]98,321
196,608low59.565[59.515, 59.614]10,014
196,608high1093.422[1093.078, 1093.766]147,470
262,144low74.433[74.365, 74.500]10,014
262,144high1951.087[1950.473, 1951.700]196,620

qwen36-dual

CONFIGURED CONTEXTWORKLOADMEAN PREFILL (s)95% t INTERVAL (s)ACTUAL PROMPT TOKENS
32,768low2.753[2.610, 2.897]10,013
32,768high7.255[7.116, 7.393]24,587
65,536low2.402[2.398, 2.406]10,013
65,536high14.569[14.534, 14.604]49,165
98,304low2.406[2.402, 2.410]10,013
98,304high24.670[24.609, 24.731]73,742
131,072low2.423[2.413, 2.433]10,014
131,072high36.732[36.652, 36.812]98,321
196,608low2.434[2.419, 2.449]10,014
196,608high66.709[66.517, 66.901]147,470
262,144low2.717[2.704, 2.731]10,014
262,144high114.372[114.050, 114.695]196,620

Downloads and provenance

Rendered from the Project Phoenix local benchmark-dashboard card. The three lanes remain separate: Qwen3.8 GPU0 and GPU1 are pinned-placement configurations, while Qwen3.6 dual is a different model/card-count cohort. Incomplete sweeps are excluded; raw result packets and full telemetry are not included.