20ca297378
His earlier 15-cap sweep (260-400W, 16832066) is superseded by a 38-cap sweep at 10W resolution from 230-600W (16854218) — strictly richer envelope. Two findings worth surfacing: - Firmware boost-clock plateau at SM 2610 MHz: caps 400-600W produce identical 51.96 TPS at 393W actual draw. Auto-plateau detector in power-cap-sweep.sh caught both sub-plateaus cleanly (400-470W and 480-600W). - decode-concurrent N=4 plateau lower than decode-single (46 vs 52 TPS) — on a 4090 + 27B Q3_K_XL the GPU is under-load at c=4 even at full TDP, so single-stream wins. Useful cross-rig signal. HARDWARE.md table + chart caption updated; chart regenerated with both load-mode curves overlaid + plateau zone shaded. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
211 lines
7.7 KiB
Python
211 lines
7.7 KiB
Python
"""Generate a tweet-ready chart from @laurimyllari's RTX 4090 power-cap sweep on llama.cpp.
|
||
|
||
Source data: https://github.com/noonghunna/club-3090/discussions/62#discussioncomment-16854218
|
||
Setup: 1× RTX 4090 air-cooled, llama.cpp default, Qwen3.6-27B-UD-Q3_K_XL GGUF, sweep on club-3090@aa99173.
|
||
|
||
Two load modes are overlaid:
|
||
- decode-single (10s × 2 timed streams per cap)
|
||
- decode-concurrent (concurrency=4, 4s/run × 3 runs × 2 batches per cap)
|
||
|
||
The two curves together show: (a) where the efficiency knee is (250-260W on either mode),
|
||
(b) the firmware boost-clock plateau locks at SM 2610 MHz / 393W draw from cap=400W onwards,
|
||
and (c) on this 27B model concurrency=4 sits below single-stream — i.e. the 4090 is in
|
||
under-load territory at c=4 on Q3_K_XL.
|
||
"""
|
||
import matplotlib.pyplot as plt
|
||
|
||
# decode-single: 38-cap sweep, 230-600W. Columns: cap, narr_tps, code_tps, actual_W, narr_tps_per_W
|
||
# Source: disc #62 16854218 (decode-single 14:42:23Z run)
|
||
decode_single = [
|
||
(230, 41.47, 41.47, 229.72, 0.181),
|
||
(240, 43.47, 43.77, 240.21, 0.181),
|
||
(250, 46.37, 46.36, 249.84, 0.186),
|
||
(260, 48.26, 48.16, 259.90, 0.186), # sweet spot (tied with 250W)
|
||
(270, 48.96, 48.96, 269.92, 0.181),
|
||
(280, 49.36, 49.36, 279.72, 0.176),
|
||
(290, 49.76, 49.86, 289.63, 0.172),
|
||
(300, 50.16, 50.16, 299.69, 0.167),
|
||
(310, 50.46, 50.56, 309.58, 0.163),
|
||
(320, 50.76, 50.76, 319.34, 0.159),
|
||
(330, 51.06, 51.06, 329.55, 0.155),
|
||
(340, 51.26, 51.26, 339.47, 0.151),
|
||
(350, 51.36, 51.36, 349.67, 0.147),
|
||
(360, 51.56, 51.56, 359.38, 0.143),
|
||
(370, 51.66, 51.66, 369.43, 0.140),
|
||
(380, 51.86, 51.86, 378.50, 0.137),
|
||
(390, 51.95, 51.96, 387.96, 0.134),
|
||
(400, 51.96, 51.96, 392.17, 0.132), # firmware plateau begins
|
||
(410, 51.96, 51.96, 391.69, 0.133),
|
||
(420, 51.96, 51.96, 392.14, 0.133),
|
||
(430, 51.96, 51.96, 392.13, 0.133),
|
||
(440, 51.95, 51.96, 393.26, 0.132),
|
||
(450, 51.96, 51.96, 393.25, 0.132),
|
||
(460, 51.86, 51.86, 393.58, 0.132),
|
||
(470, 51.96, 51.96, 393.19, 0.132),
|
||
(480, 51.85, 51.96, 394.23, 0.132),
|
||
(490, 51.86, 51.96, 394.13, 0.132),
|
||
(500, 51.86, 51.96, 393.15, 0.132),
|
||
(510, 51.86, 51.86, 394.03, 0.132),
|
||
(520, 51.86, 51.96, 393.96, 0.132),
|
||
(530, 51.85, 51.96, 394.25, 0.132),
|
||
(540, 51.86, 51.96, 394.76, 0.131),
|
||
(550, 51.86, 51.96, 394.88, 0.131),
|
||
(560, 51.86, 51.96, 394.58, 0.131),
|
||
(570, 51.85, 51.96, 394.33, 0.131),
|
||
(580, 51.86, 51.96, 394.32, 0.132),
|
||
(590, 51.86, 51.96, 394.52, 0.131),
|
||
(600, 51.85, 51.96, 394.87, 0.131),
|
||
]
|
||
|
||
# decode-concurrent N=4: 38-cap sweep, 230-600W.
|
||
# Source: disc #62 16854218 (decode-concurrent 14:58:09Z run, concurrency=4 auto-calibrated)
|
||
decode_concurrent = [
|
||
(230, 36.96, 36.46, 229.58, 0.161),
|
||
(240, 38.92, 38.69, 239.79, 0.162),
|
||
(250, 41.14, 40.66, 249.65, 0.165), # under-load knee (tied with 260W)
|
||
(260, 42.13, 42.35, 259.50, 0.162),
|
||
(270, 43.11, 43.34, 269.13, 0.160),
|
||
(280, 43.60, 43.61, 278.92, 0.156),
|
||
(290, 44.10, 43.85, 289.16, 0.153),
|
||
(300, 44.58, 44.35, 299.48, 0.149),
|
||
(310, 44.59, 44.83, 309.32, 0.144),
|
||
(320, 44.83, 45.08, 319.19, 0.140),
|
||
(330, 45.08, 45.07, 328.98, 0.137),
|
||
(340, 45.34, 45.33, 339.07, 0.134),
|
||
(350, 45.59, 45.58, 349.02, 0.131),
|
||
(360, 45.82, 45.57, 358.57, 0.128),
|
||
(370, 45.82, 45.81, 368.40, 0.124),
|
||
(380, 45.82, 46.06, 378.29, 0.121),
|
||
(390, 46.29, 46.07, 387.93, 0.119),
|
||
(400, 46.08, 46.06, 393.64, 0.117), # plateau begins
|
||
(410, 46.31, 46.06, 394.30, 0.117),
|
||
(420, 46.06, 46.07, 394.04, 0.117),
|
||
(430, 46.07, 46.07, 395.27, 0.117),
|
||
(440, 46.08, 46.07, 394.50, 0.117),
|
||
(450, 46.08, 46.07, 394.64, 0.117),
|
||
(460, 46.07, 46.04, 394.67, 0.117),
|
||
(470, 46.04, 46.08, 394.48, 0.117),
|
||
(480, 46.07, 46.08, 394.49, 0.117),
|
||
(490, 46.08, 46.07, 394.24, 0.117),
|
||
(500, 46.32, 46.07, 394.13, 0.118),
|
||
(510, 46.07, 46.08, 394.65, 0.117),
|
||
(520, 46.08, 46.07, 395.27, 0.117),
|
||
(530, 46.08, 46.06, 394.79, 0.117),
|
||
(540, 46.06, 46.07, 395.07, 0.117),
|
||
(550, 46.07, 46.07, 395.58, 0.116),
|
||
(560, 46.08, 46.07, 395.36, 0.117),
|
||
(570, 46.07, 46.07, 395.51, 0.116),
|
||
(580, 46.07, 46.07, 394.87, 0.117),
|
||
(590, 46.08, 46.07, 395.44, 0.117),
|
||
(600, 46.07, 46.06, 395.51, 0.116),
|
||
]
|
||
|
||
ds_caps = [d[0] for d in decode_single]
|
||
ds_narr = [d[1] for d in decode_single]
|
||
ds_eff = [d[4] for d in decode_single]
|
||
|
||
dc_caps = [d[0] for d in decode_concurrent]
|
||
dc_narr = [d[1] for d in decode_concurrent]
|
||
dc_eff = [d[4] for d in decode_concurrent]
|
||
|
||
plt.rcParams.update({
|
||
"font.family": "sans-serif",
|
||
"font.size": 12,
|
||
"axes.titlesize": 16,
|
||
"axes.titleweight": "bold",
|
||
"axes.labelsize": 13,
|
||
"figure.facecolor": "white",
|
||
"axes.facecolor": "white",
|
||
})
|
||
|
||
fig, ax1 = plt.subplots(figsize=(12, 6.5), dpi=150)
|
||
|
||
# Firmware boost-clock plateau shading (caps 400-600W)
|
||
ax1.axvspan(395, 605, color="#ffd6a5", alpha=0.35, zorder=0,
|
||
label="Firmware boost-clock plateau (SM 2610 MHz / 393W draw)")
|
||
|
||
# Left axis: TPS (both modes)
|
||
color_ds = "#1f77b4"
|
||
color_dc = "#2ca02c"
|
||
ax1.plot(ds_caps, ds_narr, "o-", color=color_ds, linewidth=2.2, markersize=5,
|
||
label="decode-single (narr TPS)", zorder=3)
|
||
ax1.plot(dc_caps, dc_narr, "s-", color=color_dc, linewidth=2.2, markersize=5,
|
||
label="decode-concurrent N=4 (narr TPS)", zorder=3)
|
||
ax1.set_xlabel("Power cap (W)", fontsize=13)
|
||
ax1.set_ylabel("Narrative wall TPS", fontsize=13)
|
||
ax1.set_xlim(225, 605)
|
||
ax1.set_ylim(35, 54)
|
||
ax1.grid(True, alpha=0.3, zorder=0)
|
||
ax1.tick_params(axis="both", labelsize=11)
|
||
|
||
# Right axis: TPS/W efficiency (both modes)
|
||
ax2 = ax1.twinx()
|
||
color_eff_ds = "#d62728"
|
||
color_eff_dc = "#9467bd"
|
||
ax2.plot(ds_caps, ds_eff, "^--", color=color_eff_ds, linewidth=1.5, markersize=4,
|
||
alpha=0.85, label="decode-single TPS/W", zorder=2)
|
||
ax2.plot(dc_caps, dc_eff, "v--", color=color_eff_dc, linewidth=1.5, markersize=4,
|
||
alpha=0.85, label="decode-concurrent TPS/W", zorder=2)
|
||
ax2.set_ylabel("Efficiency: TPS/W (narrative)", fontsize=13, color="#555")
|
||
ax2.tick_params(axis="y", labelcolor="#555", labelsize=11)
|
||
ax2.set_ylim(0.10, 0.20)
|
||
|
||
# Sweet spot annotation: 260W on decode-single
|
||
ax1.axvline(260, color="goldenrod", linestyle=":", alpha=0.5, linewidth=1.5)
|
||
ax1.annotate(
|
||
"★ 260W\n0.186 TPS/W\n(decode-single knee)\n42% below stock TDP",
|
||
xy=(260, 48.26),
|
||
xytext=(295, 50.5),
|
||
fontsize=10.5,
|
||
fontweight="bold",
|
||
bbox=dict(boxstyle="round,pad=0.4", facecolor="#fff3cd", edgecolor="goldenrod", linewidth=1.2),
|
||
arrowprops=dict(arrowstyle="->", color="goldenrod", lw=1.5),
|
||
zorder=4,
|
||
)
|
||
|
||
# Plateau annotation
|
||
ax1.annotate(
|
||
"Caps 400-600W functionally identical\n(SM 2610 MHz / 393W actual draw)",
|
||
xy=(500, 51.96),
|
||
xytext=(450, 47.5),
|
||
fontsize=9.5,
|
||
ha="center",
|
||
color="#a85a00",
|
||
fontstyle="italic",
|
||
arrowprops=dict(arrowstyle="->", color="#cc8800", lw=1),
|
||
)
|
||
|
||
# Title
|
||
ax1.set_title(
|
||
"RTX 4090 + Qwen3.6-27B + llama.cpp — power-cap efficiency curves",
|
||
pad=14,
|
||
)
|
||
|
||
# Subtitle
|
||
fig.text(
|
||
0.5, 0.92,
|
||
"1× 4090 air-cooled, llama.cpp default, Q3_K_XL GGUF, 38-cap sweep at 10W | data: @laurimyllari (club-3090 disc #62)",
|
||
ha="center", fontsize=10, color="#666",
|
||
style="italic",
|
||
)
|
||
|
||
# Combined legend
|
||
lines1, labels1 = ax1.get_legend_handles_labels()
|
||
lines2, labels2 = ax2.get_legend_handles_labels()
|
||
ax1.legend(lines1 + lines2, labels1 + labels2,
|
||
loc="lower right", fontsize=10, framealpha=0.95,
|
||
edgecolor="#ccc", ncol=1)
|
||
|
||
# Footer
|
||
fig.text(
|
||
0.99, 0.01,
|
||
"github.com/noonghunna/club-3090",
|
||
ha="right", fontsize=9, color="#888", style="italic",
|
||
)
|
||
|
||
plt.tight_layout(rect=(0, 0.02, 1, 0.92))
|
||
|
||
out = "docs/img/power-cap-4090-qwen36.png"
|
||
plt.savefig(out, dpi=150, bbox_inches="tight", facecolor="white")
|
||
print(f"Saved: {out}")
|