diff --git a/docs/diagrams/allreduce_latency_plots/overview_broken.png b/docs/diagrams/allreduce_latency_plots/overview_broken.png
new file mode 100644
index 0000000..3d56281
Binary files /dev/null and b/docs/diagrams/allreduce_latency_plots/overview_broken.png differ
diff --git a/docs/diagrams/allreduce_latency_plots/overview_log.png b/docs/diagrams/allreduce_latency_plots/overview_log.png
new file mode 100644
index 0000000..8c8c2b9
Binary files /dev/null and b/docs/diagrams/allreduce_latency_plots/overview_log.png differ
diff --git a/docs/diagrams/cube_mesh_view.png b/docs/diagrams/cube_mesh_view.png
new file mode 100644
index 0000000..2736c29
Binary files /dev/null and b/docs/diagrams/cube_mesh_view.png differ
diff --git a/docs/diagrams/kernbench2_overview.pptx b/docs/diagrams/kernbench2_overview.pptx
new file mode 100644
index 0000000..0941ab1
Binary files /dev/null and b/docs/diagrams/kernbench2_overview.pptx differ
diff --git a/scripts/build_overview_slides.py b/scripts/build_overview_slides.py
new file mode 100644
index 0000000..1219762
--- /dev/null
+++ b/scripts/build_overview_slides.py
@@ -0,0 +1,171 @@
+"""Generate a 5-slide PPTX summarizing the kernbench2 model.
+
+Slides (in order):
+  1. Overall architecture — how PEs are connected (cube_mesh_view)
+  2. Model correctness — DMA vs P2P latency (pe2pe overview)
+  3. PE-to-PE IPCQ communication (ipcq_two_pe_dma)
+  4. 6-device allreduce — model vs theoretical vs ext-sim (overview_broken)
+  5. IPCQ buffer-kind sweep — TCM vs SRAM vs HBM (buffer_kind_sweep)
+
+This is a derived-artifact generator — no production code touched.
+"""
+from __future__ import annotations
+
+from pathlib import Path
+
+from PIL import Image
+from pptx import Presentation
+from pptx.dml.color import RGBColor
+from pptx.enum.shapes import MSO_SHAPE
+from pptx.util import Emu, Inches, Pt
+
+ROOT = Path(__file__).resolve().parent.parent
+DIAG = ROOT / "docs" / "diagrams"
+OUT = DIAG / "kernbench2_overview.pptx"
+
+# 16:9 widescreen — 13.333 × 7.5 in
+SLIDE_W_IN = 13.333
+SLIDE_H_IN = 7.5
+
+SLIDES = [
+    {
+        "title": "1. CUBE Architecture: NOC Router Mesh + PE Connectivity",
+        "image": DIAG / "cube_mesh_view.png",
+        "bullets": [
+            "Each CUBE holds an 8-PE NOC mesh wired through routers (R0..R7)",
+            "Every PE has IO_CPU, M_CPU, PE_CPU + IPCQ engine + DMA engine",
+            "Inter-cube traffic exits via UCIe/UAL ports; SIPs stitch into ring/torus/mesh",
+            "Foundation for every latency, IPCQ, and allreduce experiment that follows",
+        ],
+    },
+    {
+        "title": "2. Model Correctness: DMA vs P2P Latency Sweep",
+        "image": DIAG / "pe2pe_latency_plots" / "overview.png",
+        "bullets": [
+            "Sweeps payload size across PE-to-PE paths and compares to DMA",
+            "Confirms the simulator reproduces the expected DMA/P2P crossover",
+            "Acts as the per-hop ground truth that feeds collective-level models",
+        ],
+    },
+    {
+        "title": "3. IPCQ: How Two PEs Communicate (DMA + Slot Memory)",
+        "image": DIAG / "ipcq_diagram_plots" / "ipcq_two_pe_dma.png",
+        "bullets": [
+            "Sender pushes payload through PE_DMA → fabric → receiver IPCQ slot",
+            "Slot memory (TCM/SRAM/HBM) charges a write on arrival, a read on consume",
+            "Credit return rides the fabric path back (16 B packet, no slot-IO)",
+            "This is the building block the multi-device allreduce composes",
+        ],
+    },
+    {
+        "title": "4. 6-Device Allreduce: Model vs Theoretical vs External Simulator",
+        "image": DIAG / "allreduce_latency_plots" / "overview_broken.png",
+        "bullets": [
+            "Three SIP topologies (ring / torus / mesh) swept 16 B → 96 KB per PE",
+            "Dashed red curve: hand-derived theoretical model for torus_2d (6 SIPs)",
+            "Top panel (broken y-axis): single-device reduce on ext-sim ≈ 366 µs",
+            "Our 6-device collective lands at ~17–22 µs — ~17× faster than ext-sim baseline",
+        ],
+    },
+    {
+        "title": "5. IPCQ Slot Memory: TCM vs SRAM vs HBM",
+        "image": DIAG / "allreduce_latency_plots" / "buffer_kind_sweep.png",
+        "bullets": [
+            "Same allreduce with slot memory swapped: TCM (per-PE local) / SRAM / HBM (cube-shared, behind router link)",
+            "Cost = NoC drain + slot-IO + PE↔bank hop; only TCM skips the bank hop",
+            "Topology link BWs set the order: SRAM bank link 128 GB/s is the narrowest in the system, HBM 256 GB/s",
+            "At 64 KB / PE: TCM 12.0 µs < HBM 21.4 µs < SRAM 24.3 µs — SRAM is slowest because of its narrow bank link",
+        ],
+    },
+]
+
+
+def _add_title(slide, text):
+    left = Inches(0.4)
+    top = Inches(0.25)
+    width = Inches(SLIDE_W_IN - 0.8)
+    height = Inches(0.7)
+    box = slide.shapes.add_textbox(left, top, width, height)
+    tf = box.text_frame
+    tf.margin_left = tf.margin_right = Emu(0)
+    tf.margin_top = tf.margin_bottom = Emu(0)
+    p = tf.paragraphs[0]
+    run = p.add_run()
+    run.text = text
+    run.font.size = Pt(26)
+    run.font.bold = True
+    run.font.color.rgb = RGBColor(0x10, 0x2A, 0x55)
+    return box
+
+
+def _add_image_centered(slide, img_path, *, left_in, top_in, max_w_in, max_h_in):
+    with Image.open(img_path) as im:
+        iw, ih = im.size
+    max_w_emu = Inches(max_w_in)
+    max_h_emu = Inches(max_h_in)
+    scale = min(max_w_emu / iw, max_h_emu / ih)
+    w = int(iw * scale)
+    h = int(ih * scale)
+    left = Inches(left_in) + (max_w_emu - w) // 2
+    top = Inches(top_in) + (max_h_emu - h) // 2
+    slide.shapes.add_picture(str(img_path), left, top, width=w, height=h)
+
+
+def _add_bullets(slide, bullets, *, left_in, top_in, width_in, height_in):
+    box = slide.shapes.add_textbox(
+        Inches(left_in), Inches(top_in), Inches(width_in), Inches(height_in),
+    )
+    tf = box.text_frame
+    tf.word_wrap = True
+    for i, line in enumerate(bullets):
+        p = tf.paragraphs[0] if i == 0 else tf.add_paragraph()
+        p.level = 0
+        run = p.add_run()
+        run.text = "• " + line
+        run.font.size = Pt(15)
+        run.font.color.rgb = RGBColor(0x22, 0x22, 0x22)
+        p.space_after = Pt(6)
+
+
+def _add_footer(slide, idx, total):
+    box = slide.shapes.add_textbox(
+        Inches(SLIDE_W_IN - 1.2), Inches(SLIDE_H_IN - 0.45),
+        Inches(1.0), Inches(0.3),
+    )
+    p = box.text_frame.paragraphs[0]
+    run = p.add_run()
+    run.text = f"{idx} / {total}"
+    run.font.size = Pt(10)
+    run.font.color.rgb = RGBColor(0x88, 0x88, 0x88)
+
+
+def build():
+    prs = Presentation()
+    prs.slide_width = Inches(SLIDE_W_IN)
+    prs.slide_height = Inches(SLIDE_H_IN)
+    blank = prs.slide_layouts[6]
+
+    for i, cfg in enumerate(SLIDES, start=1):
+        slide = prs.slides.add_slide(blank)
+        _add_title(slide, cfg["title"])
+
+        # Layout: image on the left (8.4 in wide), bullets on the right (4.4 in).
+        _add_image_centered(
+            slide, cfg["image"],
+            left_in=0.3, top_in=1.05,
+            max_w_in=8.3, max_h_in=5.9,
+        )
+        _add_bullets(
+            slide, cfg["bullets"],
+            left_in=8.8, top_in=1.2,
+            width_in=4.3, height_in=5.7,
+        )
+        _add_footer(slide, i, len(SLIDES))
+
+    OUT.parent.mkdir(parents=True, exist_ok=True)
+    prs.save(OUT)
+    print(f"wrote {OUT}")
+
+
+if __name__ == "__main__":
+    build()
diff --git a/scripts/emit_overview_with_external_ref.py b/scripts/emit_overview_with_external_ref.py
new file mode 100644
index 0000000..a8a2863
--- /dev/null
+++ b/scripts/emit_overview_with_external_ref.py
@@ -0,0 +1,192 @@
+"""One-shot: render overview.png with an external 366 µs reference, in two
+variants — log scale and broken y-axis. Reads docs/diagrams/allreduce_latency_plots/summary.csv
+and writes overview_log.png and overview_broken.png alongside it.
+
+This is a derived-artifact generator (per CLAUDE.md): plotting only, no production
+or test logic touched.
+"""
+from __future__ import annotations
+
+import csv
+from pathlib import Path
+
+import matplotlib.pyplot as plt
+import matplotlib.ticker as mticker
+
+ROOT = Path(__file__).resolve().parent.parent
+PLOT_DIR = ROOT / "docs" / "diagrams" / "allreduce_latency_plots"
+CSV_PATH = PLOT_DIR / "summary.csv"
+
+EXT_LABEL = "ext-sim single-device reduce: 366 µs"
+EXT_LATENCY_NS = 366_000.0
+
+COLORS = {
+    "ring_1d": "tab:blue",
+    "torus_2d": "tab:orange",
+    "mesh_2d_no_wrap": "tab:green",
+}
+
+# Hand-derived theoretical model for torus_2d (6 SIPs). Mirrors
+# _aggregate_sweep_plots in tests/test_allreduce_multidevice.py.
+NOC_PACKET_BYTES = 128
+PES_PER_CUBE = 8
+T_STARTUP_NS = 1346.0
+TAU_NS = (8741.0 - 1346.0) / (6144 - 1)
+
+
+def _theoretical_torus_2d_ns(bytes_per_pe: int) -> float:
+    bytes_per_cube = int(bytes_per_pe) * PES_PER_CUBE
+    n_packets = max(1, -(-bytes_per_cube // NOC_PACKET_BYTES))
+    return T_STARTUP_NS + (n_packets - 1) * TAU_NS
+
+
+def _plot_theoretical(ax, records):
+    torus_rs = sorted(
+        [r for r in records if r["sip_topology"] == "torus_2d"],
+        key=lambda r: r["bytes_per_pe"],
+    )
+    if not torus_rs:
+        return
+    ax.plot(
+        [r["bytes_per_pe"] for r in torus_rs],
+        [_theoretical_torus_2d_ns(r["bytes_per_pe"]) for r in torus_rs],
+        color="tab:red", linestyle="--", linewidth=1.6, marker="x",
+        label="theoretical torus_2d (6 SIPs)",
+    )
+
+
+def _bytes_fmt(x, _pos):
+    if x >= 1024 * 1024:
+        return f"{x / (1024 * 1024):.0f}M"
+    if x >= 1024:
+        return f"{x / 1024:.0f}K"
+    return f"{int(x)}"
+
+
+def _load_records():
+    rows = []
+    with open(CSV_PATH, newline="") as f:
+        r = csv.DictReader(f)
+        for row in r:
+            rows.append({
+                "sip_topology": row["sip_topology"],
+                "bytes_per_pe": int(row["bytes_per_pe"]),
+                "latency_ns": float(row["latency_ns"]),
+            })
+    return rows
+
+
+def _ext_x(records):
+    """Anchor the external reference at the largest payload (96 KB / PE)."""
+    return max(r["bytes_per_pe"] for r in records)
+
+
+def _plot_curves(ax, records, topologies):
+    for topo in topologies:
+        rs = sorted([r for r in records if r["sip_topology"] == topo],
+                    key=lambda r: r["bytes_per_pe"])
+        if not rs:
+            continue
+        ax.plot(
+            [r["bytes_per_pe"] for r in rs],
+            [r["latency_ns"] for r in rs],
+            marker="o",
+            label=f"{topo}",
+            color=COLORS.get(topo),
+        )
+
+
+def emit_log(records):
+    topologies = sorted({r["sip_topology"] for r in records})
+    fig, ax = plt.subplots(figsize=(9, 6))
+    _plot_curves(ax, records, topologies)
+    _plot_theoretical(ax, records)
+    ax.scatter(
+        [_ext_x(records)], [EXT_LATENCY_NS],
+        marker="*", s=220, color="tab:red", zorder=5,
+        label=EXT_LABEL,
+    )
+    ax.set_xscale("log", base=2)
+    ax.set_yscale("log")
+    ax.set_xlabel("Bytes per PE (log scale)")
+    ax.set_ylabel("Time (ns) — log scale")
+    ax.set_title("Multi-device allreduce latency vs external single-device reference")
+    ax.grid(True, which="both", alpha=0.3)
+    ax.xaxis.set_major_formatter(mticker.FuncFormatter(_bytes_fmt))
+    ax.legend(loc="upper left")
+    fig.tight_layout()
+    out = PLOT_DIR / "overview_log.png"
+    fig.savefig(out, dpi=120)
+    plt.close(fig)
+    print(f"wrote {out}")
+
+
+def emit_broken(records):
+    topologies = sorted({r["sip_topology"] for r in records})
+    max_local = max(r["latency_ns"] for r in records)
+
+    fig, (ax_top, ax_bot) = plt.subplots(
+        2, 1, sharex=True,
+        gridspec_kw={"height_ratios": [1, 4], "hspace": 0.05},
+        figsize=(9, 6.5),
+    )
+
+    # Bottom panel: today's three curves + theoretical, linear y.
+    _plot_curves(ax_bot, records, topologies)
+    _plot_theoretical(ax_bot, records)
+    ax_bot.set_ylim(0, max_local * 1.10)
+
+    # Top panel: only the external reference marker, linear y around 366 µs.
+    ax_top.scatter(
+        [_ext_x(records)], [EXT_LATENCY_NS],
+        marker="*", s=240, color="tab:red", zorder=5,
+        label=EXT_LABEL,
+    )
+    ax_top.set_ylim(EXT_LATENCY_NS * 0.93, EXT_LATENCY_NS * 1.05)
+
+    # Hide the spine between the two panels and draw diagonal "break" ticks.
+    ax_top.spines["bottom"].set_visible(False)
+    ax_bot.spines["top"].set_visible(False)
+    ax_top.tick_params(labeltop=False, bottom=False)
+    ax_bot.xaxis.tick_bottom()
+
+    d = 0.012  # diagonal-tick size, in axis-fraction
+    kw = dict(transform=ax_top.transAxes, color="k", clip_on=False, lw=1)
+    ax_top.plot((-d, +d), (-d, +d), **kw)
+    ax_top.plot((1 - d, 1 + d), (-d, +d), **kw)
+    kw.update(transform=ax_bot.transAxes)
+    ax_bot.plot((-d, +d), (1 - d * 4, 1 + d * 4), **kw)
+    ax_bot.plot((1 - d, 1 + d), (1 - d * 4, 1 + d * 4), **kw)
+
+    ax_bot.set_xscale("log", base=2)
+    ax_bot.set_xlabel("Bytes per PE (log scale)")
+    ax_bot.set_ylabel("Time (ns)")
+    ax_top.set_ylabel("Time (ns)")
+    ax_bot.grid(True, alpha=0.3)
+    ax_top.grid(True, alpha=0.3)
+    ax_bot.xaxis.set_major_formatter(mticker.FuncFormatter(_bytes_fmt))
+
+    # One legend covering both axes.
+    handles_bot, labels_bot = ax_bot.get_legend_handles_labels()
+    handles_top, labels_top = ax_top.get_legend_handles_labels()
+    ax_bot.legend(handles_bot + handles_top, labels_bot + labels_top,
+                  loc="upper left")
+
+    fig.suptitle("Multi-device allreduce latency vs external single-device reference (broken y-axis)")
+    fig.tight_layout()
+    out = PLOT_DIR / "overview_broken.png"
+    fig.savefig(out, dpi=120)
+    plt.close(fig)
+    print(f"wrote {out}")
+
+
+def main():
+    records = _load_records()
+    if not records:
+        raise SystemExit(f"no rows in {CSV_PATH}")
+    emit_log(records)
+    emit_broken(records)
+
+
+if __name__ == "__main__":
+    main()