diff --git a/himbaechel/uarch/xilinx/pack.cc b/himbaechel/uarch/xilinx/pack.cc
index df802fd8..2b54bbcd 100644
--- a/himbaechel/uarch/xilinx/pack.cc
+++ b/himbaechel/uarch/xilinx/pack.cc
@@ -808,6 +808,42 @@ void XC7Packer::pack_bram()
             }
         }
     }
+
+    // Cluster cascaded RAMB36 pairs: CASCADEOUT->CASCADEIN are dedicated
+    // wires that only reach the vertically adjacent RAMB36 site, so the
+    // pair must be placed together (cf. carry chains). BRAM tiles are 5
+    // grid rows apart within a column.
+    int cascade_pairs = 0;
+    for (auto &cell : ctx->cells) {
+        CellInfo *lower = cell.second.get();
+        if (lower->type != id_RAMB36E1_RAMB36E1)
+            continue;
+        for (auto pn : {ctx->id("CASCADEOUTA"), ctx->id("CASCADEOUTB")}) {
+            NetInfo *cn = lower->getPort(pn);
+            if (cn == nullptr)
+                continue;
+            for (auto &usr : cn->users) {
+                CellInfo *upper = usr.cell;
+                if (upper == lower || upper->type != id_RAMB36E1_RAMB36E1)
+                    continue;
+                if (upper->cluster != ClusterId())
+                    continue; // already clustered (e.g. via the A cascade)
+                if (lower->cluster == ClusterId())
+                    lower->cluster = lower->name;
+                else if (lower->cluster != lower->name)
+                    continue; // lower is a child of another cluster; unsupported
+                upper->cluster = lower->name;
+                lower->constr_children.push_back(upper);
+                upper->constr_x = 0;
+                upper->constr_y = 5;
+                upper->constr_z = 0;
+                upper->constr_abs_z = false;
+                ++cascade_pairs;
+            }
+        }
+    }
+    if (cascade_pairs > 0)
+        log_info("    Clustered %d cascaded BRAM pairs\n", cascade_pairs);
 }
 
 void XilinxPacker::pack_inverters()
diff --git a/himbaechel/uarch/xilinx/xilinx.cc b/himbaechel/uarch/xilinx/xilinx.cc
index 6e9fdb77..0f8f4772 100644
--- a/himbaechel/uarch/xilinx/xilinx.cc
+++ b/himbaechel/uarch/xilinx/xilinx.cc
@@ -101,6 +101,49 @@ void XilinxImpl::init(Context *ctx)
         auto extra_data = tile_extra_data(i);
         tile_status.at(i).site_variant.resize(extra_data->sites.ssize());
     }
+
+    // Build the next-RAMB36-up-the-column map for cascade placement
+    std::map<int, std::map<int, BelId>> bram36_by_col; // x -> y -> bel
+    for (BelId bel : ctx->getBels()) {
+        if (ctx->getBelType(bel) != id_RAMB36E1_RAMB36E1)
+            continue;
+        Loc l = ctx->getBelLocation(bel);
+        bram36_by_col[l.x][l.y] = bel;
+    }
+    for (auto &col : bram36_by_col) {
+        BelId prev;
+        int prev_y = 0;
+        for (auto &entry : col.second) {
+            if (prev != BelId() && (entry.first - prev_y) <= 6)
+                next_bram36_up[entry.second] = prev; // grid Y increases downwards: 'up' = smaller y
+            prev = entry.second;
+            prev_y = entry.first;
+        }
+    }
+}
+
+bool XilinxImpl::getClusterPlacement(ClusterId cluster, BelId root_bel,
+                                     std::vector<std::pair<CellInfo *, BelId>> &placement) const
+{
+    CellInfo *root_cell = ctx->getClusterRootCell(cluster);
+    if (root_cell->type == id_RAMB36E1_RAMB36E1) {
+        if (ctx->getBelType(root_bel) != id_RAMB36E1_RAMB36E1)
+            return false;
+        placement.clear();
+        placement.emplace_back(root_cell, root_bel);
+        BelId cursor = root_bel;
+        for (auto child : root_cell->constr_children) {
+            auto fnd = next_bram36_up.find(cursor);
+            if (fnd == next_bram36_up.end())
+                return false;
+            cursor = fnd->second;
+            if (ctx->getBelType(cursor) != id_RAMB36E1_RAMB36E1)
+                return false;
+            placement.emplace_back(child, cursor);
+        }
+        return true;
+    }
+    return HimbaechelAPI::getClusterPlacement(cluster, root_bel, placement);
 }
 
 SiteIndex XilinxImpl::get_bel_site(BelId bel) const
diff --git a/himbaechel/uarch/xilinx/xilinx.h b/himbaechel/uarch/xilinx/xilinx.h
index 973ee221..140e2880 100644
--- a/himbaechel/uarch/xilinx/xilinx.h
+++ b/himbaechel/uarch/xilinx/xilinx.h
@@ -123,6 +123,14 @@ struct XilinxImpl : HimbaechelAPI
     bool isBelLocationValid(BelId bel, bool explain_invalid = false) const override;
     bool xc7_logic_tile_valid(IdString tileType, const LogicTileStatus &lts) const;
 
+    // BRAM cascade support: CASCADEOUT->CASCADEIN only reaches the next
+    // RAMB36 site up the column; bel z indices are not uniform across BRAM
+    // tiles, so relative (x,y,z) constraints cannot express this. Precompute
+    // the next-RAMB36-in-column map and use it in getClusterPlacement.
+    dict<BelId, BelId> next_bram36_up;
+    bool getClusterPlacement(ClusterId cluster, BelId root_bel,
+                             std::vector<std::pair<CellInfo *, BelId>> &placement) const override;
+
     // Pips
     bool is_pip_unavail(PipId pip) const;
     bool checkPipAvail(PipId pip) const override { return !is_pip_unavail(pip); }