diff --git a/hp/hp1/configuration.nix b/hp/hp1/configuration.nix
index 45f4592..4563f2b 100644
--- a/hp/hp1/configuration.nix
+++ b/hp/hp1/configuration.nix
@@ -52,6 +52,8 @@
       # design 32: urp acceptor endpoint + PTP grandmaster (lab time sync)
       ./urp.nix
       ./ptp.nix
+      # design 37: persistent jumbo MTU (9700) on the 25 GbE RoCE links
+      ./roce-mtu.nix
     ];
 
   # Bootloader.
diff --git a/hp/hp1/roce-mtu.nix b/hp/hp1/roce-mtu.nix
new file mode 100644
index 0000000..9573b89
--- /dev/null
+++ b/hp/hp1/roce-mtu.nix
@@ -0,0 +1,26 @@
+# hp/hp1/roce-mtu.nix
+#
+# Persistent jumbo MTU on the 25 GbE RoCEv2 links (uds-rdma-proxy design 37
+# throughput work). The ConnectX-4 Lx driver reports maxmtu 9978; NVIDIA's
+# practical max for this NIC (MT27710) is 9700. A large L2 MTU lifts the RoCE
+# path MTU (ibv active_mtu) from 1024 to its 4096 ceiling — the real RDMA win —
+# and also raises the single-stream TCP baseline.
+#
+# Set via a systemd-udevd .link file so it applies at device appearance,
+# independent of the imperative xdp2.testbed NIC wiring (whose `jumbo` option
+# would only hardcode 9000). A 10-* prefix sorts ahead of the empty
+# 40-<iface>.link, so this is the matching .link for these ports.
+#
+# Verify after `nixos-rebuild switch` + reboot:
+#   ip -d link show enp1s0f0np0        # mtu 9700
+#   ibv_devinfo | grep active_mtu      # 4096 (was 1024)
+#   ping -M do -s 9672 10.10.2.3       # jumbo path passes end-to-end
+# If the path cannot pass 9700 (switch/peer limit), drop MTUBytes to "9000".
+{ ... }:
+
+{
+  systemd.network.links."10-roce-jumbo" = {
+    matchConfig.OriginalName = "enp1s0f0np0 enp1s0f1np1";
+    linkConfig.MTUBytes = "9700";
+  };
+}
diff --git a/hp/hp3/configuration.nix b/hp/hp3/configuration.nix
index b21aac6..ff1d7e9 100644
--- a/hp/hp3/configuration.nix
+++ b/hp/hp3/configuration.nix
@@ -41,6 +41,8 @@
       # design 32: urp initiator endpoint + PTP slave (lab time sync)
       ./urp.nix
       ./ptp.nix
+      # design 37: persistent jumbo MTU (9700) on the 25 GbE RoCE links
+      ./roce-mtu.nix
     ];
 
   # Bootloader.
diff --git a/hp/hp3/roce-mtu.nix b/hp/hp3/roce-mtu.nix
new file mode 100644
index 0000000..83c36d0
--- /dev/null
+++ b/hp/hp3/roce-mtu.nix
@@ -0,0 +1,26 @@
+# hp/hp3/roce-mtu.nix
+#
+# Persistent jumbo MTU on the 25 GbE RoCEv2 links (uds-rdma-proxy design 37
+# throughput work). The ConnectX-4 Lx driver reports maxmtu 9978; NVIDIA's
+# practical max for this NIC (MT27710) is 9700. A large L2 MTU lifts the RoCE
+# path MTU (ibv active_mtu) from 1024 to its 4096 ceiling — the real RDMA win —
+# and also raises the single-stream TCP baseline.
+#
+# Set via a systemd-udevd .link file so it applies at device appearance,
+# independent of the imperative xdp2.testbed NIC wiring (whose `jumbo` option
+# would only hardcode 9000). A 10-* prefix sorts ahead of the empty
+# 40-<iface>.link, so this is the matching .link for these ports.
+#
+# Verify after `nixos-rebuild switch` + reboot:
+#   ip -d link show enp1s0f0np0        # mtu 9700
+#   ibv_devinfo | grep active_mtu      # 4096 (was 1024)
+#   ping -M do -s 9672 10.10.2.1       # jumbo path passes end-to-end
+# If the path cannot pass 9700 (switch/peer limit), drop MTUBytes to "9000".
+{ ... }:
+
+{
+  systemd.network.links."10-roce-jumbo" = {
+    matchConfig.OriginalName = "enp1s0f0np0 enp1s0f1np1";
+    linkConfig.MTUBytes = "9700";
+  };
+}