diff --git a/.github/workflows/ipa-ramdisk-build.yml b/.github/workflows/ipa-ramdisk-build.yml
index 9b353a1..f42437f 100644
--- a/.github/workflows/ipa-ramdisk-build.yml
+++ b/.github/workflows/ipa-ramdisk-build.yml
@@ -34,6 +34,11 @@ jobs:
             release: trixie
             image_prefix: ipa-debian-trixie
             extra_elements: "fluidstack-ironwood"
+          - name: gpu-noble
+            distro: ubuntu
+            release: noble
+            image_prefix: ipa-ubuntu-noble-gpu
+            extra_elements: "fluidstack-gpu"
     steps:
       - name: Checkout
         uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
diff --git a/dib/fluidstack-gpu/README.rst b/dib/fluidstack-gpu/README.rst
new file mode 100644
index 0000000..b2bf22e
--- /dev/null
+++ b/dib/fluidstack-gpu/README.rst
@@ -0,0 +1,70 @@
+fluidstack-gpu
+==============
+
+DIB element that bundles NVIDIA drivers, CUDA runtime, NCCL, and pre-compiled
+``nccl-tests`` binaries into an IPA ramdisk for GPU qualification workloads.
+
+Intended base distribution: **Ubuntu Noble 24.04** (``ubuntu`` + ``--release noble``).
+
+What is installed
+-----------------
+
+Runtime (present in final ramdisk):
+
+* ``cuda-drivers-<VER>`` — NVIDIA metapackage (kernel modules, ``nvidia-smi``, compute libs, all from NVIDIA repo at matching versions)
+* ``libnccl2`` — NCCL multi-GPU communication library
+* ``cuda-cudart-<VER>`` — CUDA runtime library
+* ``/usr/local/bin/nccl-tests/`` — pre-compiled nccl-tests binaries
+
+Build-time only (stripped in ``post-install.d``):
+
+* ``cuda-toolkit-<VER>`` — full toolkit used to compile nccl-tests (nvcc, headers)
+* ``libnccl-dev`` — NCCL headers
+
+Configuration
+-------------
+
+All versions are overridable via environment variables at build time:
+
+.. list-table::
+   :header-rows: 1
+
+   * - Variable
+     - Default
+     - Description
+   * - ``DIB_NVIDIA_DRIVER_VERSION``
+     - ``570``
+     - NVIDIA driver series (e.g. ``535``, ``550``, ``570``)
+   * - ``DIB_CUDA_VERSION``
+     - ``12-8``
+     - CUDA version used for apt package names (dash-separated, e.g. ``12-6``)
+   * - ``DIB_NCCL_TESTS_REF``
+     - ``master``
+     - Git branch/tag of `NVIDIA/nccl-tests <https://github.com/NVIDIA/nccl-tests>`_ to compile
+
+Example build command
+---------------------
+
+.. code-block:: bash
+
+    export DIB_DHCP_TIMEOUT=60
+    export DIB_IPA_ENABLE_RESCUE=false
+
+    ironic-python-agent-builder \
+        --lzma \
+        --output ipa-ubuntu-noble-gpu-stable-2026.1-fs \
+        --release noble \
+        --branch stable/2026.1 \
+        --verbose \
+        --element fluidstack-gpu \
+        ubuntu
+
+Verification
+------------
+
+Once booted:
+
+.. code-block:: bash
+
+    nvidia-smi
+    /usr/local/bin/nccl-tests/all_reduce_perf -b 1G -e 1G -f 2 -g <N>
diff --git a/dib/fluidstack-gpu/element-deps b/dib/fluidstack-gpu/element-deps
new file mode 100644
index 0000000..0a65984
--- /dev/null
+++ b/dib/fluidstack-gpu/element-deps
@@ -0,0 +1 @@
+install-static
diff --git a/dib/fluidstack-gpu/environment.d/10-versions.bash b/dib/fluidstack-gpu/environment.d/10-versions.bash
new file mode 100644
index 0000000..f73abec
--- /dev/null
+++ b/dib/fluidstack-gpu/environment.d/10-versions.bash
@@ -0,0 +1,19 @@
+# GPU stack version pins — edit this file to update component versions.
+# All DIB_* variables can be overridden at build time via environment.
+
+# NVIDIA driver — must match an available nvidia-driver-pinning-* package
+export DIB_NVIDIA_DRIVER_VERSION=${DIB_NVIDIA_DRIVER_VERSION:-590.48.01}
+
+# CUDA toolkit (dash-separated apt package suffix, e.g. 13-1 → cuda-toolkit-13-1)
+export DIB_CUDA_VERSION=${DIB_CUDA_VERSION:-13-1}
+
+# NCCL tests git ref compiled into the image
+export DIB_NCCL_TESTS_REF=${DIB_NCCL_TESTS_REF:-master}
+
+# GDRCopy (package revision format: <semver>-<rev>)
+export DIB_GDRCOPY_VERSION=${DIB_GDRCOPY_VERSION:-2.5.1-1}
+
+# Feature flags
+export DIB_ENABLE_GDRCOPY=${DIB_ENABLE_GDRCOPY:-true}
+export DIB_ENABLE_DCGM=${DIB_ENABLE_DCGM:-true}
+export DIB_ENABLE_PEERMEM=${DIB_ENABLE_PEERMEM:-false}
diff --git a/dib/fluidstack-gpu/environment.d/20-cuda-paths.bash b/dib/fluidstack-gpu/environment.d/20-cuda-paths.bash
new file mode 100644
index 0000000..dc12f8d
--- /dev/null
+++ b/dib/fluidstack-gpu/environment.d/20-cuda-paths.bash
@@ -0,0 +1,3 @@
+export CUDA_HOME=/usr/local/cuda
+export PATH=$CUDA_HOME/bin:$PATH
+export LD_LIBRARY_PATH="${CUDA_HOME}/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}"
diff --git a/dib/fluidstack-gpu/install.d/40-install-gpu-packages b/dib/fluidstack-gpu/install.d/40-install-gpu-packages
new file mode 100755
index 0000000..48273c7
--- /dev/null
+++ b/dib/fluidstack-gpu/install.d/40-install-gpu-packages
@@ -0,0 +1,77 @@
+#!/bin/bash
+set -eux
+
+NVIDIA_VER="${DIB_NVIDIA_DRIVER_VERSION:-590.48.01}"
+CUDA_VER="${DIB_CUDA_VERSION:-13-1}"
+CUDA_MAJOR=$(echo "${CUDA_VER}" | cut -d'-' -f1)
+
+# --- Kernel headers first ------------------------------------------------
+# Install headers for the Ubuntu kernel in the image BEFORE any DKMS package.
+# When headers are present at DKMS package install time, the modules are built
+# immediately against the right kernel. Installing headers afterwards triggers
+# the postinst DKMS hook on already-registered modules, which fails in a chroot.
+KERNEL_VERSION=$(ls /lib/modules | head -1)
+DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
+    "linux-headers-${KERNEL_VERSION}"
+
+# DKMS 3.4.x from the NVIDIA CUDA repo passes ARCH=amd64 (Debian package arch)
+# to the kernel build system. The kernel has no arch/amd64/ — it only knows
+# x86_64 → x86. Create a symlink so the kernel Makefile finds its arch files.
+ln -sf x86 "/usr/src/linux-headers-${KERNEL_VERSION}/arch/amd64"
+
+# --- Drivers and CUDA toolkit -------------------------------------------
+# Pin the driver version first, then install nvidia-open. DKMS builds the
+# kernel module for the headers we just installed.
+DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
+    "nvidia-driver-pinning-${NVIDIA_VER}" \
+    nvidia-open \
+    "cuda-toolkit-${CUDA_VER}" \
+    libnccl2 \
+    libnccl-dev \
+    "cuda-cudart-${CUDA_VER}"
+
+# --- DCGM ----------------------------------------------------------------
+if [ "${DIB_ENABLE_DCGM:-true}" = "true" ]; then
+    if [ "${CUDA_MAJOR}" -ge 13 ]; then
+        DCGM_PKG="datacenter-gpu-manager-4-cuda${CUDA_MAJOR}"
+    else
+        DCGM_PKG="datacenter-gpu-manager"
+    fi
+    DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
+        "${DCGM_PKG}"
+    systemctl enable nvidia-dcgm
+fi
+
+# --- nvidia-fabricmanager: explicitly disabled ---------------------------
+# Fabric Manager requires NVSwitch hardware. Single-GPU and PCIe-attached
+# nodes don't have NVSwitch, so it will always fail to start.
+if systemctl list-unit-files nvidia-fabricmanager.service &>/dev/null; then
+    systemctl disable nvidia-fabricmanager
+fi
+
+# --- GDRCopy -------------------------------------------------------------
+if [ "${DIB_ENABLE_GDRCOPY:-true}" = "true" ]; then
+    GDRCOPY_VER="${DIB_GDRCOPY_VERSION:-2.5.1-1}"
+    GDRCOPY_SEMVER="${GDRCOPY_VER%-*}"          # 2.5.1-1 → 2.5.1
+    GDRCOPY_CUDA="${CUDA_MAJOR}.0"
+    GDRCOPY_BASE="https://developer.download.nvidia.com/compute/redist/gdrcopy/CUDA%20${GDRCOPY_CUDA}/ubuntu24_04/x64"
+
+    mkdir -p /tmp/gdrcopy
+
+    # Install all GDRCopy packages including gdrdrv-dkms. The postinst builds
+    # and installs gdrdrv.ko via DKMS (works because of the arch/amd64 symlink).
+    # post-install.d/96-reset-gdrdrv-dkms uninstalls it via DKMS before the
+    # DIB 97-dkms element runs, so 97-dkms can install it cleanly without the
+    # version mismatch (MODULE_VERSION "2.5" vs DKMS package version "2.5.1").
+    for deb in \
+        "gdrdrv-dkms_${GDRCOPY_VER}_amd64.Ubuntu24_04.deb" \
+        "libgdrapi_${GDRCOPY_VER}_amd64.Ubuntu24_04.deb" \
+        "gdrcopy-tests_${GDRCOPY_VER}_amd64.Ubuntu24_04+cuda${GDRCOPY_CUDA}.deb" \
+        "gdrcopy_${GDRCOPY_VER}_amd64.Ubuntu24_04.deb"
+    do
+        curl -fsSL "${GDRCOPY_BASE}/${deb}" -o "/tmp/gdrcopy/${deb}"
+        DEBIAN_FRONTEND=noninteractive dpkg -i "/tmp/gdrcopy/${deb}"
+    done
+
+    rm -rf /tmp/gdrcopy
+fi
diff --git a/dib/fluidstack-gpu/install.d/45-nvidia-services-setup b/dib/fluidstack-gpu/install.d/45-nvidia-services-setup
new file mode 100755
index 0000000..8fe5d39
--- /dev/null
+++ b/dib/fluidstack-gpu/install.d/45-nvidia-services-setup
@@ -0,0 +1,19 @@
+#!/bin/bash
+set -eux
+
+# --- nvidia-persistenced -------------------------------------------------
+# Create a dedicated system user (mirrors the Ansible role) and enable the
+# service with an override that runs with --persistence-mode.
+useradd --system --shell /usr/sbin/nologin \
+    --comment "NVIDIA Persistence Daemon user" \
+    nvidia-persistenced 2>/dev/null || true
+
+mkdir -p /etc/systemd/system/nvidia-persistenced.service.d
+# override.conf is laid down by install-static from static/
+systemctl enable nvidia-persistenced
+systemctl enable nvidia-gpu-reset
+
+# --- peermem -------------------------------------------------------------
+if [ "${DIB_ENABLE_PEERMEM:-false}" = "true" ]; then
+    echo "nvidia-peermem" > /etc/modules-load.d/nvidia-peermem.conf
+fi
diff --git a/dib/fluidstack-gpu/install.d/50-nccl-tests-build b/dib/fluidstack-gpu/install.d/50-nccl-tests-build
new file mode 100755
index 0000000..2d14093
--- /dev/null
+++ b/dib/fluidstack-gpu/install.d/50-nccl-tests-build
@@ -0,0 +1,18 @@
+#!/bin/bash
+set -eux
+
+# Compile nccl-tests against the CUDA toolkit and NCCL installed in the image.
+# Only the resulting binaries are kept; sources and build artifacts are removed.
+NCCL_TESTS_REF="${DIB_NCCL_TESTS_REF:-master}"
+
+git clone --depth=1 --branch "${NCCL_TESTS_REF}" \
+    https://github.com/NVIDIA/nccl-tests /tmp/nccl-tests
+
+make -C /tmp/nccl-tests \
+    CUDA_HOME=/usr/local/cuda \
+    MPI=0 \
+    -j"$(nproc)"
+
+install -d /usr/local/bin/nccl-tests
+install -m 755 /tmp/nccl-tests/build/*_perf /usr/local/bin/nccl-tests/
+rm -rf /tmp/nccl-tests
diff --git a/dib/fluidstack-gpu/post-install.d/96-reset-gdrdrv-dkms b/dib/fluidstack-gpu/post-install.d/96-reset-gdrdrv-dkms
new file mode 100755
index 0000000..b03a132
--- /dev/null
+++ b/dib/fluidstack-gpu/post-install.d/96-reset-gdrdrv-dkms
@@ -0,0 +1,11 @@
+#!/bin/bash
+set -eux
+
+# The gdrdrv-dkms postinst builds gdrdrv.ko via DKMS and places the .ko.zst
+# on disk, but does NOT record it in the DKMS state database ("dkms uninstall"
+# reports "not installed"). The DIB 97-dkms element then calls "dkms install",
+# which finds the orphaned .ko.zst at version "2.5" (module's internal version,
+# not the package version "2.5.1") and refuses to overwrite without --force.
+# Remove the orphaned file so 97-dkms can install cleanly.
+KERNEL_VERSION=$(ls /lib/modules | head -1)
+rm -f "/lib/modules/${KERNEL_VERSION}/updates/dkms/gdrdrv.ko.zst"
diff --git a/dib/fluidstack-gpu/post-install.d/99-cleanup-gpu-build b/dib/fluidstack-gpu/post-install.d/99-cleanup-gpu-build
new file mode 100755
index 0000000..6f077dc
--- /dev/null
+++ b/dib/fluidstack-gpu/post-install.d/99-cleanup-gpu-build
@@ -0,0 +1,20 @@
+#!/bin/bash
+set -eux
+
+# Remove the CUDA development toolkit that was needed only to compile nccl-tests.
+# Keep: cuda-cudart-*, libnccl2, nvidia-headless-no-dkms-*, nvidia-utils-*,
+#       libnvidia-compute-* (these are all needed at runtime).
+DEBIAN_FRONTEND=noninteractive apt-get remove -y --auto-remove \
+    libnccl-dev \
+    linux-headers-\* \
+    cuda-toolkit-\* \
+    cuda-compiler-\* \
+    cuda-cudart-dev-\* \
+    cuda-nvcc-\* \
+    cuda-libraries-dev-\* \
+    gdrcopy-tests \
+    || true
+
+apt-get autoremove -y
+apt-get clean
+rm -rf /var/lib/apt/lists/*
diff --git a/dib/fluidstack-gpu/pre-install.d/10-add-nvidia-repo b/dib/fluidstack-gpu/pre-install.d/10-add-nvidia-repo
new file mode 100755
index 0000000..cf0a88e
--- /dev/null
+++ b/dib/fluidstack-gpu/pre-install.d/10-add-nvidia-repo
@@ -0,0 +1,12 @@
+#!/bin/bash
+set -eux
+
+# Add the NVIDIA CUDA apt repository using the official cuda-keyring package.
+# This repo provides nvidia-headless-no-dkms-*, libnccl2, cuda-cudart-*, etc.
+KEYRING_DEB=cuda-keyring_1.1-1_all.deb
+curl -fsSL \
+    "https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/${KEYRING_DEB}" \
+    -o /tmp/${KEYRING_DEB}
+dpkg -i /tmp/${KEYRING_DEB}
+rm /tmp/${KEYRING_DEB}
+apt-get update -q
diff --git a/dib/fluidstack-gpu/static/etc/ld.so.conf.d/60-cuda.conf b/dib/fluidstack-gpu/static/etc/ld.so.conf.d/60-cuda.conf
new file mode 100644
index 0000000..751cd0e
--- /dev/null
+++ b/dib/fluidstack-gpu/static/etc/ld.so.conf.d/60-cuda.conf
@@ -0,0 +1,2 @@
+/usr/local/cuda/lib64
+/usr/local/cuda/extras/CUPTI/lib64
diff --git a/dib/fluidstack-gpu/static/etc/modprobe.d/nvidia.conf b/dib/fluidstack-gpu/static/etc/modprobe.d/nvidia.conf
new file mode 100644
index 0000000..80da664
--- /dev/null
+++ b/dib/fluidstack-gpu/static/etc/modprobe.d/nvidia.conf
@@ -0,0 +1 @@
+options nvidia NVreg_EnableStreamMemOPs=1 NVreg_RegistryDwords="PeerMappingOverride=1;"
diff --git a/dib/fluidstack-gpu/static/etc/modules-load.d/nvidia.conf b/dib/fluidstack-gpu/static/etc/modules-load.d/nvidia.conf
new file mode 100644
index 0000000..8bcae1d
--- /dev/null
+++ b/dib/fluidstack-gpu/static/etc/modules-load.d/nvidia.conf
@@ -0,0 +1,3 @@
+nvidia
+nvidia_uvm
+nvidia_drm
diff --git a/dib/fluidstack-gpu/static/etc/systemd/system/nvidia-gpu-reset.service b/dib/fluidstack-gpu/static/etc/systemd/system/nvidia-gpu-reset.service
new file mode 100644
index 0000000..eadd203
--- /dev/null
+++ b/dib/fluidstack-gpu/static/etc/systemd/system/nvidia-gpu-reset.service
@@ -0,0 +1,14 @@
+[Unit]
+Description=Reset NVIDIA GPUs via PCIe FLR before driver load
+DefaultDependencies=no
+Before=systemd-modules-load.service
+
+[Service]
+Type=oneshot
+ExecStart=/usr/local/bin/nvidia-gpu-reset.sh
+RemainAfterExit=yes
+StandardOutput=journal+console
+StandardError=journal+console
+
+[Install]
+WantedBy=sysinit.target
diff --git a/dib/fluidstack-gpu/static/etc/systemd/system/nvidia-persistenced.service.d/override.conf b/dib/fluidstack-gpu/static/etc/systemd/system/nvidia-persistenced.service.d/override.conf
new file mode 100644
index 0000000..7633473
--- /dev/null
+++ b/dib/fluidstack-gpu/static/etc/systemd/system/nvidia-persistenced.service.d/override.conf
@@ -0,0 +1,6 @@
+[Service]
+ExecStart=
+ExecStart=/usr/bin/nvidia-persistenced --user nvidia-persistenced --persistence-mode --verbose
+
+[Install]
+WantedBy=multi-user.target
diff --git a/dib/fluidstack-gpu/static/usr/local/bin/nvidia-gpu-reset.sh b/dib/fluidstack-gpu/static/usr/local/bin/nvidia-gpu-reset.sh
new file mode 100755
index 0000000..b66a7ed
--- /dev/null
+++ b/dib/fluidstack-gpu/static/usr/local/bin/nvidia-gpu-reset.sh
@@ -0,0 +1,17 @@
+#!/bin/bash
+# Reset all NVIDIA GPUs via PCIe FLR before the nvidia driver loads.
+# This clears any leftover GSP/WPR2 state from a previous driver session
+# (e.g. after kexec) that would cause RmInitAdapter to fail on boot.
+set -eu
+
+for dev in /sys/bus/pci/devices/*/; do
+    vendor=$(cat "${dev}vendor" 2>/dev/null || true)
+    [ "${vendor}" = "0x10de" ] || continue
+
+    reset_method=$(cat "${dev}reset_method" 2>/dev/null || true)
+    if echo "${reset_method}" | grep -q "flr"; then
+        pci_id=$(basename "${dev}")
+        echo "nvidia-gpu-reset: FLR on ${pci_id}"
+        echo 1 > "${dev}reset"
+    fi
+done