Expanding GPU Capabilities on Notebooks and Mini PCs Without PCIe Slots via M.2 NVMe Slots

Introduction

Modern AI, machine learning, and scientific computing demand high-performance GPUs, but many notebooks and mini PCs lack traditional PCIe slots for expansion. While it is possible to add one or even multiple GPUs to these systems using M.2 NVMe-to-PCIe adapters and PCIe expansion boards, this approach is mostly practical for retrofitting under specific constraints. For general-purpose or high-performance use, a standard desktop system with a full-size case and proper PCIe slots remains the more robust and scalable solution. This article details how to achieve this, potential challenges, and how to optimize power and connectivity for multi-GPU setups.

Why Expand GPUs on a Notebook or Mini PC?

Notebooks and mini PCs are compact and energy-efficient, but their limited expandability due to the lack of PCIe slots often restricts their computational potential. However, adding high-end GPUs with large dedicated memory in a modular fashion can significantly enhance their capabilities by enabling:

• Performing scientific computations and modeling—for example, in quantum physics.
• Running Large Language Models (LLMs) and Stable Diffusion locally.
• Accelerating custom image processing tasks.
• Enhancing machine learning workloads.
• Performing large-scale finite element simulations (FEM).

The Hardware Setup

To connect GPUs to a system without standard PCIe slots, we utilize:

• M.2 NVMe to PCIe x4 Adapter – Converts an NVMe slot into a PCIe x4 interface.
• PCIe x1 to 4x PCIe x1 Expansion Board – Allows connecting up to four GPUs via risers.
• External ATX Power Supply – Ensures sufficient power delivery for GPUs (e.g., ATX PSU with breakout board for GPUs like A100s).

The actual hardware components in our build are (note: all links provided are Amazon affiliate links, this pages author profits from qualified purchases):

• Mini PC (Ιntel Alder Lake-N97) with an M.2 NVMe slot.
• M.2 NVMe to PCIe x4 riser.
• PCIe x1 to 4x PCIe x1 expander board.
• GPUs: Starting with a budget-friendly RTX 3090 Ti 24GB (or in case of very low budget an RTX 3060 12GB), with potential for up to 4x A100s featuring 80GB VRAM per card—totaling up to 320GB VRAM, suitable for most modern applications.
• External 1200W PSU to power multiple GPUs (A single A100 requires approximately 300W, bringing the total power requirement for four GPUs to about 1200W).
• A power switch adapter that powers up the external power supply whenever the 12V rail of the main computer is energized. This synchronizes the two power supplies.
• A HDMI dummy plug to circumvent mining lock at graphics cards to which we do not attach displays.
• Some thermal paste when mounting heat spreaders or fans

The following photograph shows a hacked Zombie setup based on an Intel Alder Lake-N97, an additional 400W ATX power supply (with bridged power on pin) as well as an NVidia RTX 3060 with 12 GB VRAM:

.

Of course one should think about a proper enclosure when not mounting inside a server rack or similar - and ensure proper airflow.

Step-by-Step Installation

Step 1: Identify M.2 Slot

Most modern mini PCs have at least one NVMe M.2 slot (key M). To ensure compatibility, check if your system’s BIOS supports PCIe bifurcation or external GPU detection. However, many budget-friendly systems allow utilizing the PCIe lanes of the NVMe slot even if they do not explicitly support PCIe bifurcation.

Before purchasing hardware, ensure that your system has an available M.2 slot. Discovering that your only M.2 slot is occupied by a storage device after acquiring all components can be frustrating.

Step 2: Connect M.2 NVMe to PCIe Adapter

Insert the M.2 NVMe to PCIe adapter into the mini PC’s M.2 slot. These adapters generally offer either a PCIe x4 connector or a hardware x16 connector that operates with up to four lanes. Some models even utilize two M.2 slots to provide an x8 connection. To extend the connection, attach a PCIe riser cable, which not only facilitates the use of multiple GPUs but also allows you to test if the system detects the additional PCIe host controller before making significant hardware investments. Simply connecting the PCIe riser will help verify whether the system properly recognizes the host controller.

A setup utilizing the PCIe expansion board is often used in the cryptocurrency mining world. In that domain, it’s common to chain multiple stages of PCIe expanders to connect dozens of GPUs per host.

Step 3: Install the GPU(s)

Insert the GPU into one of the available PCIe slots. Some GPUs may refuse to function correctly if all PCIe lanes are not populated or if no display is attached. This restriction, commonly referred to as a mining lock, is intended to prevent cryptocurrency mining but unfortunately also limits the usability of such devices for scientific computing and media processing. There are dummy plugs that you can utilize to emulate an attached display though.

Proper mounting and adequate spacing are essential to avoid overheating. Additionally, ensure a mechanically stable setup, as troubleshooting hardware contact issues alongside software-related problems can be highly frustrating. 

Step 4: Powering the GPUs

First a word of caution: Warning: Always take precautions when working with high-wattage PSUs. Improper grounding or overloads can damage hardware or pose electrical risks. Also only manipulate high wattage devices or line powered devices when you are qualified to do so. They impose a significant risk that could lead to lethal injuries. Also note that working without proper grounding may destroy components by electro static discharge.

In most cases, an external ATX PSU is required to ensure stable power delivery to the GPUs. Mini PC systems and notebooks generally come with power supplies limited to under 100W, which is insufficient for powering a dedicated GPU. Since these GPUs are powered independently from the main system while maintaining a common ground through PCIe, the PSU must be manually enabled before the PC starts. One approach to achieve this is by bridging the ATX power supply’s enable pin to ground. Alternatively, a more convenient solution is using an adapter to synchronize power activation with the main computer’s 12V rail, provided the mini PC or computer includes a 12V rail connection.

It is also important to note that many graphics adapters cannot be reset without a full power cycle, especially during system reboots. This is a well-known limitation in virtualization systems but can also present challenges in this setup. If automated power control is not available, manually cycling the GPU’s power may be required to restore functionality.

To prevent connectivity issues, ensure that all PCIe power connections are secure, including those for the expansion board and GPUs.

Step 5: Configure Software

• Install the appropriate GPU drivers, whether NVIDIA or AMD, based on your hardware. If performing GPU computations, consider installing the CUDA toolkit for NVIDIA GPUs.
• If required, enable external GPU support in the BIOS to ensure proper functionality.
• Use PCI device listing tools to verify the presence of the host bridge and CPUs. Additionally, utilities such as nvidia-smi can help confirm GPU detection and monitor performance.

Challenges and Workarounds

• BIOS and PCIe Detection Issues: Some systems may not recognize external GPUs via M.2 slots. Possible solutions include:
  • Updating the BIOS (though some BIOS versions intentionally blacklist unsupported PCIe devices in NVMe slots - then updating will not work most likely).
  • Enabling Above 4G Decoding in BIOS. This setting allows the system to address PCIe devices requiring large memory allocations beyond the 4GB limit of traditional 32-bit addressing. It is particularly important when using multiple GPUs, as each GPU demands dedicated memory-mapped I/O space. Without this enabled, some GPUs may not be recognized or function properly.
  • Check if the slot actually supports NVMe - or only SATA. This approach does not work on SATA only ports.
• Limited Bandwidth on PCIe x1: PCIe x1 bandwidth is restrictive for certain applications, particularly graphics-intensive workloads.
  • Running LLMs and machine learning tasks is still feasible since these workloads primarily rely on internal GPU memory once the model is loaded. But note that loading of models takes about 16 times longer than on a proper port. If you use multiple GPUs in parallel on the same x1 port this multiplies with the number of GPUs - so using 4 GPUs in parallel means 64 times slower loading of the models.
  • Cryptocurrency mining is another use case where PCIe bandwidth is less critical.
  • Direct GPU-to-GPU memory transfers can be leveraged to minimize data movement through the system’s PCIe bus. This can be done utilizing proper memory copy routines in your OpenCL or CUDA code instead of always transferring synchronously to host memory.
• Power Constraints:
  • A single A100 GPU consumes between 300-400W, meaning a four-GPU setup could require anywhere from 1.2kW to 1.6kW of power. If your power supply cannot handle this load, you may experience power failures and system instability.
  • Proper power distribution is crucial to prevent overloading and to ensure stable operation. The power supply type plays a role in this; some require an even distribution of power across all rails, while single-rail supplies do not face this issue.
  • For multi-rail power supplies, ensure that you do not exceed the maximum power draw on any individual rail to avoid shutdowns or component failures.
• Heat Management:
  • The substantial power consumption of high-end GPUs results in considerable heat output, making effective cooling solutions essential. Underestimating the heat generated can lead to system instability and reduced hardware longevity, so proper thermal calculations should not be overlooked.
  • Implementing additional case fans or liquid cooling systems can help maintain optimal thermal performance and prevent overheating.

Conclusion

We tested our dual RTX 3060 (24GB useable with applications that are capable of splitting their workload on two GPUs) configuration with:

• Stable Diffusion for image generation.
• GPT-based LLM inference (running models like DeepSeek-R1:70b or mistral-large:123b).
• Custom FEM simulations and math algorithms (like QR decomposition, GMRES, SVD decomposition, Eigenvector calculations, etc.)

Results demonstrated significant performance improvements over CPU-only processing, despite limited PCIe bandwidth.

Expanding GPU capabilities on a notebook or mini PC is not only feasible but also a cost-effective solution. By utilizing the right adapters, power supplies, and configurations, even an affordable mini PC can accommodate up to four high-end GPUs for AI and scientific computing. With a slightly larger investment, the expansion potential can scale almost indefinitely per host.

For deeper insights into the mathematics behind computations that you can speed up, check out our related articles on machine learning, finite element simulations, and numerical methods.

This article is tagged:

• DIY
• Hardware
• Machine learning
• Electronics
• System administration