External GPU (eGPU) Enclosures & Docks for Running Local AI Models on Laptops

External GPU (eGPU) Enclosures & Docks for Running Local AI Models on Laptops

What an eGPU Enclosure Actually Does

An external GPU enclosure (eGPU) is a small chassis with its own power supply that holds a desktop graphics card and tunnels its PCIe lanes over a Thunderbolt or USB4 cable back to a laptop. The laptop’s Thunderbolt controller negotiates a PCIe link with the enclosure, the operating system enumerates the card as if it were installed internally, and the GPU driver stack loads normally. For local AI work this matters because laptop-class GPUs are capped at 8GB–16GB of VRAM and modest power limits, while a desktop card in an enclosure can carry far more memory and a higher sustained wattage budget for running quantized LLMs.

Docks and enclosures are not interchangeable terms in practice. A “dock” like the MINISFORUM DEG1 or DEG2 usually ships with a GPU already mounted and sealed, aimed at plug-and-play simplicity. An “enclosure” like the Razer Core X V2 or Sonnet Breakaway Box is an empty chassis you populate yourself, trading convenience for the freedom to upgrade the card later.

Thunderbolt Generation Sets the Bandwidth Ceiling

The Thunderbolt or USB4 generation on the laptop’s port is the single biggest limiter on eGPU performance, because it determines how many PCIe lanes’ worth of throughput actually reach the card. Thunderbolt 3 and Thunderbolt 4 both carry a 40 Gbps signaling rate, but PCIe traffic shares that link with display and USB data, so real-world throughput lands closer to a PCIe x4 connection rather than the x16 a desktop slot provides. Thunderbolt 5 raises the ceiling to 80 Gbps bidirectional, with a Bandwidth Boost mode pushing transmit to 120 Gbps and roughly doubling the PCIe throughput available to the enclosure — enough to narrow, though not eliminate, the gap with an internal slot.

That narrowed-but-present bandwidth bottleneck shows up differently depending on the workload. Loading a large quantized model from disk into VRAM is bound by the link’s raw throughput, so a Thunderbolt 5 port measurably shortens load times versus Thunderbolt 3/4. Token generation during inference, once the model is resident in VRAM, depends far more on the GPU’s own memory bandwidth and compute than on the external link, so the practical penalty for running inference over Thunderbolt versus a native PCIe x16 slot is smaller than most buyers expect.

Enclosure Classes and What Separates Them

Four rough classes show up in this category:

Which class fits depends on whether the GPU is a known fixed asset (buy a dock) or something you expect to swap as VRAM needs grow (buy an empty enclosure).

GPU Compatibility, Power, and Physical Clearance

An enclosure’s spec sheet needs to be checked against the card, not just the port. Card length and slot width (two-slot, three-slot, or four-slot blower and triple-fan designs) must fit the enclosure’s internal clearance, and taller or longer cards common on workstation GPUs can simply not close the lid on a compact box. Power matters just as much: the enclosure’s built-in PSU, or the user-supplied ATX/SFX PSU in an open chassis, needs headroom above the card’s rated TDP — enough margin to cover transient spikes from a 12VHPWR or 8-pin PCIe power connector without tripping overcurrent protection.

Laptop-side power delivery is a separate constraint from the enclosure’s own PSU. Most Thunderbolt ports can also charge the laptop over the same cable, but that charging wattage is unrelated to how much power the GPU itself draws from the enclosure’s internal supply, so a weak laptop charger is not a reason to assume the eGPU setup will underperform.

VRAM, Quantization, and Local AI Workloads

For running local AI models, VRAM capacity is usually the deciding spec, not raw GPU clock speed. Tools like Ollama and LM Studio load quantized model weights (commonly GGUF format) into VRAM and offload as many transformer layers to the GPU as will fit; models that don’t fully fit still run with partial offload, but fall back to slower CPU/RAM paths for the remainder. A desktop-class card in an enclosure with 16GB, 24GB, or more VRAM lets a laptop run larger quantized models or wider context windows than its internal GPU ever could, which is the core reason this category exists.

Driver stack support is uneven across operating systems. Windows and Linux both have mature eGPU driver paths, including CUDA for NVIDIA cards used by llama.cpp-based runtimes, while native eGPU support on Apple Silicon Macs is far more limited, so Mac buyers should confirm OS-level eGPU support before assuming a dock will work as advertised.

Setup, Drivers, and Hotplug Behavior

First-time setup typically involves authorizing the Thunderbolt controller to trust the new device, installing the GPU’s driver stack as if it were a fresh desktop install, and in some cases enabling Resizable BAR in firmware so the CPU can address the full VRAM pool efficiently rather than in small windows. Firmware updates for the enclosure itself are common in the first weeks of ownership and often fix early hotplug issues — cases where disconnecting and reconnecting the Thunderbolt cable doesn’t cleanly re-enumerate the GPU and requires a reboot.

Cable quality is not cosmetic here: Thunderbolt and USB4 cables are certified to specific bandwidth and length limits, and an uncertified or overly long cable can silently negotiate down to a lower link speed, which looks identical to a slow enclosure but is actually a cabling problem. Minding the cable’s bend radius near the connector also matters for long-term reliability, since repeated flexing at a tight bend is a common point of intermittent disconnects.

Noise, Heat, and Everyday Ownership

Sustained local AI inference keeps a GPU’s fans spinning for longer stretches than typical gaming bursts, so enclosure airflow design is worth weighing: open-frame and larger chassis generally cool a given card more quietly than compact boxes, which compensate for a smaller internal volume with higher fan speeds under load. Thermal throttling in a cramped enclosure shows up as gradually slowing token generation during long sessions rather than an outright failure, which can be mistaken for a software problem when it’s actually an airflow one.

Day-to-day ownership also means managing an extra power brick, an extra cable, and a chassis that needs a stable spot on the desk rather than living in a bag — a tradeoff against the raw VRAM and compute gains the category exists to deliver.

FAQ

Does Thunderbolt 5 make a real difference for local AI versus Thunderbolt 4?

Yes, mainly for model loading and VRAM transfer speed, where Thunderbolt 5’s higher PCIe throughput noticeably shortens the time to get a large quantized model into memory. Once a model is resident in VRAM, token generation speed depends much more on the GPU’s own memory bandwidth, so the inference-speed gap between Thunderbolt generations is smaller than the loading-speed gap.

Can I just buy any desktop GPU and put it in any enclosure?

Only if the card’s length, slot width, and power connector match the enclosure’s internal clearance and PSU wattage. Empty enclosures like the Razer Core X V2 or Sonnet Breakaway Box are built for this kind of GPU swap, while pre-populated docks like the MINISFORUM DEG1/DEG2 or GIGABYTE AORUS AI Box ship with a fixed card and are not meant to be reconfigured.

Will an eGPU work with a MacBook for running local LLMs?

Native eGPU support on Apple Silicon Macs is far more limited than on Windows or Linux, so a dock that works perfectly for Ollama or LM Studio on a Windows laptop may not be usable the same way on an M-series Mac. Check the specific OS-level eGPU support claims for the Mac generation in question before buying hardware on the assumption it will just work.