External GPUs: PCI-SIG Brings Performance Close to 100%, but the Connection Costs More Than an RTX 5090

Anyone who has approached external GPUs or eGPUs will have noticed their biggest limitation: bandwidth. Regardless of the connection protocol and the chosen enclosure, the performance of the graphics card reduces by 30% to 50% compared to its full potential when paired with a desktop PC. Well, it seems that PCI-SIG has finally managed to develop a connection that allows for performance almost equal to that of a PCIe connection: CopprLink.

Preliminary tests were conducted by colleagues at PCWorld who adopted a configuration with an RTX 5090 connected via CopprLink and the same model installed in a traditional PCIe slot. The result shows an average deviation of just 2.29%, a figure that is well below any previous solution.

Moreover, CopprLink supports a complete PCIe 5.0 x16 connection, with a transfer speed of 32 GT/s per lane and a total bandwidth of 64 GB/s. This is a value four times higher than OCuLink, which stops at 16 GB/s with a PCIe 4.0 x8 configuration. This increase virtually eliminates the typical bottleneck of eGPUs.

The configuration used in the test is based on the HighPoint RocketStor 8631D chassis, a server-grade solution designed for professional environments and AI workloads. It integrates a 1300W power supply and supports large GPUs thanks to optimized airflow. The price of the chassis is around $1,300.

Additionally, there is the HighPoint Rocket 7634D adapter card, necessary to connect the host system via the CDFP connector, priced at about $999. Thus, the entire CopprLink ecosystem exceeds $2,300, excluding the GPU. With an RTX 5090, the total cost of the setup easily surpasses $5,000.

From an operational standpoint, the behavior is identical to that of a traditional connection. The system automatically recognizes the external GPU without the need for manual configurations. For the operating system, the graphics card appears as if installed directly on the motherboard, with full access to resources.

Benchmarks confirm the robustness of the solution. In various scenarios, the difference between eGPU and internal GPU is minimal, with variations sometimes regarded as mere margins of error. In fact, some of those lower values could be attributed to the fact that PCWorld used a PCIe 4.0 riser cable for the CDFP adapter card; if it had been PCIe 5.0, probably no difference would have been recorded.

Naturally, as it stands, the main limitation remains the cost. No consumer would spend more than the cost of today’s top-tier solution for an ecosystem that allows them to connect a GPU to a laptop. It’s even less likely that anyone would do so for a lower-end card.

Moreover, this is infrastructure designed for racks, data centers, or professional applications, which are poorly suited for gaming on-the-go. However, it still represents a fundamental step forward that virtually eliminates the gap between standard configurations and eGPUs. In the near future – perhaps with increased hardware availability that is currently lagging – it could lead to more affordable consumer solutions.