A patch submitted today to the AMD graphics driver mailing list officially opens the Linux chapter for Barco’s professional MXRT graphics cards. Long deployed in multi-display medical imaging systems—where ‘mission-critical’ is no exaggeration—these cards have until now been tightly bound to Windows drivers. That limitation kept the hardware out of the Linux ecosystem, depriving medical facilities of a potential local AI inference resource exactly where clinical data privacy is sacrosanct.

Barco, a European technology company, builds its MXRT cards around AMD Radeon Pro GPUs (and MXRV on NVIDIA architecture). The decision to release Windows drivers first, and only now begin Linux development, reflects a long-standing medical-sector pattern: certified diagnostic software runs almost exclusively on Windows. But the landscape is shifting.

The driver bottleneck and the rise of local inference

Linux support is the prerequisite for pulling these cards into modern compute stacks built on containers and Kubernetes, where generative AI and computer vision models leverage GPU-accessible frameworks like PyTorch and libraries like ROCm. Without Linux drivers, MXRT cards remain confined to closed Windows workstations, ruling out the possibility of running quantized LLMs or convolutional networks directly on the same systems that drive radiology displays.

For those designing on-prem AI deployments in healthcare, Linux driver availability means the possibility of reusing existing hardware—or equipment already accounted for in budgets—without having to deploy costly dedicated servers packed with enterprise-grade GPUs. This goes right to the heart of TCO and data sovereignty debates.

Medical imaging and on-premise: a forced marriage

GDPR and national healthcare regulations mandate that diagnostic images and reports cannot be processed outside institutional or regional boundaries without complex anonymization measures. Inference of AI models for diagnostic support (fracture detection, lesion classification, organ segmentation) therefore cannot rely on public cloud APIs without legal risk. In this context, having Linux-compatible GPUs already installed in radiology carts or reporting stations avoids additional purchases and simplifies regulatory validation.

Barco has yet to publish detailed computational specifications, but the lineage from Radeon Pro GPUs points to a good balance of computational power and power consumption suited for near-real-time processing of high-resolution images. The arrival on Linux may also encourage the open-source community to develop optimized drivers, possibly building on work AMD has already done for FirePro and Radeon PRO WX series.

Implications for on-premise LLM deployment

While MXRT cards are not designed for large model training, their floating-point capability and VRAM (presumably between 8 and 16 GB, extrapolating from Radeon Pro equivalents) make them interesting candidates for 4-bit quantized language model inference or for feature extraction pipelines applied to radiology text reports. In a hospital, a local LLM analyzing written radiology reports to suggest standard coding or flag linguistic anomalies can operate securely with no external communication.

For system administrators and IT managers, the driver opening signals a broader direction: AMD is progressively expanding Linux support for its professional GPU lineup, narrowing the gap with NVIDIA in the “sovereign AI” segment. The news is one piece of a puzzle we track regularly at AI-RADAR, where we offer analytical frameworks for those evaluating how to bring LLM workloads onto on-premise infrastructure.

Looking ahead

Today’s patch is only a first step. It will take time for the drivers to be accepted into the mainline kernel or AMD’s repositories, and for medical software vendors to start certifying Linux configurations. Yet the door is ajar: for sector professionals who want full control over data and reduced cloud vendor lock-in, having a proven hardware option that is now also opening up to Linux represents an evolution worth watching closely.