Exploring Multi-GPU Setups with Intel Arc Pro B70 on Ubuntu 26.04: A Q&A Guide

Welcome to an in-depth look at Intel's recent progress in multi-GPU support under Linux, specifically using four Arc Pro B70 graphics cards on Ubuntu 26.04. Nearly a year ago, Intel launched Project Battlematrix to enhance its Linux driver for the Arc Pro B-Series, aiming to support up to eight GPUs per system. We’ve now tested this setup and compiled answers to the most pressing questions, covering hardware, driver improvements, and real-world benefits. Use the links below to jump to specific topics:

• What is Project Battlematrix?
• How does multi-GPU work with Intel Arc Pro GPUs on Linux?
• What hardware was used in this test?
• What driver improvements did Project Battlematrix bring?
• What are the benefits of running four Arc Pro B70 cards?
• What challenges might you face when setting up multiple Intel GPUs on Linux?

What is Project Battlematrix?

Project Battlematrix is an Intel initiative introduced about a year ago to improve Linux driver support for the Arc Pro B-Series graphics cards. Its primary goal is to enhance multi-GPU capabilities, allowing systems to harness up to eight Arc Pro GPUs simultaneously. Beyond just enabling multiple cards, the project focuses on open-source driver optimizations, particularly in the era of artificial intelligence workloads. By refining the open-source driver stack, Intel aims to make the Arc Pro lineup more appealing for compute-intensive tasks—such as machine learning training or scientific simulations—where multiple GPUs can dramatically boost performance. This initiative marks a significant step for Intel in the Linux ecosystem, as it directly addresses the demand for scalable, high-performance GPU clusters without relying on proprietary drivers.

How does multi-GPU work with Intel Arc Pro GPUs on Linux?

Multi-GPU support in Intel’s open-source driver for Arc Pro GPUs leverages the Linux kernel’s Direct Rendering Manager (DRM) and the Mesa graphics library. When you install multiple cards—up to the eight allowed by Project Battlematrix—the driver treats each GPU as an independent device, but workloads can be distributed across them via frameworks like OpenCL, SYCL, or Vulkan. This is a key difference from older SLI/CrossFire setups, which required explicit game profiles. Here, the focus is on compute and professional rendering, so applications must be designed for multi-device execution. Intel has also worked on improving memory sharing and inter-GPU communication, reducing overhead for parallel tasks. In our test with four Arc Pro B70 cards, Ubuntu 26.04 recognized each one immediately, and basic compute tasks showed near-linear scaling without driver modifications.

What hardware was used in this test?

To evaluate the multi-GPU state of Intel’s Arc Pro GPUs on Linux, we obtained four Arc Pro B70 review samples. Each B70 is a workstation-class card based on the Alchemist architecture, equipped with 16GB of VRAM and support for ray tracing, AI acceleration, and multi-display outputs. The system ran Ubuntu 26.04 with the latest upstream kernel and Mesa drivers. We used a motherboard capable of supporting four PCIe x16 slots and a power supply rated for the combined 300W TDP of the cards. This configuration allowed us to test not only driver stability but also thermal and power management under sustained compute loads. The B70’s robust cooling solution ensured that even when all four cards were under 100% utilization, temperatures stayed within safe limits, confirming that the hardware is well-suited for dense multi-GPU builds.

What driver improvements did Project Battlematrix bring?

Project Battlematrix delivered several key enhancements to the open-source driver stack for Arc Pro GPUs on Linux. First, it addressed the multi-GPU scheduling in the kernel’s DRM subsystem, ensuring that up to eight cards could be enumerated and managed without conflicts. Second, it optimized memory fragmentation handling, which previously caused performance degradation when multiple GPUs shared system resources. Third, the project included firmware-level tweaks for better power balancing and clock synchronization across cards. These changes are transparent to the user—after updating to the latest kernel and Mesa, the driver automatically supports the new capabilities. For our test with four B70 cards, we observed stable multi-GPU rendering in Blender and accelerated inferencing in PyTorch, with no need for manual configuration beyond installing the cards and booting the system. This is a dramatic improvement from earlier iterations where multi-GPU setups were notoriously tricky under Linux.

What are the benefits of running four Arc Pro B70 cards?

Running four Arc Pro B70 cards provides significant advantages for compute-heavy workflows. In AI/ML tasks, batch sizes can be increased, reducing training time nearly in proportion to the number of GPUs (assuming the model fits in combined memory). For example, a deep learning model that takes 8 hours on one B70 could be trained in roughly 2 hours when split across four cards. For 3D rendering, Cycles-based engines distribute tiles across GPUs, doubling performance over a dual-card setup. Additionally, the 16GB VRAM per card offers up to 64GB total, enabling larger datasets or scenes to be held in video memory. Professional applications like DaVinci Resolve and OctaneRender also benefit from multiple GPUs for video encoding and ray tracing. Finally, mining and scientific computing tasks see similar scaling, making this configuration cost-effective for workstations that require high parallel throughput without investing in server-grade hardware.

What challenges might you face when setting up multiple Intel GPUs on Linux?

Despite the improvements, some challenges remain when deploying multiple Intel Arc Pro GPUs on Linux. First, physical hardware constraints: you need a motherboard with enough PCIe lanes and physical x16 slots, plus sufficient cooling and power. On the software side, while Ubuntu 26.04’s out-of-the-box support is good, older distributions may require manual kernel and Mesa updates to unlock all features. Not all applications are multi-GPU aware, so you must use frameworks that explicitly leverage multiple devices—for instance, OpenCL or SYCL instead of generic OpenGL. Debugging inter-GPU synchronization can be tricky, especially when memory allocation isn’t balanced. Finally, because the open-source driver is still maturing, occasional stability issues (like GPU hangs during heavy workloads) may arise. Intel’s active contributions to the kernel help mitigate this, but users in production environments should verify application compatibility and keep their driver stack current. Our test showed these are manageable, but they still require attention from system integrators.