Gunnir Puts Two Intel Arc Pro B70 GPUs in a Compact AI Workstation

Gunnir has demonstrated a compact workstation built around two Intel Arc Pro B70 professional graphics cards at Intel Connection 2026 in China. The larger system pairs the two GPUs with an Intel Core Ultra 200HX Arrow Lake-HX processor platform and uses liquid cooling; Gunnir also showed a smaller single-GPU workstation using a vapor-chamber cooling design.

Each Arc Pro B70 carries 32GB of GDDR6 ECC memory, so the dual-GPU machine contains 64GB of graphics memory in aggregate. That is the headline capacity, but it needs an important qualification: the two cards retain separate memory spaces rather than behaving like one conventional 64GB GPU.

Intel specifies the Arc Pro B70 with 32GB of GDDR6 ECC on a 256-bit interface and 608GB/s of memory bandwidth per card. Two cards therefore provide 64GB of installed VRAM, but software has to support multi-GPU execution and distribute work appropriately to benefit from both devices.

That distinction matters for local AI workloads. A model or workload that cannot partition its data and compute across multiple GPUs cannot simply assume one contiguous 64GB allocation. Intel lists oneAPI support for the B70 and positions the product for scalable multi-GPU AI, but application-level support still determines whether a specific workload can use both cards effectively.

Intel's official B70 specification lists 32 Xe cores, 256 XMX engines, 32 ray-tracing units and up to 367 TOPS of dense INT8 XMX throughput. The Intel-branded card is rated at 230W, while Intel documents a broader 160W-to-290W range for partner implementations.

Those specifications explain why cooling becomes a central design problem in a compact dual-card workstation. Gunnir's liquid-cooled design is a system-level response to fitting two workstation-class GPUs into a smaller chassis; the event demonstration does not by itself establish sustained clocks, acoustic behavior or performance under long AI workloads.

The demonstration is technically interesting because Arc Pro B70 offers a relatively large ECC memory allocation per card and Intel explicitly supports oneAPI and OpenVINO on the GPU. A compact dual-B70 system could therefore target developers who need more aggregate accelerator memory without moving to a much larger workstation chassis.

Gunnir has not established a broad retail price or shipping schedule for these demonstrated systems in the cited coverage. Independent testing is also still needed for sustained thermals, noise, power draw and real multi-GPU scaling across common local-AI frameworks. Until those measurements exist, the workstation is best read as a concrete compact-system design rather than evidence of a particular performance-per-dollar advantage.