Intel Arc Pro B65 brings 32GB GDDR6 and 608GB/s published bandwidth to a partner-card workstation GPU. Its 20 Xe-cores, 160 XMX engines and 200W specification place it below B70’s larger compute configuration while retaining the same nominal memory capacity and bandwidth.
Capacity is only the first gate
Official Qwen2.5-32B files show the practical range: Q4_K_M at 19.9GB leaves 12.1GB of nominal space, Q6_K at 26.9GB leaves 5.1GB, and Q8_0 at 34.8GB exceeds the pool. Those margins exclude cache, runtime workspace, display use and concurrency.
Host memory can affect a GPU load
A closed Intel issue documented a B65 load failure through SYCL/Level Zero even though the projected model allocation fit in VRAM; Vulkan loaded the same model. Follow-up found only 16GB host RAM, and increasing swap allowed the SYCL load. That makes host RAM, swap and staging behavior part of a reproducible report—not evidence that nominal VRAM alone decides success.
Read the full Arc Pro B65 32GB local-AI guide for exact specifications, model-file margins, backend caveats and the validation checklist.
- Calculate complete model fit
- Plan a 32B workload
- Load the B65 benchmark starter
- Watch Arc Pro B65
- Compare Intel Arc Pro tiers
- Compare Arc Pro B70
Published October 25, 2026. HighVRAM has not independently benchmarked Arc Pro B65; verify the current partner board, driver and application build.