Short answer: Intel Arc Pro B60 provides 24GB of dedicated GDDR6 memory in a workstation GPU, placing it in the same nominal capacity tier as other 24GB cards while using Intel’s Xe2, XMX, oneAPI and OpenVINO ecosystem. It can hold many quantized 32B artifacts that exceed 16GB, but capacity alone does not prove that your exact runtime, driver, backend, operators or quantization work correctly.
Check both gates: calculate weights, KV cache, workspace, display use and concurrency with the hardware recommender, then verify the exact Intel driver and oneAPI/OpenVINO/SYCL or Vulkan path for the application and model format.
Official Intel Arc Pro B60 specifications
| Official specification | Arc Pro B60 | Local-AI planning meaning |
|---|---|---|
| Microarchitecture | Xe2 | Requires a current Intel software stack and compatible backend |
| Dedicated memory | 24GB GDDR6 | Weights, cache, workspace, display use and concurrency share capacity |
| Memory interface | 192-bit | Interface width alone does not establish application speed |
| Published bandwidth | 456GB/s | Peak bandwidth is not a token-rate measurement |
| Xe-cores / XMX engines | 20 / 160 | Do not normalize directly against CUDA cores or AMD compute units |
| Graphics clock | 2,400MHz | Sustained clocks depend on the specific board and workload |
| Total board power | 200W; 120–200W low-power range | Partner implementations and power limits vary |
| PCIe configuration | PCIe 5.0 x8 | Confirm motherboard slot wiring and the exact board design |
| Display support | Up to four displays | Desktop display use can consume part of the 24GB pool |
| Published APIs | oneAPI, OpenVINO, IPEX, Vulkan 1.3, OpenCL 3.0 | API availability is not blanket application compatibility |
What 32B files fit in 24GB?
These official Qwen2.5-32B-Instruct-GGUF repository sizes establish file-only boundaries. They exclude KV cache, runtime workspace, temporary buffers, display use, modality, context and concurrency.
| Official artifact | Displayed size | 24GB file-only margin | Planning result |
|---|---|---|---|
| Q2_K | 12.3GB | 11.7GB | Capacity fits; validate quality and the complete allocation |
| Q3_K_M | 15.9GB | 8.1GB | Leaves more nominal room for cache and workspace |
| Q4_K_M | 19.9GB | 4.1GB | Capacity-sensitive; calculate the full run |
| Q5_K_M | 23.3GB | 0.7GB | Not a safe resident target |
| Q6_K | 26.9GB | Does not fit wholly | Requires offload, another placement or more device memory |
| Q8_0 | 34.8GB | Does not fit wholly | Move to another artifact or capacity tier |
Read the 32B allocation guide, compare the 24GB buyer paths, or see where B60 sits in the Intel Arc Pro B-Series guide.
Software support must be verified on the exact stack
Intel lists oneAPI, OpenVINO and Intel Extension for PyTorch support for B60. Upstream llama.cpp’s SYCL documentation covers Intel Arc B-Series and now discusses B60-specific optimization work, but its verified-device table is not a guarantee for every Arc Pro board, operating system, driver, build or quantization. Confirm that the B60 appears through Level Zero or the intended backend, test the required operators, and record any host-memory fallback or CPU offload.
Board and system checks
- Verify exact partner board, 24GB runtime-visible capacity, cooling, dimensions and slot width.
- Confirm connector, PSU, sustained power limit and case airflow.
- Check PCIe slot wiring; the product specification is PCIe 5.0 x8.
- Record OS, driver, oneAPI/runtime/backend and application versions.
- Do not assume two B60 cards form one automatic 48GB allocation; verify explicit multi-GPU support and placement.
Measure instead of inferring
Hold model revision, quantization, runtime/version, backend, kernels, context, cache precision, prompt/output profile, batch, concurrency and timing definition constant. Record peak device memory, system RAM, CPU fallback/offload, time to first token, generation rate, wall power and failed runs. Use the controlled workload method.
Load the exact Arc Pro B60 benchmark starter; it stages hardware identity, 24GB capacity and the topology checklist only, leaving every measured field blank.
Continue the decision
Compare larger capacity tiers · Browse GPU comparisons · Check screened observations · Stage the exact B60 watchlist family
Primary sources
- Intel Arc Pro B60 official specifications
- Intel Arc Pro B60 datasheet
- llama.cpp SYCL backend documentation
- Official Qwen2.5-32B GGUF repository
- HighVRAM methodology
Reviewed October 25, 2026. HighVRAM has not independently benchmarked Arc Pro B60. Specifications, board implementations, drivers and software support can change.
Choosing between compact 16GB and partner 24GB? Compare Arc Pro B50 and B60 by model fit, power, form factor and software validation.
Need more than 24GB without changing listed compute resources? Compare Arc Pro B60 and B65 by exact model-fit boundary, memory system, partner board and measured evidence.
Comparing the 24GB B60 with CUDA? Use the Arc Pro B60 versus RTX 3090 guide to separate identical capacity from Intel-stack support, CUDA maturity, power and used-card risk.
Same-capacity cross-platform comparison: Compare Intel Arc Pro B60 24GB with GeForce RTX 4090 24GB by software support, power, system requirements and matched evidence.