Intel® Arc™ Pro B-Series Graphics Architecture

The Intel® Arc™ Pro B-Series Graphics family brings Intel Xe2-HPG graphics to professional workstations and AI inference systems. It includes the B50, B60, B65, and B70 models, which share XMX matrix engines and ray tracing hardware acceleration while scaling memory capacity, compute resources, and power.

Architecture overview

The series is based on Intel Xe2 architecture. The GPU is organized into Render Slices, each combining four Xe cores with graphics backends and associated caches. Xe cores contain Xe Vector Engines (XVEs) for programmable vector execution and XMX engines for matrix operations, while a shared L2 cache connects the execution units to dedicated graphics memory.

All four GPUs provide XMX-based AI acceleration and hardware ray tracing, and support up to four displays. Hardware video encoding and decoding supports AV1, HEVC, H.264, and VP9 on the B50, B60, B65, and B70.

Each Intel® Arc™ Pro B-Series Graphics model is designed for a distinct workstation or AI deployment profile:

  • B50: A compact, low-profile, dual-slot workstation card suited to local AI inference and professional graphics workloads.

  • B60: Designed for AI inference workstations and servers, including Linux multi-GPU large language model deployments.

  • B65: A professional workstation option for AI inference and visualization workloads.

  • B70: Targets advanced AI and high-end workstation workloads, including scalable multi-GPU large language model deployments on Linux.

Specification

The table consolidates the listed specifications for each GPU.

Specification

B50

B60

B65

B70

Xe cores

16

20

20

32

XMX engines

128

160

160

256

Peak INT8 throughput (pTOPS)

170

197

197

367

Peak FP32 throughput (TFLOPS)

10.65

12.28

12.28

22.94

Ray tracing units

16

20

20

32

PCIe interface

Gen 5 x8

Gen 5 x8

Gen 5.0 x16

Gen 5 x16

Dedicated memory capacity (GB)

16

24

32

32

Memory type

GDDR6

GDDR6

GDDR6

GDDR6

Memory bandwidth (GB/s)

224

456

608

608

Memory interface width (bits)

128

192

256

256

Total board power (W)

70

120–200

200

150–300

Peak INT8 throughput is expressed in pTOPS (peak tera operations per second) for dense INT8 workloads using XMX engines. Peak FP32 throughput is based on maximum clock frequency and peak single-precision operations. Actual application performance varies by workload and configuration.

For more comparison details and specifications, see Intel products comparison.