Skip to main content

Intel Arc Pro B70

Product Overview​

Intel Arc Pro B70 is the newest member of Intel's Arc Pro B series of professional workstation graphics cards. Its official status is Launched, with a release date of Q1 2026. It is built on the Xe2 (Battlemage) microarchitecture using TSMC N5 process technology, extending the second-generation Xe architecture shared with Intel's consumer discrete GPUs.

In specifications, B70 is the direct successor to Arc Pro B60: memory capacity increases from 24GB to 32GB (GDDR6, 256-bit bus, 608 GB/s bandwidth), FP32 compute rises from 12.28 to 22.94 TFLOPS, INT8 compute rises from 197 to 367 TOPS, and TBP increases from 200W to 230W. The GPU contains 32 Xe-cores, 8 render slices, 32 ray-tracing units, plus 256 XMX (Xe Matrix Extensions) AI engines and 256 Xe vector engines.

Product positioning: the combination of 32GB memory and 608 GB/s bandwidth makes it a sweet spot for LLM inference and multi-GPU workstation platforms. Like the B60, B70 targets AI inference, professional graphics (AEC/CAD/3D rendering), and ISV-certified stability scenarios, with support for the IPEX (PyTorch extension), OpenVINO, and oneAPI software stacks. Note that Intel's official specification table states B70's peak compute figure in INT8 (TOPS) terms and does not publish FP8 or FP4 AI-training precision numbers.

Key Specifications​

ItemSpecification
ArchitectureIntel Xe2 (Battlemage)
ProcessTSMC N5
Memory32 GB GDDR6
Memory TypeGDDR6 (256-bit)
Memory Bandwidth608 GB/s
FP8Not disclosed
FP4Not disclosed
FP16Not disclosed
FP3222.94 TFLOPS
INT8367 TOPS
TDP230 W
InterconnectPCIe 5.0 x16
Release / ProductionQ1 2026, official status Launched

📌 Additional official specification fields (from Intel's spec table):

  • Xe-cores: 32; render slices: 8; ray-tracing units: 32
  • XMX AI engines: 256; Xe vector engines: 256
  • Clocks: maximum dynamic graphics frequency 2800 MHz, base graphics clock 2280 MHz
  • TDP (TBP): 230 W; Intel separately lists a low-power range of 160-290 W
  • ECC: the official specification table lists ECC memory support
  • Device ID: 0xE223; display output supports 4 monitors (DP 2.1 up to UHBR 13.5 / HDMI), max resolution 7680 x 4320@120Hz
  • Warranty: 3 years; no embedded options offered

⚠️ Precision caveat: Intel officially publishes only INT8 (367 TOPS) and FP32 (22.94 TFLOPS), and does not publish FP8 / FP4 / FP16 compute. Direct horizontal comparison with other GPUs' FP8 figures is therefore not possible; use the INT8 basis when evaluating.

Key Features​

  • 32GB GDDR6 at 608 GB/s: an increase over the B60 (24GB / 456 GB/s) in both capacity and bandwidth, key for serving larger models and longer contexts
  • 256 XMX AI engines: purpose-built for matrix operations and AI inference workloads, reaching 367 TOPS at INT8
  • 22.94 TFLOPS FP32: graphics and general compute performance suited to 3D rendering and visualization
  • TSMC N5 process: consistent with the Battlemage architecture generation
  • ISV certification: stability-optimized for AEC (architecture, engineering, construction) and professional graphics workstations
  • Full software ecosystem: oneAPI, OpenVINO, Intel Extension for PyTorch (IPEX), and XeSS super-resolution
  • ECC memory support: advantageous for long-running inference services
  • PCIe 5.0 x16: full host bandwidth, enabling multi-GPU workstation platforms

Use Cases​

  • ✅ LLM inference workstations (32GB can hold larger quantized models)
  • ✅ Multi-GPU parallel inference (scaling via Project Battlematrix-class workstation platforms)
  • ✅ AI development and model fine-tuning (IPEX + oneAPI software stack)
  • ✅ Professional graphics and 3D rendering (32 RTUs, 22.94 TFLOPS FP32)
  • ✅ AEC/CAD engineering workstations (ISV certified, ECC memory)
  • ❌ Large-scale model training (no official FP8/FP4 figure, limited memory capacity)
  • ❌ Direct compute comparison against datacenter GPUs (INT8 and FP8 bases are not directly convertible)

Comparison​

Data Sources​