Intel Launches New GPUs and AI Accelerators for AI Workstations

At Computex Taipei 2025, Intel introduced two new AI-focused hardware products: the Arc™ Pro B-Series GPUs for creative and edge applications, and the third-generation Gaudi AI accelerator for large-scale model training. At the same time, Intel upgraded its development toolchain to enhance hardware-software synergy and further strengthen its AI ecosystem.

New Intel Arc Pro B-Series GPUs: Built for Edge AI and Creative Work

Intel’s Arc Pro B60 and B50 graphics processors are based on the second-generation Xe architecture (Xe2) and feature integrated XMX (Xe Matrix eXtensions) AI engines. These GPUs are optimized for professional workloads such as architectural design, engineering modeling, graphic creation, and AI inference, delivering high performance and energy efficiency for edge devices and creative workstations.

Arc Pro B60 (flagship model)

Featuring 20 Xe Cores and 160 XMX AI Engines

24GB GDDR6 memory, 192-bit bus, up to 456 GB/s of bandwidth

Up to 197 TOPS of inference performance at INT8 precision

TDP range of 120W to 200W, supporting multi-GPU configurations

Certified for mainstream ISV applications such as SolidWorks, Maya, and Blender

Arc Pro B50 (energy efficient version)

Equipped with 16 Xe cores and 128 XMX engines

16GB GDDR6 memory, 128-bit bus, up to 224 GB/s bandwidth

Inference performance up to 170 TOPS at INT8 precision

Just 70W TDP, ideal for small form factor workstations or light AI tasks such as image enhancement and CAD acceleration

Both GPUs support professional and consumer driver stacks on Windows platforms and are fully compatible with ISV-certified software, enabling seamless AI integration in local workstations.

Intel Gaudi 3 AI Accelerators: Powering Enterprise AI Training

Designed for large model training and high-efficiency inference, Intel’s third-generation Gaudi AI accelerator—Gaudi 3—brings significant improvements in compute, memory, and networking, directly competing with NVIDIA’s H100.

Hardware Specifications

TSMC 5nm process, dual-die architecture

8 Matrix Multiplication Engines (MMEs) and 64 Tensor Processing Cores (TPCs)

128GB HBM2e memory with 3.7 TB/s bandwidth

Supports FP8 and BF16 formats with up to 1.8 PFLOPs peak performance

Performance Advantages

2× the MMEs, 1.5× memory bandwidth, and 40% higher energy efficiency vs. Gaudi 2

Intel claims 1.7× faster LLM training and 2.3× more efficient inference than NVIDIA H100 (based on Llama2-13B benchmarks)

Networking and Scaling Architecture

24 integrated 200GbE Ethernet ports with RDMA per chip

Supports All2All topology with up to 150 GB/s inter-node bandwidth

Scales to 512 nodes using open Ethernet standards, eliminating the need for proprietary switching.

Software Support: Habana Synapse AI SDK

The Habana Synapse AI SDK provides native support for PyTorch, TensorFlow, and ONNX, enabling optimized training workflows and flexible workload scheduling across MME and TPC engines.

Supports Both PCIe and Rack-Scale Deployment

Intel Gaudi 3 offers two deployment formats to meet AI compute needs across different scales:

Gaudi 3 PCIe Accelerator Card: Designed for mainstream datacenter servers, it supports flexible expansion and is ideal for SMBs and R&D teams to run inference workloads across models like Llama 3.1 8B to Llama 4 Maverick. Available in the second half of 2025.

Gaudi 3 Rack-Scale System: Supports up to 64 accelerators per rack, delivering a total of 8.2 TB of high-bandwidth memory. Built with an open, modular design and liquid cooling, it’s compatible with Open Compute Project (OCP) standards, enabling CSPs and large enterprises to build scalable, low-latency infrastructure for large-scale training and inference.

This dual-format deployment strategy reflects Intel’s vision of building an “open, flexible and secure” AI infrastructure, supporting smooth expansion from a single machine to a large cluster.

Intel AI Assistant Builder: Open Source Release

Alongside the hardware launch, Intel officially open-sourced its AI Assistant Builder tool. The framework enables developers to run lightweight AI agents locally on Intel platforms, with support for containerized deployment on Linux and Windows.

Intel’s software stack continues to evolve to take advantage of MME and TPC architecture, enabling task-level scheduling and efficient mixed-load processing for developers building edge or enterprise AI solutions.
This open-source initiative aligns with Intel’s commitment to empowering developers with accessible, efficient, and flexible AI tools.

Intel Challenges the Status Quo in AI Compute

While NVIDIA continues to lead in AI acceleration, Intel’s Arc Pro GPUs and Gaudi 3 AI accelerators offer a compelling alternative—focusing on open standards, scalable infrastructure, and power-efficient performance.

Both Arc Pro and Gaudi 3 are expected to be commercially available in the second half of 2025, driving the next generation of AI performance across workstations, datacenters, and enterprise infrastructure.