Intel Debuts 32GB Arc Pro B70 And B65 For AI Workstations

Intel is expanding its professional GPU lineup with two cards aimed squarely at AI workstations, the Arc Pro B70 and Arc Pro B65. Both arrive with a headline feature that speaks directly to modern model-serving needs: 32GB of onboard GDDR6 memory.

This is a deliberate play for data scientists, MLOps teams, and content studios that want fast local inference, larger context windows, and higher concurrency without jumping to datacenter-class hardware. The company is also pushing aggressive value, particularly with the Arc Pro B70’s sub-$1,000 price point.

Why Memory Capacity Is The Battleground For AI Workstations Today

AI workloads tend to be memory-bound. The size of a model you can host in a single card often determines throughput, latency, and even feasibility. A 13B-parameter LLM typically occupies roughly 26GB in FP16, according to guidance around deployments of models like Llama, leaving little headroom on 24GB cards. With 32GB, there’s room for larger context windows and heavier tokenization pipelines, or for running 13B models without trims. Quantized INT8 deployments can push even higher parameter counts per card.

That’s the calculus behind Intel’s move. Analysts at firms such as TrendForce have flagged tight supply in graphics and HBM memory; yet the demand curve is clear: more local VRAM means simpler, faster pipelines and fewer compromises when deploying on-prem.

Arc Pro B70 Specs And AI Throughput For Pro-Grade Inference

The Arc Pro B70 introduces Intel’s fastest Arc Pro silicon to date, built on the Battlemage architecture. It steps up to 32 Xe2-HPG cores totaling 4,096 shaders, with 32 ray-tracing units and 256 XMX AI engines. XMX blocks are Intel’s matrix accelerators, analogous in concept to Nvidia’s Tensor Cores, and they underpin the card’s stated 367 TOPS for INT8 inference.

Feeding those engines is a 256-bit memory controller paired with 32GB of GDDR6, delivering a quoted 608GB/s of bandwidth. In practice, that combination targets common workstation inference stacks—retrieval-augmented generation, multi-turn assistants, computer vision preprocessing—where memory footprint and bandwidth tend to dominate.

Intel says the B70 is tuned for multi-user serving. In internal demonstrations, the company reported higher token throughput under concurrent load than a rival in its price tier, highlighting the impact of both memory headroom and XMX acceleration. As always, real-world results will vary by framework, kernel optimizations, and the balance of CPU and storage in the host.

Arc Pro B65 Brings 32GB To The Midrange For Affordable AI Workstations

The Arc Pro B65 takes the graphics core used in Intel’s Arc B580 consumer card—also seen in the existing Arc Pro B60—and pairs it with a larger 32GB memory pool. That’s a notable jump from the B60’s 24GB ceiling and is aimed at teams standardizing on quantized LLMs, diffusion models with sizable UNets, or CAD/CAE projects that mix simulation and ML inference on the same box.

Intel is leaving B65 pricing to board partners, which will differentiate with cooling, form factors, and power connectors to fit varied workstation chassis.

Positioning Against Nvidia Workstation GPUs

Intel is clearly targeting Nvidia’s RTX Pro 4000, a Blackwell-based workstation card with 24GB of memory. Intel’s pitch centers on two levers: capacity and cost. With 32GB vs. 24GB, the Arc Pro B70 can host larger models or run the same models with more context and concurrent sessions. And at a list price of $949 for Intel’s own B70 variant, it undercuts the RTX Pro 4000, which lists at $1,899.

In Intel’s testing, the B70 posted stronger multi-user token-per-second metrics than the RTX Pro 4000. Independent MLPerf Inference submissions and cross-vendor, same-framework comparisons will offer a fuller picture once available, but the value proposition is unambiguous: bring more VRAM to the workstation tier at roughly half the cost.

Ecosystem Cooling And Power Details Across Partner Workstation Cards

Expect a broad slate of partner cards from ARKN, ASRock, Gunnir, Maxsun, and Sparkle. Designs vary from traditional dual- and triple-fan shrouds to a notable pass-through, fanless Maxsun prototype aimed at densely packed systems with strong case airflow. Given the B70’s higher power profile, truly silent operation may require performance trade-offs, but the thermal approach suits multi-GPU rigs.

Power connectors will differ by vendor: some boards use standard 8-pin PCIe, others adopt 12VHPWR. Intel also highlighted support for 2-, 4-, or 8-card configurations for scaling out model serving. Memory remains local to each GPU, but frameworks can shard or pipeline workloads across cards via PCIe peer-to-peer, making multi-adapter setups practical for larger models and higher concurrency.

What It Means For Studios And Labs Running On-Prem AI Pipelines

For small to midsize teams building private copilots, on-prem search, or media pipelines, the Arc Pro B70’s mix of 32GB VRAM, strong INT8 throughput, and sub-$1,000 pricing is compelling. It lowers the bar for running 13B FP16 models natively, or larger quantized models, without offloading to cloud GPUs. That can improve data governance, reduce latency, and cut ongoing costs.

On the software side, expect tight integration with Intel’s OpenVINO, alongside growing support in PyTorch and DirectML paths, plus oneAPI tooling for developers who want to squeeze more out of XMX. If supply constraints in graphics memory persist, early procurement will matter, but for now Intel has drawn a clear target: bring bigger models and better concurrency to the workstation, and do it at a price that reorders the segment.