Intel's Arc B70: 32GB of VRAM for AI, Not Gaming

Intel just shipped one of the weirdest GPUs [you can buy right now. The Arc Pro B70 isn't trying to be an RTX killer, and it's definitely not aiming for the data center. It's something in between—officially a workstation card, but really built for people who want to run large language models locally without remortgaging their house.

The specs tell most of the story: 32GB of VRAM, roughly 600 GB/sec memory bandwidth, and a 250-watt power envelope. That's not earth-shattering if you're thinking about gaming performance or raw compute. But if you've been trying to run modern AI models on consumer hardware, those numbers suddenly get very interesting.

Level1Techs got their hands on the B70 (technically four of them) and the testing reveals both the promise and the problem with Intel's approach. The promise: this card can actually handle serious local AI workloads. The problem: Intel's software ecosystem is perpetually playing catch-up with the model releases everyone actually wants to use.

The Hardware Intel Built (Maybe By Accident?)

There's something fascinating about the B70's positioning. Intel won't call it the successor to the A770 that enthusiasts have been begging for, but if you squint at the specs, it's suspiciously close. "This might be the most interesting GPU that Intel has done in years, but facing the steepest of software uphill battles," the Level1Techs team notes.

Intel's pitch centers on XMX acceleration (their answer to NVIDIA's tensor cores), open software, and enough memory to run models that would choke on cards with less VRAM. The Battlemage architecture underneath is Intel's Xe2, which showed up first in the B50 and B60 workstation cards launched at Computex almost a year ago.

What makes the B70 compelling isn't the TOPS number or ray tracing capabilities—it's purely about memory. 32GB in a consumer-adjacent price point (Intel hasn't published official pricing, but Level1Techs mentions it costs less than AMD's Instinct solutions) changes the calculation for anyone running local inference.

The benchmarks back this up. Running Qwen 3.5 27B (a dense model that's genuinely challenging), the four-card setup achieved 369 tokens per second with peaks hitting 550 tokens per second under parallel load. That's with 50 concurrent requests and a 200,000 token context window. Not bad for hardware that arrived literally yesterday.

The Software Problem That Won't Go Away

Here's where things get messy. Intel maintains LLM Scaler, which is described in their own repo as "an extension of vLLM or a version of vLLM." If you're not deep in the AI infrastructure world, vLLM is the de facto standard for serving large language models efficiently. Intel's fork means they can optimize for their hardware, but it also means they're always trailing upstream development.

This matters more than it sounds. When a hot new model drops—like Qwen 3.5 did recently—everyone wants to try it immediately. But Qwen 3.5 was built against nightly vLLM builds and uses SGLang, which meant Intel's stack wasn't ready. "The hardware is finally getting interesting at the exact moment the model ecosystem is moving too fast for lagging forks and compatibility layers," Level1Techs explains.

Intel apparently got Qwen 3.5 support working right before this hardware hit reviewers' hands (the video literally got edited to mention Intel's new partnership with vLLM that dropped the same day). That's cutting it close. Like, uncomfortably close.

The comparison to AMD's ROCm situation is instructive. ROCm has matured significantly, but the experience varies wildly depending on which RDNA hardware generation you're running. Intel's in a similar position, except their software surface area is smaller since they have fewer SKUs to support. That's good for focus, but only if they can maintain momentum.

"For this card to matter, 'works eventually' is not good enough," the Level1Techs review emphasizes. "It has to work when the model is hot, not after the hype cycle has moved on."

CUDA's dominance isn't about technical superiority at this point—it's pure inertia. Everyone builds for it first because everyone else builds for it first. Intel's asking developers and power users to accept more friction in exchange for openness, better memory per dollar, and an alternative to NVIDIA's ecosystem. Whether that trade-off makes sense depends entirely on how reliably Intel can ship software updates.

What You're Actually Getting

If you're thinking about the B70 for gaming, stop. The Windows drivers weren't even functional at the time of Level1Techs' testing. This isn't a gaming card wearing a workstation badge—it's genuinely designed for a different workload.

What it is good at: local AI inference, especially when you start stacking cards. Two B70s gives you 64GB of VRAM. Four gives you 128GB. That's getting into territory where you can run models that normally require renting cloud compute. Level1Techs points out this four-card setup costs less than AMD's Strix Halo while delivering better inference performance (though admittedly, that's comparing dedicated GPUs to an APU, so it's somewhat apples-to-oranges).

The B70 also supports SR-IOV (single root I/O virtualization), meaning you can partition the card into virtual GPUs for different workloads or VMs. That's not a feature most consumer cards bother with, and it makes sense for workstation scenarios where you might want to isolate different AI projects.

Memory bandwidth sits around 600 GB/sec, which isn't class-leading but paired with 32GB capacity, it's enough for the models this card is targeting. The 250-watt power envelope is also remarkably sane—you're not going to need a separate breaker box for a multi-card setup.

The Bet Intel's Making

Intel's positioning the Arc Pro B series around a specific thesis: that there's a market for people who want to run AI workloads locally, who don't need data center features, but who need more memory than consumer gaming cards offer. It's a narrow slice of the market, but it might be a real one.

The enterprise world has figured out that not everything needs to run in the cloud. Privacy concerns, latency requirements, and cost optimization all push toward local compute for certain workloads. The B70 slots into that gap—too much card for hobbyists, not enough for true data center deployment, but maybe just right for small teams or power users who've hit the ceiling on consumer hardware.

Intel's partnership with vLLM (announced literally as this hardware launched) suggests they understand the software problem. Whether they can execute on it is the open question. The AI model ecosystem moves fast. New architectures, new training techniques, new optimizations drop constantly. Intel needs to be responsive in days, not quarters.

The B70 represents something potentially important: a GPU designed explicitly for the local AI use case that's actually interesting on paper. The memory configuration is right, the power envelope is reasonable, and the pricing (when it's revealed) might be competitive. But hardware is only half the product. The other half is whether Intel's software team can keep up with an ecosystem that never stops moving.

— Yuki Okonkwo