Intel Arc Pro B70 & B65 Cards Debut: B70 AI Workstation GPU Starts At $949

Well, it's not coming quite the way some might have hoped, but today is the day for Intel's BMG-G31 GPU, releasing now as the Intel Arc Pro B70. This (very) long-rumored card sports Intel's most powerful Arc GPU to date, with a full 32 Xe2 cores and a 256-bit memory interface that wires up the big Battlemage GPU to 32GB of GDDR6 memory running at 19Gbps. There's also a cut-down Arc Pro B65 model that we'll talk about in a bit.


So why now, and why as Arc Pro? Well, because Arc happens to be pretty darn good at AI compute, and Intel is willing to max out the memory capacity of a graphics card without charging $2,000 for it. The Arc Pro B70 comes with 32GB of GDDR6 memory, and that's a boon for local AI processing. If you're the sort of fellow to run local LLMs or the latest high-end image generation models, the Arc Pro B70 is easily the most affordable way to get going.


Intel directly compares the Arc Pro B70 against competitor NVIDIA's twice-as-expensive RTX PRO 4000 Blackwell GPU and finds the competition wanting. More specifically, because Intel's GPU has 25% more memory onboard, it can apparently support a significantly larger context window with the Llama 3.1 8B language model. That means the AI can hold more information in memory, which in turn makes it better able to handle complex queries.


We won't go over all of Intel's benchmarks here; they're all AI-related and generally tell the same story. The blue team consistently shows its GPU offering superior responsiveness and ability to handle multiple users versus NVIDIA's card, even though peak throughput is a bit lower (as seen in the last graph.)


A key detail with the Arc Pro GPUs is that Intel is emphasizing multi-GPU performance. This is mostly at the software level, as there's no fancy interconnect like with NVLink. Still, assuming Intel's numbers hold up under independent testing, the scalability is impressive, and validates setups like the "Battlematrix" configuration with four Arc Pro GPUs in a single workstation. It will be interesting to see what four Arc Pro B70s can do.


In terms of the card itself, well, it's a graphics card. The Arc Pro B70 has, as we noted above, 32 Xe2-cores, which gives it 60% more peak per-clock compute throughput versus the Arc Pro B60, although the actual performance gain will be even higher than that, as it also runs higher clock rates, with the GPU clock peaking at an impressive 2.8GHz.

Intel puts the card down for 367 TOPS, versus 197 TOPS for its extant Arc Pro B60—and of the Arc Pro B65, which actually uses the same BMG-G31 GPU as the Arc Pro B70 but has the same 20-Xe2 core configuration as the Arc Pro B60. It's essentially the current top-end card, just with 32GB of memory and the associated increase in memory bandwidth.

Intel rates that card for 200W TDP, same as the Arc Pro B60, while the Arc Pro B70 is rated for "160-290W." The range is because Intel is allowing board vendors to set the power limits on their own GPUs. This is pretty cool, and it will be fascinating to see if anyone actually sells an Arc Pro B70 card rated for 290W.


We're surprised not to see a single 3D or HPC benchmark of any kind in Intel's materials; 3D rendering workloads can also be quite VRAM intensive, and so can things like computational fluid dynamics, big data analytics, and photogrammetry. We'll have to see if we can wrangle a card from Intel or one of its partners to see how it performs in those kinds of tasks—and maybe in a few games, too.

Intel says the Arc Pro B70 is available starting from today, both as an Intel-branded model and through the company's partners, like ASRock, Gunnir, and Sparkle (among others.) Intel's card retails for $949, while cards from AIB partners will vary in price depending on configuration and features. If you're more concerned about the memory than the speed, the Arc Pro B65 is launching in the middle of next month, exclusively from Intel's AIB partners.