Intel Announces Arc Pro B50 and B60 Graphics Cards for Pro-Vis and AI Inferencing

[btarunr]

Intel at the 2025 Computex unveiled its Arc Pro "Battlemage" line of graphics cards powered by its latest Xe2 graphics architecture, and based on its 5 nm BMG-G21 silicon. The Arc Pro B50 is targeted squarely for professional visualization and graphics workstations; and comes with 12 GB of memory. The Arc Pro B60, on the other hand, has 24 GB of memory and has additional use-cases in the area of AI inferencing. Unlike AMD and NVIDIA, Intel is going to market for its Arc Pro B-series with board partner-based custom designs. These partners include ASRock, Gunnir, Maxsun, Sparkle, Onix, Senao, and Lanner.

As a pro-vis solution, the Arc Pro B50 series comes with a comprehensive set of certifications and validation by leading content creation applications. Intel is working on bolstering its AI inferencing product stack, with the debut of the new Project Battlematrix Linux software stack, and workstation platform, which enables up to 8 Arc Pro GPUs with 192 GB of memory between them, for accelerating 70 billion+ parameter models.



The Arc Pro B50 is based on a slightly cut down BMG-G21 silicon, with hardware features resembling the Arc B570 graphics card, but with a more thorough software and support stack; while the Arc Pro B60 is a maxed out card resembling the B580, but with double the memory across its 192-bit memory bus. Some partners, such as Maxsun, even unveiled dual-GPU cards based on the B60. These use two BMG-G21 chips, there is no bridge chip, but PCIe lane segmentation similar to M.2 NVMe risers is used. Each card exposes two GPUs to the host. Most custom-design Arc Pro B60 cards implement 600 W 12V2x6 power connectors for sufficient power for the two GPUs.



In the following slides, Intel put out its first party performance claims for the various use-cases of the B50 and B60.



View at TechPowerUp Main Site

 

[sudothelinuxwizard]

What's the pricing?

 

[Verpal]

I tried to keep an open mind but the memory bandwidth looks extremely anemic for running LLMs or other professional workload, that being said, more VRAM on the cheap is always welcomed.

 

[Dragokar]

What's the pricing?

"Suggested End User Price” (SEP) for B50 is 299$ US. No SEP for B60 yet.

 

[ErikG]

B50 is low profile, nice.

 

[vantila]

Wonderfull news! This confirms my suspicion: Intel doesn't see AMD as main competitor. Intel is directly aiming for Nvidia which is the right call.

AMD either as a result of lack of vision and incompetence, or because of practicing an unspoken duopoly with Nvidia, is either unwilling or incapable of competing with Nvidia. This is the only explanation that makes sense when AMD Nvidia competition history is taken into account.

AMD has to be either unable or unwilling to compete for de facto AMD competition to Nvidia to be this bad and miserable.

A bit of context: Gaming and AI are only two sector support Nvidia is providing. Wherever you can somewhat find a use case for gpu compute, Nvidia is there providing rigorous support.

What AMD is doing? NOTHING. EITHER SHEER INCOMPETENCE OR DUOPOLY.

What Intel is doing, beggining to induce competiton! Wonderfull!

 

[Sunlight91]

@vantila
Maybe AMD employees own nvidia shares. They make more money by releasing bad products and making nvidia look good.

 

[lemonadesoda]

Single Slot Low Profile designs. Nice. Shame there isn't more memory on a single GPU - their AI benchmarking doesn't run on a single GPU due to lack of memory? So much more efficient to have the whole model sitting in one onboard contiguous memory area rather than sucking up PCIe bandwidth and chopped liver single models sitting across multi GPUs. IMO. Gimme the 32GB edition. Why aren't they doing that?

 

[Stephen.]

while the Arc Pro B60 is a maxed out card resembling the B580, but with double the memory across its 192-bit memory bus.

Does this mean the B580 is a maxed out chip? this information was not clear before

 

[bgx]

surprised by the bandwith difference between B50 and B60, almost doubled??

300 USD for a pro card is nice though, even if of course the GPU is not very capable.

Single Slot Low Profile designs. Nice. Shame there isn't more memory on a single GPU - their AI benchmarking doesn't run on a single GPU due to lack of memory? So much more efficient to have the whole model sitting in one onboard contiguous memory area rather than sucking up PCIe bandwidth and chopped liver single models sitting across multi GPUs. IMO. Gimme the 32GB edition. Why aren't they doing that?

Indeed, i am not sure how they cut the LLM to use several GPUs.

May be a pipeline, where the first GPU takes care of the first few layers, then pass the answer to the other GPU to compute the next few layers?
You only have to pass the value of 1 layer around?

A quick read o, deepseek: 61 layers, so each layer would not be too large. Doable, but communication would be costly.

Other possibility, as it is expert based (1+8), you load some experts on some GPU, and some other on other GPUs, and you only need to merge the answer at the end.

Hey, what do I know, I only have a spotlight paper on "LLM" (CLIP) at ICML.

 

[10tothemin9volts]

In the pic INTEL is comparing GDDR6: their 192-bit B60 24GB (192-bit -> 12GB VRAM native) vs 128-bit 5060 Ti 16GB (clamshell) and 128-bit RTX 2000 Ada (clamshell), of course the 192-bit one will be always faster, but if price is the lowest common denominator and is lower, then that comparison would be ok, I guess.

Nice to see Linux stack mantioned.

I have said it previously and with more and more MOE LLMs coming out (e.g. Llama 4, Qwen 3 MOE ones), which although are bigger in size, run multiple times faster, I personally see less pressure to get a high VRAM GPU now, when reasonable inferencing speeds are achievable using RAM only.
24 GB unfortunatelly can't fill Qwen3-30B-A3B-Q6_K.gguf (MOE) or Qwen*-32B-Instruct-Q6_K.gguf (dense), but there's a B60 dual 48GB VRAM one.

.. Gimme the 32GB edition. ..

I agree, at least 32GB VRAM (preferably 36GB), is what I'd be interested it. Saw INTEL and B770 (256-bit: 16GB native, 32GB clamshell) mentioned recently, so a 32GB VRAM GPU may come out at some point if not canceled (of course it will, bc NVIDIA has a 32GB VRAM 5090).

 

[igormp]

Does this mean the B580 is a maxed out chip? this information was not clear before

Yes, it has always been, it's a fully unlocked BMG-G21 chip.
If you're thinking about the B770, that was rumored to be a G31 chip, which is a bigger one.

 

[PLAfiller]

So the B50 is expected to hit the shelves Q3 this year, which is right around the corner. Let's see how the MSRP of $300 translates into the real world. Cause the A50 is still selling ~€300 where I live, so I can't imagine the "new version" will be at the same price....