A 48GB dual-GPU Intel Arc B580 is reportedly in the works — Computex reveal rumored

DGIX

May 13, 2025

An AIB is reportedly preparing a dual-GPU variant of Intel’s rumored Arc B580 24GB, totalling 48GB of VRAM on a single board, via VideoCardz. According to the source, this model is slated to be revealed at Computex 2025, which is just days away at this point. Specific details like the AIB, interconnect technology, and price are in the dark, but we can expect more details at the trade show, if the rumors are true.

This is the third leak that references a 24GB edition of the Arc B580, following EEC filings from Maxsun and an earlier slip-up from Sparkle. It’s quite surprising that a GPU that’s supposed to rival the RTX 4060 is getting a 24GB memory configuration. Many PCs don’t even possess that much system memory. These cards are intended as a cost-effective solution for AI/ML developers, where similar capacity cards from AMD and Nvidia carry a steeper price tag. The most affordable Blackwell GPU with 24GB of memory is the RTX Pro 4000, costing over $1,500 based on preliminary listings.

An AIB is apparently doubling down on this approach by building a dual-GPU solution, housing two of these rumored 24GB B580 GPUs on a single PCB. Keep in mind, this product is reportedly a one-time design from the AIB, not a standard reference model from Intel.

Traditionally, multi-GPU setups nowadays, like Nvidia’s B200 and Apple’s M1 Ultra, rely on their own advanced interconnect solutions like NVLink and UltraFusion. While Intel does have Xe Link, it likely isn’t compatible with the B580 and would be too costly for a one-off project. The most probable contender is a PCIe bridge linking the GPU’s interfaces, allowing them to communicate through one slot.

Even so, the system will likely recognize the card as two separate GPUs. Some simulation programs where latency isn’t a concern could potentially utilize all 48GB as one large resource. However, as the memory isn’t pooled, using both GPUs for AI may require some optimizations on your end, including techniques like model parallelism and data parallelism.

With that in mind, don’t expect this GPU to be cheap. My estimate puts it around $700-$800, which should still be significantly lower than the rumored $4,500 figure for Nvidia’s RTX PRO 5000 Blackwell 48GB. However, you’d be sacrificing compute power and the convenience of a shared VRAM pool. Alternatively, you may consider mini-PCs or laptops powered by AMD’s Strix Halo, with up to 128GB of unified memory.

Follow Tom’s Hardware on Google News to get our up-to-date news, analysis, and reviews in your feeds. Make sure to click the Follow button.

Source link