A few days ago we covered the Intel Arc Pro B50 SFF GPU and B60 GPU launch at Computex 2025. At $299 for a 16GB SFF GPU, the Arc Pro B50 has many folks interested, but the B60 with 24GB is going to be the better AI GPU. Taking that a step further, Intel and Maxsun showed a GPU that had not one but two GPUs in the Maxsun Intel Arc Pro B60 Dual GPU 48GB card. We found the card later on the show floor.
Maxsun Intel Arc Pro B60 Dual GPU 48GB at Computex 2025
On full display is the dual GPU card. The idea here is that one can put two 24GB GPUs on a single PCIe dual slot card and get 48GB of GPU memory per card. Something a bit strange here is that each GPU gets one DisplayPort and one HDMI port in the rear.
One of the other key enablers here is that this utilizes a blower-style cooler. That is important since that allows the GPU to be used in workstations (and even servers.)
350W or so is usually when we see cards transition to passive cooling instead using chassis fans to hit up to 600W TDP in dual slot PCIe cards, but for the workstation market, the blower-style cooler is easier to integrate since you do not need to do the chassis design around passively cooled GPUs.
Here is a shot at the card. Since the Intel Arc Pro B60 is an x8 GPU, it maks the design easy. Each GPU can get an x8 from the x16 connector so long as the CPU and slot support x8 bifurcation.
The real use for this is not a single card, but instead something that looks like this. A workstation with four of these cards with eight B60 24GB GPUs total. That gives us 192GB of video memory in a system that still has a shot to be powered by a standard wall circuit. Since the B60 can power down to 120W, a 120W to 150W per GPU solution, or 240-300W per card, would be more interesting since it is easier to power and cool in workstations.
Maxsun has another detail that may not be intentional at first, but might be useful. With only two video outputs per GPU, this is only 16 display outputs. Add one for the BMC’s VGA or iKVM video output and that is 17 in a system. Old versions of Windows had issues above 32 display outputs that we found when putting 10x GPUs in systems years ago. Having only two per card ensures that eight GPUs on four cards works a bit better for compatibility with older OSes.
Final Words
Just for some comparison, a modern generation RTX 6000 Ada with 48GB is still around $5500-6000 or so these days, so if these come in at $1000 or so (the single Arc Pro B60 is supposed to start around $375 MSRP) then that becomes an inexpensive way to get GPUs and, more importantly, GPU memory in a system. One could do this with two NVIDIA RTX Pro 6000 cards, but it would cost closer to $18-20K just for the two GPUs. Coincidentally many in the industry still call the RTX Pro 6000 the B40 as that was the name up until days before NVIDIA’s GTC 2025 announcement, which would have made this more confusing.
This could be a really neat and lower-cost way to add GPUs capabilities for graphics, virtualization, and AI workloads to smaller scale systems.
Yes, but will it SLI? Can it Crossfire? Inquiring minds want to know.
One IMHO very welcome development shown by the Maxsun Dual is that Intel seem to be quite open to let boardpartners develop their own board designs, and collaborate with them on those. Certainly very different from the tight-fisted control practiced by Nvidia.
Also @STH: could you ask Intel for (a lot) more detail what exactly they mean with their “Battlematrix” approach, and how they plan to implement spreading an LLM across up to 8 x B60. If the overhead isn’t eating a lot of the performance, getting 192 GB VRAM on a budget would be very attractive for running a strictly local LLM. Please follow up, this could be interesting!
@Stephen Beets SLI/Crossfire isn’t required because it supports PCIe P2P (Peer to Peer) and doesn’t bottleneck the host.
@sfjdfjhk Hey, that’s even better! No latency to bottleneck the host at all! Though my comment was mostly in jest, I am aware that SLI and Crossfire are kind of old hat now and not really supported any more (Nvidia replaced SLI with NVLink, a supposedly better protocol we’ve seen here on STH lots of times). So, maybe I’ll end it with “Two GPUs, one slot”. Anyway, thanks for playing along. :-)
8 bit per monitor on a “AAA” Game? please….you are only gonna have the game on one monitor……