B60 Dual style graphics card close up

Maxsun Arc Pro B60 Dual Review: 48GB on One Card for Local AI, With Strings Attached

The Maxsun Arc Pro B60 Dual packs two Intel GPUs and 48GB of VRAM onto one board. What it runs, what it needs from your motherboard and whether it beats used 3090s.

This review is built from Maxsun’s and Intel’s published specifications, retail listings and third-party coverage rather than hands-on use. Figures are attributed in the text, and some speed figures are estimates derived from bandwidth, which are labelled as such.

The Maxsun Arc Pro B60 Dual is the only new graphics card that puts 48GB of memory on one board for around $1,200. It does that by doing something nobody had done in years: putting two complete GPUs, two Intel Arc Pro B60 chips with 24GB each, on a single card.

The honest answer is that the B60 Dual is a great idea with real strings attached. It is two GPUs, not one big one, so your motherboard must split its slot, and your software must split each model across both chips. Each half has only 456 GB/s of bandwidth, so large models run slowly. And it sits on Intel’s AI software stack, which is still maturing. If you want 48GB per slot on a budget and enjoy tinkering, it is unique. If you want something that just works, two used RTX 3090s remain the safer path.

Table of Contents

B60 Dual workstation PC tower build
Not a Maxsun card in this photo. 48GB on one board only helps if your motherboard can split the slot and your software can split the model.

What the B60 Dual Offers

Maxsun unveiled the card at Computex 2025 under its Cut the Cloud slogan, and it began reaching buyers later that year. Wccftech, via Let’s Data Science, reported a formal launch with Intel on April 2, 2026.

Specification Maxsun Arc Pro B60 Dual 48G Turbo
GPUs 2 x Intel BMG-G21 (Battlemage)
Memory 48GB GDDR6, 24GB per GPU
Memory bus 2 x 192-bit
Bandwidth 456 GB/s per GPU
Shader units 5,120 FP32 in total
Interface PCIe 5.0, x8 per GPU
Board power About 400W
Form factor Dual slot, blower cooler

PC Games Hardware listed the two BMG-G21 chips, 5,120 shader units and two 192-bit memory interfaces. 24wireless confirmed the per-GPU bandwidth of 456 GB/s. Jon Peddie Research noted that the card supports PyTorch, Intel’s ISV certifications and vLLM.

Dual-GPU cards were common in gaming a decade ago, then disappeared. AllMind pointed out that the last notable examples were NVIDIA’s GTX Titan Z in 2014 and AMD’s Radeon Pro Duo in 2016. This one exists for a different reason: memory capacity for AI.

Two GPUs, Not One: Why It Matters

The single most important thing to understand about the B60 Dual is that your computer sees two separate graphics cards. Each has its own 24GB of memory. A model larger than 24GB must be split across both, exactly as it would be across two physical cards.

There are two ways to split a model. Layer splitting puts some layers on each GPU and passes data between them once per token. It is simple and supported by llama.cpp, but each GPU sits idle while the other works, so speed is limited by one GPU’s bandwidth at a time. Tensor parallelism splits each layer across both GPUs so they work simultaneously, which can nearly double speed, but it needs fast communication and support in your inference engine, such as vLLM.

So 48GB on one card is not the same as 48GB on one GPU. It is a convenient, compact way to get two 24GB GPUs into one slot.

B60 Dual Motherboard Requirements

StorageReview explained that each GPU connects through its own PCIe 5.0 x8 interface. For both to work, your motherboard must support splitting the x16 slot into two x8 links, a feature called bifurcation.

Many workstation and server boards support this. Many consumer boards do not, or support it only on specific slots. Without bifurcation, only one of the two GPUs will appear, and you will have paid for a 24GB card. Check your motherboard manual for x8/x8 bifurcation on the slot you plan to use before buying.

This is the most common reason the card disappoints its owners, and it is entirely avoidable with five minutes of reading.

B60 Dual Price and Availability

StorageReview tested a unit that cost $1,200 directly from Maxsun, in line with Intel’s original announcement, while warning that memory price swings could change that. Hardware Corner called it the new local LLM value king at that price.

European prices ran higher. PC Games Hardware and Profesional Review reported German listings at about €1,499 to €1,500 when the card arrived there. Availability has been patchy, and the 2026 memory shortage pushed many graphics card prices up, so check current listings before you plan a build.

At $1,200, the card costs roughly $25 per gigabyte of VRAM, lower than any other new option.

B60 Dual Performance: What to Expect

Independent, model-by-model benchmarks for this exact card remain limited. The best guide is bandwidth, which sets the ceiling for single-user text generation.

Each GPU has 456 GB/s. A model that fits in one GPU’s 24GB runs as it would on a single Arc Pro B60. For a model split across both with layer splitting, the time per token is roughly the time to read all the weights at 456 GB/s. A 70B model at 4-bit is around 40GB, which puts the theoretical ceiling at about 11 tokens per second, before software overhead. Real results will be lower.

For comparison, the PMZFX benchmark repository measured two Arc Pro B70 cards, each with 608 GB/s, running Llama 3.3 70B at Q4 at 11.5 tokens per second. The B60 Dual has less bandwidth per GPU, so expect it to be slower on the same model. That is usable for careful work, not fast.

Mixture-of-experts models are much better suited. Because they read only a few billion parameters per token, a 30B to 35B MoE model that fits on one 24GB half runs quickly, and larger MoE models split across both halves stay responsive.

Intel Software in Late 2026

Intel’s AI software is the main risk. llama.cpp supports Arc GPUs through its SYCL and Vulkan backends, and Intel maintains a vLLM branch with Battlemage support, which is what you need for tensor parallelism across both GPUs.

The stack is still young. StorageReview described Intel’s LLM Scaler vLLM branch as closer to beta in its Arc Pro B70 review, with limited model support and some quantisation paths not yet working. Intel also archived its ipex-llm repository in January 2026, one of the libraries Jon Peddie Research listed as supported when the card launched. Software plans have shifted before and may shift again.

Ollama and LM Studio support for Intel GPUs is less mature than for NVIDIA and AMD. Expect to run llama.cpp or vLLM directly, and to spend time on setup. Our Arc Pro B70 review covers the Intel stack in more detail.

Models That Suit 2 x 24GB

Choose models with the card’s structure in mind.

  • Two different models at once. Run a coding model on one GPU and a chat or embedding model on the other. This uses the hardware perfectly, with no splitting at all.
  • 30B to 35B mixture-of-experts models on one half. Fast, with the other half free.
  • Large MoE models across both halves. Good quality at usable speed.
  • 70B dense models at 4-bit across both. They fit, as Maxsun advertised with DeepSeek R1 70B, but expect single-digit to low double-digit tokens per second.

The first option is underrated. For a small team or a developer who wants a coding assistant and a general assistant running side by side, the B60 Dual is a tidy solution.

B60 Dual vs Used RTX 3090s and Arc Pro B70

Option Memory Bandwidth Slots Software
Maxsun B60 Dual 2 x 24GB 456 GB/s per GPU One dual-slot card Intel, maturing
Two used RTX 3090s 2 x 24GB 936 GB/s per GPU Two triple-slot cards CUDA, everything
Arc Pro B70 32GB 608 GB/s One card Intel, maturing
Radeon AI PRO R9700 32GB 640 GB/s One card ROCm, broad support

Two used RTX 3090s are about twice as fast per GPU and run every tool through CUDA. Hardware Corner’s original comparison framed the B60 Dual against a used 3090 at about $800 in 2025; prices for used 3090s rose in 2026, which narrows that gap. The 3090s take far more space and power, though, and you would need a board and case that fit two large cards.

The B60 Dual’s real advantage is density: 48GB in one dual-slot space. That matters most when you want four cards, and 192GB, in a single workstation.

Scaling to 96GB, 192GB and Battlematrix

Intel’s Project Battlematrix is built around exactly this idea. Let’s Data Science reported that it targets up to eight GPUs and 192GB of VRAM in one system. Maxsun showed a workstation with four B60 Dual cards for 192GB, as TweakTown reported, and StorageReview previewed a 192GB Battlematrix configuration for on-premises AI.

At that scale, the card’s density pays off. No other affordable route puts 192GB of GPU memory into a single workstation. The trade-offs scale too: eight GPUs to manage, a platform with enough PCIe lanes and bifurcation, and a power supply able to feed 1,600W of cards.

StorageReview’s assessment of a single dual card is useful here: 48GB provides sufficient capacity for quantised models and covers a large share of popular open-source LLM applications.

The Single-Slot Liquid Version

Maxsun also announced a B60 Dual 48G Liquid, which Igor’s Lab covered in March 2026. It is a single-slot, liquid-cooled version of the same two-GPU design, aimed at high-density workstations.

In a chassis built for it, that could fit twice as many cards in the same space. It needs a liquid cooling loop or a system designed around one, so it is a product for integrators and specialist builds rather than a weekend upgrade.

Power, Cooling and Noise

At around 400W for the board, the card draws more than any single-GPU option at this price. Size the power supply accordingly: at least 850W for a single card in a normal system, and much more for multiple cards. Our power supply guide explains how.

The blower cooler exhausts heat out of the back of the case, which suits multi-card builds, but blowers are loud under sustained load. If the machine will sit near you, plan for the noise or put it in another room.

Display Outputs and Desktop Use

TechRadar noted that each GPU gets one DisplayPort and one HDMI output, which avoids confusing the operating system in multi-GPU workstations. You can use the card as your display adapter, though for an AI server you may prefer to run headless.

Gaming is not the point of this card, and dual-GPU gaming support is effectively dead in modern games. Treat any gaming as a single B60.

Cost, Resale and Risk

On paper, the value case is overwhelming. At $1,200 for 48GB, you pay about $25 per gigabyte. Two used RTX 3090s cost more than that in late 2026, and any new 32GB card costs more per gigabyte as well.

Three things narrow the gap. The first is your time. Bifurcation settings, Intel drivers and model splitting all take effort that a CUDA setup avoids. The second is resale value. Used NVIDIA cards sell easily and hold their price. A niche dual-GPU card from a smaller manufacturer has a thin second-hand market, so expect to lose more when you upgrade. The third is support. If one of the two GPUs fails, the whole card goes back for repair, and Maxsun’s warranty service varies by region.

None of these make the card a bad buy. They make it a buy for people who value capacity per dollar enough to accept some extra work and risk. If you plan to keep it for several years and use both GPUs every day, the numbers hold up well. If you expect to resell within a year, a used NVIDIA card is the safer store of value.

Who Should Buy What

If you are Start with Why Link
Wanting 48GB in one slot, or 192GB in one workstation Maxsun Arc Pro B60 Dual Unmatched VRAM density per dollar Check price
Wanting maximum speed and software support Two used RTX 3090s Twice the bandwidth per GPU, CUDA Check price
Wanting 32GB on a single GPU Radeon AI PRO R9700 No model splitting, broader tools Check price
Staying with Intel on one GPU Arc Pro B70 32GB, more bandwidth Check price

Frequently Asked Questions About the B60 Dual

Is the B60 Dual one GPU with 48GB?
No. It is two GPUs with 24GB each on one board. Models larger than 24GB must be split across both.

What does my motherboard need?
x8/x8 bifurcation on the PCIe slot. Without it, only one GPU appears.

How much does it cost?
About $1,200 from Maxsun at launch, and around €1,500 in Germany. Check current prices.

How fast is it on 70B models?
Expect single digits to low double digits of tokens per second at 4-bit, limited by 456 GB/s per GPU.

Does Ollama support it?
Support exists through llama.cpp backends but is less mature than on NVIDIA or AMD. Many owners run llama.cpp or Intel’s vLLM directly.

Is there a single-slot version?
Yes, a liquid-cooled B60 Dual 48G Liquid for dense workstation builds.

Setting Up the B60 Dual

Enable x8/x8 bifurcation on the target slot in your BIOS before installing the card. Boot into a recent Linux release with Intel’s current compute runtime, and confirm both GPUs appear.

Start simple: run a model that fits in 24GB on one GPU to confirm everything works. Then try two models, one per GPU. Only then attempt splitting a single large model, first with llama.cpp layer splitting, then with Intel’s vLLM branch for tensor parallelism.

Keep a note of which driver and software versions work, and update cautiously.

Verdict

Buy the Maxsun Arc Pro B60 Dual if you want the most VRAM per slot and per dollar, you have a motherboard with x8/x8 bifurcation, and you are comfortable with Intel’s tools. It is especially good for running two models side by side, or for building a dense 192GB workstation.

The honest caveat is that it is two modest GPUs, not one big one. Large dense models run slowly, Intel’s software is still maturing, and two used RTX 3090s are faster and simpler if you have the space.

Next step: check your motherboard manual for bifurcation support today. If it is there and your models fit the two-by-24GB pattern, price the B60 Dual against two used 3090s and decide whether density or speed matters more.

Sources and Further Reading

On this site: compare Intel options in our Arc Pro B70 review, plan multiple cards with the dual GPU build guide, see the AMD alternative in the R9700 review, and check memory needs in how much VRAM for AI.