New local LLM value king – MaxSun Arc Pro B60 Dual with 48GB VRAM for $1200

We recently discussed the upcoming single-GPU Intel Arc Pro B60 with 24GB of VRAM and its potential to shake up the local LLM hardware market. Now, its bigger brother is set to arrive. Reports indicate the MaxSun Arc Pro B60 Dual, featuring two GPUs on a single board for a total of 48GB of VRAM, will start shipping for around $1200. This development introduces a new and compelling option for enthusiasts focused on VRAM capacity.

The New VRAM Heavyweight: Specs and Price

The most important number for any local LLM enthusiast is VRAM, and the MaxSun Arc Pro B60 Dual delivers a massive 48GB of GDDR6 memory. This is achieved by placing two separate BMG-G21 GPUs on one circuit board. Each GPU has its own 24GB of VRAM connected via a 192-bit interface, resulting in a memory bandwidth of 456 GB/s for each processor. At a reported price of $1200, this card offers 48GB of VRAM at a price point previously unheard of outside the used market.

How It Stacks Up: VRAM vs. Bandwidth Trade-offs

For years, the undisputed value king for local LLM has been the used NVIDIA RTX 3090, offering 24GB of GDDR6X VRAM for about $800. The Arc Pro B60 Dual challenges this by providing double the memory for only a 50% price increase. This puts its cost at just $25 per gigabyte of VRAM, compared to the RTX 3090’s roughly $33 per gigabyte.

However, the trade-off is memory bandwidth. An RTX 3090 delivers a powerful 936 GB/s of bandwidth, more than double the 456 GB/s of a single Arc GPU on the B60 Dual. This means that for models that fit within 24GB, the RTX 3090 will be significantly faster at generating tokens. The Arc card’s value isn’t in raw speed but in its ability to load models that are simply too large for a 24GB card. Other high-end cards like the RTX 4090 or RTX 5090 offer even more bandwidth but at a much higher cost and with only 24GB or 32GB of VRAM, respectively. The leaked specs for the upcoming RTX 5080 Super and 5070 Ti Super suggest they will also stick to the 24GB VRAM size, reinforcing the B60 Dual’s unique position in the market for VRAM capacity.

The 96GB VRAM Dream: A Dual Card Setup for Massive Models

The most interesting possibility this card presents is a dual-card configuration. Installing two MaxSun Arc Pro B60 Dual cards in a system would provide a massive 96GB of VRAM for approximately $2400. This amount of memory enters the territory of elite workstation cards like the NVIDIA RTX Pro 6000, which costs between $8000 and $9000 for its 96GB of VRAM.

To put this into perspective, we recently tested the new GPT-OSS 120B MoE model on an RTX Pro 6000 Blackwell. With flash attention, it achieved an impressive 83 tokens per second for generation while using 67GB of VRAM. A dual Arc B60 setup could easily accommodate this model. Based on the much lower memory bandwidth, we can do some simple math. Even with a theoretical performance reduction of 75-80% compared to the RTX 6000, this dual Arc setup could still produce 15 to 20 tokens per second. For a 120B parameter model running locally for under $2500, that is a remarkable capability.

Conclusion: Is the Arc Pro B60 Dual a Value King?

The MaxSun Arc Pro B60 Dual is not a straightforward replacement for the RTX 3090. If your primary goal is the fastest possible speed for models up to 30 billion parameters, the higher bandwidth of NVIDIA cards remains the better choice. However, if your ambition is to run larger models like 65B and 70B models with long context windows, or to experiment with 120B MoE models that are impossible to load on 24GB cards, this GPU may become the new value king. Its value lies in pure capacity, opening the door to running truly massive models locally without spending a fortune on enterprise hardware. As MoE architectures become more popular, which often trade higher VRAM requirements for faster inference on capable hardware, the B60 Dual’s place in the market seems secure.

Trending Stories