Arc Pro B60 Price & Used Prices: Which Local LLMs Can It Run?

• Memory24 GB · GDDR6
• Bandwidth456 GB/s
• FP16 / BF1698.5 TFLOPS
• INT8197 TOPS

Bus width

192-bit

Architecture

Xe2 Battlemage (BMG-G21)

Cores

2,560

Tensor cores

160

Ecosystem

oneAPI

Interface

PCIe 5.0 x8

Power

200 W

Slots

2

Type

Workstation

Launch date

Aug 1, 2025

MSRP

$599

Model	Size	Cards	Total price	Total VRAM	Speed t/s	Total power
Qwen3.8 Flash Next176B (6B active)	111.3 GB	8 cards	$7,488	192 GB	59.9	1,600 W
DeepSeek V4 Flash284B (13B active)	155.1 GB	8 cards	$7,488	192 GB	39.6	1,600 W

Full calculation rules are on the Methodology page.

1. 1

   The 48GB B60 Dual is a different card

   There is also a 48GB “B60 Dual”, which Maxsun and Sparkle make by putting two B60 chips on one board. This page covers the single-chip 24GB version, so check which one a listing is for. Each of the Dual's two GPUs runs on PCIe x8, so the motherboard must be able to split x16 into x8/x8 to see both, and many consumer boards lack this option. The Maxsun version draws 400W for the whole card through a single 12V-2x6 connector.

2. 2

   Check the BIOS on older workstations

   Under Linux, if the motherboard gives the card too small an address space (BAR), Intel's compute stack won't recognize it: display output and Vulkan work fine, but SYCL and oneAPI can't find the card. The B60 needs a 32GB BAR. One user with a Dell Precision 5820 hit this because the BIOS default was too small, and only fixed it by changing a hidden setting. Before putting the card in an older machine, confirm the BIOS can enable Resizable BAR or enlarge the MMIO space.

3. 3

   Pick software by operating system

   On Windows, Vulkan is the easiest route: llama.cpp has a Vulkan backend, and Ollama supports Intel GPUs through Vulkan by default (on Linux you need to install extra driver components). vLLM requires Linux; Intel's packaged version is called llm-scaler and requires Ubuntu 24.04. IPEX-LLM was archived this January, and Intel Extension for PyTorch stopped being maintained at the end of March, so don't follow older tutorials built on them.

4. 4

   Check the cooling design first

   There is no reference B60; every card comes from a board partner. The common dual-slot blower versions have their own fan (Sparkle's, for example, uses one 8-pin connector and draws 200W). There are also passive versions that rely on server airflow (such as ASRock's single-slot Passive), which will overheat in a regular case. Board partners also set the power limit anywhere between 120 and 200W, so two cards both sold as a B60 may draw different power at full load.

Spec	Arc Pro B60	Arc Pro B70	GeForce RTX 3090	RTX PRO 4000
Memory	24 GB	32 GB33% more	24 GBSame	24 GBSame
Bandwidth	456 GB/s	608 GB/s33% higher	936 GB/s105% higher	672 GB/s47% higher
BF16 / FP16 compute	99 TFLOPS	184 TFLOPS86% higher	142 TFLOPS44% higher	161 TFLOPS64% higher
INT8 compute	197 TOPS	367 TOPS86% higher	285 TOPS44% higher	323 TOPS64% higher
Qwen3.8 27B speed (2-bit)	33.1 t/s	43.5 t/s31% faster	64.8 t/s96% faster	47.8 t/s44% faster
Largest dense model (4-bit)	37B	51B38% more	37BSame	37BSame
Power	200 W	230 W15% higher	350 W75% higher	145 W28% lower
GPU interconnect	—	—	NVLink 112.5 GB/s	Not supported
Interface	PCIe 5.0 x8	PCIe 5.0 x16	PCIe 4.0 x16	PCIe 5.0 x16
Released	2025-08	2026-03	2020-09	2025-03
eBay median	$936Buy Arc Pro B60	$1,850Buy Arc Pro B70	$1,799Buy GeForce RTX 3090	$2,988Buy RTX PRO 4000

As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.