More
Submit build

Builds / Intel Arc Pro B60 24GB (USB4 eGPU enclosure)

Intel Arc Pro B60 24 GB in a USB4 enclosure on an M1 Max MacBook (aarch64 Linux): Llama 2 7B Q4_0 at 20.1 tok/s, 13.0 on a newer llama.cpp Vulkan build

Last updated:

Source: github.com ↗📏 How measured: reported by the original source; RigBase re-checked the numbers at that source
Legacy models only. Every model measured on this rig now has a newer successor: Llama 2 7B → Muse Glimmer 30B, Llama 2 7B → Muse Glimmer 30B. The hardware numbers are still a useful reference.
GPU(s)
Intel Arc Pro B60 24GB (USB4 eGPU enclosure)
Amazon · eBay (used) · Newegg
VRAM / unified memory (GB)
24
CPU
Apple M1 Max (16-inch MacBook Pro, aarch64 Linux host)
Motherboard
MacBook Pro 16-inch (M1 Max), USB4
Amazon · eBay (used) · Newegg
RAM
not stated
Amazon · eBay (used) · Newegg
PSU
not stated
Other
Omarchy (aarch64 Linux), kernel 7.1.12, Mesa ANV (vulkan-intel 26.2.5), KHR_coopmat; release Vulkan-only aarch64 builds; llama-bench -p 512 -n 128 -r 3 -ngl 99, builds run minutes apart; no resizable BAR
Reported total cost
not stated
Power
not stated

Measured speed (generation tokens/sec)

ModelQuantEnginetok/sPrompt processing (tok/s)Details
Llama 2 7B legacyQ4_0llama.cpp Vulkan 65840ed (Mesa ANV, aarch64)20.051,117 @ 512-token prompt
est. TTFT @8k ≈ 7.3 s (calculated from measured prefill)
context: llama-bench tg128📏 tg128 20.05 ± 0.11, all layers on the card
Llama 2 7B legacyQ4_0llama.cpp Vulkan 29c720e (Mesa ANV, aarch64)13.01821 @ 512-token prompt
est. TTFT @8k ≈ 10.0 s (calculated from measured prefill)
context: llama-bench tg128📏 Same host and flags ~614 commits later: tg128 13.01 ± 0.13 (-35%), pp512 -27%; not reproduced on a PCIe-slot B60 (see the Ubuntu PCIe build), so the cause is unresolved

Prompt processing = prefill speed (how fast the prompt is read) as reported by the source, with the prompt length it was measured at; blank = the source does not report it. est. TTFT @8k is calculated from the measured prefill speed (8,192 ÷ prefill tok/s), shown only when prefill was measured on a prompt of 512+ tokens. It is not a measurement and ignores load time and caching. measured TTFT is the source's own time-to-first-token. Why prefill matters for coding assistants.

Reporter's own numbers in open llama.cpp issue #31557 (2 Nov 2026). The same B60 does about 61 tok/s in a normal PCIe slot on x86 (Intel's reply in the same issue), so this unsupported setup (eGPU over USB4 on an Arm host, no ReBAR) costs roughly two thirds of the speed even on the faster build. Intel notes it does not officially support Arc cards in eGPU enclosures or on aarch64 Linux.

Source: https://github.com/ggml-org/llama.cpp/issues/31557

Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.

Embed this result

Copy into a README or blog. Links back to this build. See badge docs (CC BY 4.0 attribution). The SVG does not invent speeds — numbers stay on the build page.

Markdown

[![Measured on RigBase](https://getrigbase.com/badge/intel-arc-pro-b60-24gb-usb4-egpu-m1-max-aarch64-llama2-7b-vulkan-regression.svg)](https://getrigbase.com/builds/intel-arc-pro-b60-24gb-usb4-egpu-m1-max-aarch64-llama2-7b-vulkan-regression/)

HTML

<a href="https://getrigbase.com/builds/intel-arc-pro-b60-24gb-usb4-egpu-m1-max-aarch64-llama2-7b-vulkan-regression/" rel="noopener"><img src="https://getrigbase.com/badge/intel-arc-pro-b60-24gb-usb4-egpu-m1-max-aarch64-llama2-7b-vulkan-regression.svg" alt="Measured on RigBase" height="20"/></a>

Preview: Measured on RigBase · How to use

Rent before you buy: try this model on a cloud GPU for a few dollars first: Runpod (referral link; sign up with Google to get the new-user bonus)