Builds / Intel Arc Pro B60 24GB
Intel Arc Pro B60 24 GB in a PCIe slot (Ubuntu 26.04, Mesa 26.0.3): Llama 2 7B Q4_0 at 61.0 tok/s, pp512 1,795 on llama.cpp Vulkan
Last updated:
- GPU(s)
- Intel Arc Pro B60 24GB
Amazon · eBay (used) · Newegg - VRAM / unified memory (GB)
- 24
- CPU
- not stated
Amazon · eBay (used) · Newegg - Motherboard
- not stated (card in a PCIe slot)
Amazon · eBay (used) · Newegg - RAM
- not stated
Amazon · eBay (used) · Newegg - PSU
- not stated
- Other
- Ubuntu 26.04 LTS, Linux 7.2.6, Intel xe kernel driver, Mesa 26.0.3; llama.cpp Vulkan at 65840ed and 29c720e; llama-bench -p 512 -n 128 -r 3 -ngl 99, each revision run twice in reversed order
- Reported total cost
- not stated
- Power
- not stated
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| Llama 2 7B legacy | Q4_0 | llama.cpp Vulkan 29c720e (Mesa ANV) | 61.03 | 1,795 @ 512-token prompt est. TTFT @8k ≈ 4.6 s (calculated from measured prefill) | context: llama-bench tg128📏 tg128 61.60 (run 1) and 61.03 (run 2) on 29c720e; 61.76 / 61.03 on the older 65840ed, i.e. no regression on this setup. The lower run is listed |
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.
Comparison measurements posted by an Intel engineer (Arturo-Intel) on llama.cpp issue #31557 (4 Nov 2026), answering an eGPU regression report. Same commits and flags as the report, but the card sits in a normal PCIe slot on x86 Ubuntu, and speed was unchanged between the two builds. Llama 2 7B Q4_0 is llama.cpp's standard comparison model, so this is a clean reference point for the B60, not a modern-model result.
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.