Builds / Intel Arc Pro B70 32 GB
Intel Arc Pro B70 32 GB (Core i7-12700KF, llama.cpp b12128 SYCL): gpt-oss 20B MXFP4 at 134.2 tok/s, Gemma 4 E4B Q4_K_M 97.3, Qwen3.5 9B Q6_K 66.5
Last updated:
- GPU(s)
- Intel Arc Pro B70 32 GB
Amazon · eBay (used) · Newegg - VRAM / unified memory (GB)
- 32
- CPU
- Intel Core i7-12700KF
Amazon · eBay (used) · Newegg - Motherboard
- not stated
- RAM
- not stated
- PSU
- not stated
- Other
- Ubuntu 26.04 LTS, llama.cpp b12128 SYCL (the source found Vulkan 1.5-5.3x slower on this card, and Vulkan lost the device on Gemma 4 26B); 8 fastest CPU cores pinned
- Reported total cost
- not stated
- Power
- 0.84 tokens per watt on gpt-oss 20B decode (driver sensor); about 51 W idle
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| gpt-oss 20B | MXFP4 | llama.cpp b12128 (SYCL) | 134.2 | 1,296 @ 4,096-token prompt est. TTFT @8k ≈ 6.3 s (calculated from measured prefill) | context: empty context; 110.3 at 16,384 depth📏 write 128 tokens, 5 repetitions (tg128), model fully in VRAM |
| Qwen3.5 9B | Q6_K | llama.cpp b12128 (SYCL) | 66.5 | 3,673 @ 4,096-token prompt est. TTFT @8k ≈ 2.2 s (calculated from measured prefill) | context: empty context📏 tg128, 5 repetitions |
| Gemma 4 E4B legacy | Q4_K_M | llama.cpp b12128 (SYCL) | 97.3 | 4,610 @ 4,096-token prompt est. TTFT @8k ≈ 1.8 s (calculated from measured prefill) | context: empty context📏 tg128, 5 repetitions |
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.
Vendor-run: GEARVN AI Lab (GEARVN is a Vietnamese PC retailer that sells all three cards), article published 5-6 Nov 2026, measured 6 Nov 2026. Same llama.cpp b12128, same six GGUF files (sha256 checked on all three machines), 8 fastest CPU cores pinned, other GPU services paused; re-running on the single slowest core changed speed by at most 3.5%, so the gaps are the GPUs. Each card ran its fastest backend. Multi-user, gpt-oss 20B on SYCL: 1 user 130.8, 4 users 244.1, 8 users 371.0, 16 users 283.3 tok/s total (slower at 16 than at 8; the source attributes it to this llama.cpp build, not necessarily the card). The article also gives B70 decode of 93.0 (Qwen3.6 35B-A3B, 22.1 GB file), 83.4 (Gemma 4 26B-A4B, 16.9 GB) and 21.9 (Qwen3.8 27B, 16.5 GB) vs R9700 124.7 / 115.1 / 32.7, but does not name those files' quant types, so RigBase does not list them as results.
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.
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
[](https://getrigbase.com/builds/intel-arc-pro-b70-32gb-core-i7-12700kf-sycl-gpt-oss-gemma4-qwen35-gearvn/)
HTML
<a href="https://getrigbase.com/builds/intel-arc-pro-b70-32gb-core-i7-12700kf-sycl-gpt-oss-gemma4-qwen35-gearvn/" rel="noopener"><img src="https://getrigbase.com/badge/intel-arc-pro-b70-32gb-core-i7-12700kf-sycl-gpt-oss-gemma4-qwen35-gearvn.svg" alt="Measured on RigBase" height="20"/></a>
Preview: 
· How to use
Quick answers: Can Intel Arc Pro B70 run Qwen3.5 9B? · Can Intel Arc Pro B70 run gpt-oss 20B?