Intel Arc Pro B60 24GB for local LLMs
Quick answer
How fast is the Intel Arc Pro B60 24GB for local LLMs?
The fastest plain-decode result we have is 16.13 tok/s on Qwen3.8 27B (Q4_K_L (Swift-Qwen3.8-27B GGUF, 17.72 GiB), no MTP, llama.cpp b11100 (7ab4ee7b) SYCL) in 2x Intel Arc Pro B60 (Maxsun Dual 48G): Qwen3.8 27B Q4_K_L at 16.1 tok/s, 24.0 with MTP (SYCL). Speed depends heavily on the model, quant and engine, so compare like with like in the table. RigBase lists 5 measured results from 2 builds.
Last updated:
5 measured results from 2 builds, each linked to its source.
| Memory | 24 GB per card |
|---|---|
| Memory bandwidth | 456 GB/s (memory bandwidth as listed in the arc-llm-lab README) |
| Typical EU price | not tracked yet |
All builds with the Intel Arc Pro B60 24GB
| Build | GPU(s) | Memory GB (total) | Results |
|---|---|---|---|
| 2x Intel Arc Pro B60 (Maxsun Dual 48G): Qwen3.8 27B Q4_K_L at 16.1 tok/s, 24.0 with MTP (SYCL) | 2x Intel Arc Pro B60 24GB (Maxsun Arc Pro B60 Dual 48G Turbo, one board) | 48 | 2 |
| Intel Arc Pro B60 24GB + Ryzen 7 3700X: Qwen3.8 27B Q4_K_M at 16 tok/s (14.5 at 32K), 42.3 with DFlash2 | Intel Arc Pro B60 24GB | 24 | 3 |
All measured results
(MTP, DFlash, DSpark and similar draft-and-verify methods give much higher tok/s on predictable text and are not comparable 1:1 with plain decode)
| Model | Build | Ran on | Quant | Engine | tok/s | Details |
|---|---|---|---|---|---|---|
| Qwen3.8 27B | Intel Arc Pro B60 24GB + Ryzen 7 3700X: Qwen3.8 27B Q4_K_M at 16 tok/s (14.5 at 32K), 42.3 with DFlash2 | mixed / see build | UD-Q4_K_M + DFlash2 draft model (n=7), KV q8 draft | llama.cpp Vulkan + Intel flash-attention prefill kernel (arc-llm-lab build) | 42.3 | ⚡ spec-decode (DFlash)context: 64K KV window allocated, 8 eval requests📏 Eval-suite throughput: generated tokens / total wall-clock incl. prompt processing (the source's conservative headline; per-request decode 47-58, mean 53.3) |
| Qwen3.8 27B | 2x Intel Arc Pro B60 (Maxsun Dual 48G): Qwen3.8 27B Q4_K_L at 16.1 tok/s, 24.0 with MTP (SYCL) | mixed / see build | Q4_K_L (Swift-Qwen3.8-27B GGUF) + embedded Q8_0 MTP head, KV q8_0 | llama.cpp b11100 (7ab4ee7b) SYCL, --spec-type draft-mtp, n_max 3 | 24.0 | ⚡ spec-decode (MTP)context: 131K ctx allocated, batch 1📏 Submitted decode figure; the submitter sees 20-30 t/s across context lengths in chat |
| Qwen3.8 27B | 2x Intel Arc Pro B60 (Maxsun Dual 48G): Qwen3.8 27B Q4_K_L at 16.1 tok/s, 24.0 with MTP (SYCL) | mixed / see build | Q4_K_L (Swift-Qwen3.8-27B GGUF, 17.72 GiB), no MTP | llama.cpp b11100 (7ab4ee7b) SYCL | 16.13 | context: 131K ctx allocated📏 "Stock non-MTP decode" as stated in the submission notes |
| Qwen3.8 27B | Intel Arc Pro B60 24GB + Ryzen 7 3700X: Qwen3.8 27B Q4_K_M at 16 tok/s (14.5 at 32K), 42.3 with DFlash2 | mixed / see build | UD-Q4_K_M, KV f16, no drafter | llama.cpp Vulkan b11524 (86a283532), stock | 16.01 | context: empty context, llama-bench tg32📏 llama-bench tg32, single run (-r 1) |
| Qwen3.8 27B | Intel Arc Pro B60 24GB + Ryzen 7 3700X: Qwen3.8 27B Q4_K_M at 16 tok/s (14.5 at 32K), 42.3 with DFlash2 | mixed / see build | UD-Q4_K_M, KV f16, no drafter | llama.cpp Vulkan b11524 (86a283532), stock | 14.45 | context: 32K context depth, llama-bench tg32📏 Same llama-bench row at 32K depth (sustained figure); 15.14 at 16K |
All Intel Arc Pro B60 24GB builds: see also all GPUs · best LLMs by memory size · download the dataset.
FAQ
How fast is the Intel Arc Pro B60 24GB for local LLMs?
The fastest plain-decode result we have is 16.13 tok/s on Qwen3.8 27B (Q4_K_L (Swift-Qwen3.8-27B GGUF, 17.72 GiB), no MTP, llama.cpp b11100 (7ab4ee7b) SYCL) in 2x Intel Arc Pro B60 (Maxsun Dual 48G): Qwen3.8 27B Q4_K_L at 16.1 tok/s, 24.0 with MTP (SYCL). Speed depends heavily on the model, quant and engine, so compare like with like in the table. RigBase lists 5 measured results from 2 builds.
How much memory does the Intel Arc Pro B60 24GB have?
24 GB per card. Memory bandwidth: 456 GB/s (memory bandwidth as listed in the arc-llm-lab README).
Which models have been measured on the Intel Arc Pro B60 24GB?
Qwen3.8 27B.
Where do these numbers come from?
Every row links to a build page with its original public source (benchmark logs, READMEs, forum posts). RigBase does not re-run the benchmarks. Results marked ⚡ use speculative decoding and are not comparable 1:1 with plain decode.
Referral link: if you sign up with Google and add $10, Runpod gives you and RigBase bonus credit (terms). It never affects our data.