Intel Arc Pro B60 for LLMs in November 2026: speed, software and price per gigabyte
The Arc Pro B60 is a 24 GB, 456 GB/s, 200 W card with a 3-month median price of $650 in the US (gpuprix, 4 November 2026; the cheapest new listing that day was $1,000), which makes it cheaper per gigabyte than any new or used NVIDIA card in this article. It decodes at about half a used RTX 3090's speed, as its bandwidth predicts: 16.05 tok/s on Qwen3.8-27B Q4_K_M with stock llama.cpp Vulkan (arc-llm-lab, October 2026), against 33.5 tok/s for Qwen3.5 27B on a 3090. In November 2026 it runs through llama.cpp's SYCL and Vulkan builds and through vLLM's XPU images, including Intel's llm-scaler 0.26.0-b2; IPEX-LLM was archived on 28 January 2026.
| Card | Memory | Bus, bandwidth | Board power | Interface | Launch price | US, 4 Nov 2026: cheapest new / 3-month median |
|---|---|---|---|---|---|---|
| Arc Pro B50 | 16 GB GDDR6 | 128-bit, 224 GB/s | 70 W | PCIe 5.0 x8 | $349 MSRP | $639 / $563 |
| Arc Pro B60 | 24 GB GDDR6 | 192-bit, 456 GB/s | 200 W | PCIe 5.0 x8 | about $600 per GPU (Intel) | $1,000 / $650 |
| Maxsun Arc Pro B60 Dual 48G | 2 x 24 GB (two GPUs) | 2 x 456 GB/s | 400 W | x16 slot, needs x8/x8 bifurcation | $1,200 (Dec 2025) | not tracked |
| Arc Pro B70 | 32 GB GDDR6 ECC | 256-bit, 608 GB/s | 230 W | not listed | not listed | $2,059 / $1,300 |
Specs and prices: gpuprix pages for the B50, B60 and B70, last updated 4 November 2026, US dollars as listed by retailers. B60 die details (20 Xe cores, 197 INT8 TOPS), Intel's announced price and the Maxsun dual board's power, bifurcation requirement and price are from StorageReview's Battlematrix preview of 1 December 2025. The B60's 12-month low on gpuprix was $600.
How fast is the Arc Pro B60 for LLMs?
About as fast as 456 GB/s allows on dense models, and comfortable on mixture-of-experts models. arc-llm-lab's stock llama.cpp Vulkan run of Qwen3.8-27B at UD-Q4_K_M generated 16.05 tok/s. That file is 16.46 GB per Unsloth's listing, so the bandwidth ceiling is 456 / 16.46 = 27.7 tok/s and the card reached 58% of it, by my arithmetic. A 3090 on Qwen3.5 27B Q4_K lands at 33.5 tok/s (Hardware Corner, March 2026), almost exactly the bandwidth ratio away.
| Setup | Model, quant | Engine | Reported | Source | Date |
|---|---|---|---|---|---|
| 1x B60 | Qwen3.8-27B UD-Q4_K_M | llama.cpp Vulkan b11524, stock | 16.05 tok/s generation | arc-llm-lab | 2026-10 |
| 1x B60 | same, DFlash2 draft model | llama.cpp build 11526 (Intel flash-attention patches) | 42.3 tok/s (8-request eval), 47 to 58 per request | arc-llm-lab | 2026-10 |
| 1x B60 | same, 32k-token prompt | stock Vulkan vs patched build, f16 KV | 118 vs 420 tok/s prompt processing | arc-llm-lab | 2026-10 |
| 2x B60 | Gemma 4 26B-A4B | llama.cpp SYCL b9820, split-mode layer | about 42 tok/s | CrunchyMonkies | 2026-07-15 |
| 1x B60 | Gemma 4 26B-A4B | vLLM XPU 0.22.0 | about 5 tok/s single, 96.5 tok/s total at 64 streams | CrunchyMonkies | 2026-07-15 |
| 1x B60 | Gemma 4 E2B | llama.cpp | 65 tok/s (RTX 4080: 178) | CrunchyMonkies | 2026-07-15 |
| 1x B60 | gpt-oss-20b, INT4 | early LLM-Scaler (vLLM) | 49.22 tok/s at batch 1 | StorageReview | 2025-12-01 |
| 4x B60, TP 4 | gpt-oss-120b, INT4 | early LLM-Scaler | 16.28 tok/s at batch 1 | StorageReview | 2025-12-01 |
| 4x B60, TP 4 | Qwen3-VL-30B-A3B | vLLM 0.12 vs LLM-Scaler 1.2, 16 users | 61 vs 48 ms per output token; first token 2.1 vs 4.2 s | Embedded LLM | 2026-02-14 |
| B70 + B60, PCIe 4.0 x4 adapters | Qwen3.5-27B, dynamic FP8 | llm-scaler-vllm 0.14.0-b8.1 | 13.3 tok/s generation, 1,009.7 prompt | Ben Linville | 2026-04-16 |
Single-stream unless stated. arc-llm-lab ran Windows 11 with Intel's Arc Pro driver 32.0.101.8976; its DFlash2 figures use speculative decoding with a draft model, so they depend on how predictable the text is. Ben Linville's cards sat on PCIe 4.0 x4 M.2 adapters, which he blames for slow multi-GPU generation, so his numbers understate a card in a full slot. StorageReview called its software "very early in development".
Three things stand out. Speculative decoding lifts single-user speed on a dense model 2.6 to 3.3 times (42.3 and 53.3 against 16.05 tok/s), because the card has spare compute while it waits on memory. Prompt processing depends heavily on the build: arc-llm-lab went from 118 to 420 tok/s at 32k by switching to a build with Intel's flash-attention work. And vLLM on XPU is a throughput engine: CrunchyMonkies measured about 5 tok/s for one stream and 96.5 tok/s across 64, and recommends llama.cpp for chat and vLLM for batch or API serving.
What is the software state in November 2026?
Usable through two routes, llama.cpp and vLLM, with more tuning than an NVIDIA card needs. IPEX-LLM, the library many 2024 guides used, is gone: Intel archived intel/ipex-llm on 28 January 2026, and its README says it gets no further development, bug fixes or patches and "has known security issues".
| Route | State on 7 November 2026 | Source |
|---|---|---|
| llama.cpp, SYCL | Release b12189 (7 November) ships SYCL builds for Linux and Windows. Docs list only the Arc B580 as a verified B-series card, recommend oneAPI 2026.1 (which took an Arc B570 from 434 to 1,331 tok/s prompt processing) and add an XMX flash-attention path for q4_0/q8_0 cache on Xe2. Row split is not supported. | SYCL docs, releases |
| llama.cpp, Vulkan | Same release ships Vulkan builds. Faster than SYCL for Qwen3.8-27B in arc-llm-lab's tests (its SYCL log was not kept); half SYCL's speed for Gemma 4 31B on a B70 (8.1 against 16.3 tok/s at UD-Q4_K_XL) in Ben Linville's, where SYCL broke on Qwen3.5's hybrid layers. | arc-llm-lab; Ben Linville, April 2026 |
| vLLM upstream, XPU | Docker image vllm/vllm-openai-xpu since v0.26.0, XPU wheel and image with v0.33.0 (2 November). Needs Python 3.12 and vllm-xpu-kernels. Docs say "Intel ARC GPU" without naming the B60. | vLLM install docs, releases |
| Intel llm-scaler | intel/llm-scaler-vllm:0.26.0-b2 (September 2026) lists the B60 and B70 (not the B50), FP16, dynamic FP8 and INT4, MXFP4 and pre-quantized FP8, with tensor, pipeline and data parallelism. | intel/llm-scaler |
| IPEX-LLM | Archived 28 January 2026; do not start new work on it. | intel/ipex-llm |
The rough edges reported in 2026: a q8_0 KV cache on SYCL kept only about 38% of the f16 decode speed in CrunchyMonkies' tests, against about 80% on an RTX 4080, so the usual memory-saving trick costs more here. Row split segfaulted. SR-IOV, which lets one card serve several virtual machines, arrived through a firmware update rather than at launch. StorageReview's December 2025 preview found that only MXFP4 gpt-oss models worked and only tensor parallelism ran; llm-scaler's September 2026 notes now list FP8, INT4 and MXFP4, so that part has moved on.
A minimal llama.cpp start on Linux with the release's Vulkan build looks the same as on any GPU. A model that fits is the main requirement; the memory calculator does the weights-plus-cache arithmetic.
./llama-server -m Qwen3.8-27B-UD-Q4_K_M.gguf -ngl 99 -c 32768 -fa on \
--host 127.0.0.1 --port 8080Is the dual-B60 48 GB board worth it?
As a way to fit 48 GB into one slot, yes; as a faster card, no. The Maxsun board carries two separate B60 GPUs, each on eight PCIe lanes, so the motherboard must split its x16 slot x8/x8, and the system sees two 24 GB devices rather than one 48 GB pool. It draws 400 W and cost $1,200 when StorageReview tested four of them in a Supermicro server.
The reports agree that a second B60 adds room, not speed, for one user. CrunchyMonkies ran Gemma 4 26B-A4B across two B60s with layer split at about 42 tok/s, about the same as one card. StorageReview found that at low batch sizes, using the fewest GPUs that hold the model gave the best per-user speed: gpt-oss-20b ran at 49.22 tok/s on one GPU and 22.83 tok/s spread over eight. Concurrency is where several cards pay off: eight B60s served gpt-oss-20b at 511.99 tok/s total at batch 16.
Four dual boards make Intel's 192 GB "Battlematrix". For a single model that needs 100 GB or more, compare that with a single 128 GB unified-memory machine: the DGX Spark against Strix Halo comparison covers the two 128 GB mini-PC options, and GPUwerk's models page lists what fits in 128 GB.
Arc Pro B60 against RTX 3090 and RTX 5090: which is better value?
Per gigabyte, the B60 wins easily; per tok/s, it wins only near its median price. The table uses reported single-stream llama.cpp speeds on a dense 27B model at 4-bit and the US prices gpuprix showed on 4 November 2026.
| Card | VRAM, bandwidth | Dense 27B Q4, tok/s | Price used | tok/s per $1,000 | GB per $1,000 |
|---|---|---|---|---|---|
| Arc Pro B60 | 24 GB, 456 GB/s | 16.05 (Qwen3.8-27B) | $650 median / $1,000 cheapest new | 24.7 / 16.1 | 36.9 / 24.0 |
| RTX 3090 (renewed) | 24 GB, 936 GB/s | 33.5 (Qwen3.5 27B) | $1,795 cheapest used | 18.7 | 13.4 |
| RTX 5090 | 32 GB, 1,792 GB/s | 58.8 (Qwen3.5 27B) | $3,330 median / $7,550 cheapest in stock | 17.7 / 7.8 | 9.6 / 4.2 |
| Arc Pro B70 | 32 GB, 608 GB/s | no full-slot figure | $1,300 median / $2,059 cheapest new | n/a | 24.6 / 15.5 |
My arithmetic, for example 16.05 / 0.65 = 24.7 and 24 / 0.65 = 36.9. Speeds: arc-llm-lab (B60, llama.cpp Vulkan b11524) and Hardware Corner's 3090 and 5090 pages (Q4_K at 4k context, March 2026). Qwen3.8-27B and Qwen3.5 27B are different releases of the same size class; Hardware Corner has no Qwen3.8 figure. Prices: gpuprix pages for the 3090 and 5090, 4 November 2026.
At $650 the B60 gives about a third more tok/s per dollar than a renewed 3090 and nearly three times the memory per dollar. At $1,000 the 3090 is twice as fast for 1.8 times the price, and it runs every CUDA tool, which the B60 does not. The 5090 is the fastest card here by a wide margin; at in-stock prices it is the worst value on both measures.
Should you buy one?
I would buy a B60 at or near its $650 median for a 24 GB llama.cpp box that one or two people use, or several of them for a vLLM server where concurrency matters more than single-stream speed. It is also the cheapest way to 48 GB in one slot, if the motherboard bifurcates.
I would not buy one for a single fast chat assistant on dense models (a used 3090 is twice as quick, as the sibling article on the used RTX 3090 shows), for CUDA-only tooling such as TensorRT-LLM, for NVIDIA-specific formats such as NVFP4, or at $1,000. The B50, at 224 GB/s, has half the bandwidth of an RTX 5060 Ti 16GB, so its 16 GB is for small models; the sibling guide to the best LLMs for 16 GB of VRAM lists what fits. For the AMD alternative at 32 GB, see the sibling comparison of the Radeon AI PRO R9700 and RTX 5090, and for how to weigh capacity against bandwidth in general, choosing a GPU for open-source LLMs.
If the plan is several B60s to hold a 100B-class model, try that model first on hardware that already holds it: GPUwerk rents a 128 GB DGX Spark at $0.79 an hour, and an hour with your own prompts tells you whether the model is worth a four-card build.
FAQ
Is the Intel Arc Pro B60 good for running LLMs?
It is a good capacity card and a slow single-stream card. It has 24 GB at 456 GB/s, about half an RTX 3090's bandwidth, and arc-llm-lab measured 16.05 tok/s on Qwen3.8-27B Q4_K_M with stock llama.cpp Vulkan in October 2026. Mixture-of-experts models are much quicker: CrunchyMonkies reports about 42 tok/s on Gemma 4 26B-A4B across two B60s in July 2026.
How much does the Arc Pro B60 cost?
gpuprix's US page, updated 4 November 2026, showed a 3-month median of $650 and a 12-month low of $600, with the cheapest new listing at $1,000 (an ASRock B60 Creator on Amazon). Intel announced the B60 at about $600 per GPU, per StorageReview. The Maxsun dual-B60 48 GB board cost $1,200 when StorageReview tested it in December 2025.
Does vLLM support the Intel Arc Pro B60?
Yes, two ways. Upstream vLLM publishes an XPU Docker image from v0.26.0 and an XPU wheel with v0.33.0, which need Python 3.12 and the vllm-xpu-kernels package. Intel's llm-scaler builds, at intel/llm-scaler-vllm:0.26.0-b2 in September 2026, list the B60 and B70 with FP8, INT4 and MXFP4 and tensor, pipeline and data parallelism. CrunchyMonkies found vLLM XPU slow single-stream but strong under concurrency.
Is IPEX-LLM still maintained?
No. Intel archived the intel/ipex-llm repository on 28 January 2026. Its README says Intel will no longer provide development, maintenance, bug fixes or new releases, no longer accepts patches, and that the project has known security issues. For Arc Pro cards, use llama.cpp's SYCL or Vulkan builds or a vLLM XPU or llm-scaler image instead.
Arc Pro B60 or used RTX 3090 for local LLMs?
It depends on the price you pay. At the B60's $650 median it gives 24.7 tok/s per $1,000 on a dense 27B model against 18.7 for a $1,795 renewed RTX 3090, by my calculation from arc-llm-lab and Hardware Corner figures. At the B60's $1,000 listing price the 3090 is twice as fast for 1.8 times the money and runs every CUDA tool.
Sources
Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.
- Arc Pro B60 price in United States, gpuprix, 2026-11-04
- Arc Pro B50 price in United States, gpuprix, 2026-11-04
- Arc Pro B70 price in United States, gpuprix, 2026-11-04
- Intel Arc Pro B60 Battlematrix preview: 192GB of VRAM for on-premise AI, StorageReview, 2025-12-01
- arc-llm-lab: Run LLMs on Intel Arc with llama.cpp Vulkan, edagher92-coder on GitHub, 2026-10-10
- The Intel Arc Pro B60 for self-hosted LLMs: 24 GB, SR-IOV, and a software stack that's finally catching up, CrunchyMonkies, 2026-07-15
- Intel Arc Pro B60 + B70 LLM benchmarks: most quants, every backend, Ben Linville (Substack), 2026-04-16
- Benchmarking LLM inference on Intel Arc Pro B60: a comparative analysis of vLLM, Embedded LLM, 2026-02-14
- intel/llm-scaler, Intel on GitHub, 2026-09
- intel/ipex-llm (archived), Intel on GitHub, 2026-01-28
- llama.cpp SYCL backend documentation, ggml-org/llama.cpp on GitHub
- llama.cpp releases, ggml-org/llama.cpp on GitHub, 2026-11-07
- GPU installation (ROCm, XPU and others), vLLM documentation
- vLLM releases, vllm-project/vllm on GitHub, 2026-11-02
- RTX 3090 local LLM benchmarks, Hardware Corner, 2026-03
- RTX 5090 local LLM benchmarks, Hardware Corner, 2026-03
- GeForce RTX 3090 price in United States, gpuprix, 2026-11-04
- GeForce RTX 5090 price in United States, gpuprix, 2026-11-04
- Qwen3.8-27B-GGUF files, Unsloth on Hugging Face