Dual Intel Arc Pro B70s (Battlemage), 42 benchmarks and counting!
LLM inference, 64K context scaling, power telemetry, every model we could load.
PSA: There's a build-flag trap that's been suppressing public B70 prefill by ~50%.
Take a look at the raw data, sweep configs, and the fix:
github.com/PMZFX/intel-arc-p…
I ran the same 9B model on 7 GPUs from $200 to $1,300.
RTX 3080 Ti: 113 t/s ($425)
RTX 3090: 113 t/s ($1,300)
Arc Pro B70: 54 t/s ($949)
RTX 3060: 51 t/s ($250)
RTX 4060: 43 t/s ($250)
BC-250: 26 t/s ($200)
M4 Mac Mini: 21 t/s ($499)
The 3080 Ti matches the 3090 at 1/3 the price.
But both are capped at models that fit in VRAM.
More VRAM = bigger models = smarter AI.
The B70 runs a 35B model on one card. Two cards run 80B.
PSA for Intel Arc / Battlemage multi-GPU users: Running dual-GPU inference (tensor parallel) on Intel dGPUs can eat your entire system RAM, even if the model fits entirely in VRAM.
We hit this running Qwen3.5-27B across two Arc Pro B70s (32GB each, 64GB total VRAM). Four OOM crashes in 25 minutes. 64GB system RAM, 100% consumed.
We all know that NVIDIA is the king of AI right now. CUDA is unmatched. But to get 32 GB of VRAM there, you're either spending $4,000+ on a 5090 or hunting for old, used, power-hungry hardware.
AMD has their offering around $1,350+. Solid, but still a big ask for a local AI setup.
Then Intel comes along with the Arc Pro B70. 32 GB GDDR6 ECC. 608 GB/s bandwidth. $950.
I knew it was a risk. The ecosystem is immature, the software stack has rough edges, and driver support is still catching up. But I decided to take the chance, and more importantly, share everything I learn so others can make an informed decision.
So I bought two of them.
Here's what I've found so far running local LLMs:
The good:
- 32 GB VRAM fits Qwen3.5-27B at Q8 with ~20K context, or Q4_K_M with ~63K context
- Q4 inference runs at 20 t/s on Qwen 3.5 27B. usable
- After finding and fixing a Q8_0 kernel issue in llama.cpp, Q8 now runs at 15 t/s (it was 5 before)
- The hardware is legitimately capable
The rough:
- SYCL JIT compilation on first run takes minutes. Minutes.
- Some code paths segfault on Battlemage that work fine on older Arc
- vLLM's XPU backend crashes during initialization on B70
- Intel's own IPEX-LLM doesn't even recognize this GPU's PCI device ID
- You will be debugging things that CUDA users never have to think about
Is it for everyone? No. But if you want 32 GB of VRAM for under $1,000 and you're willing to either figure things out or wait for updates, it looks to be a decent value. I'll be posting benchmarks, fixes, and honest takes as I go.
@IntelGraphics #IntelArc #Battlemage #LocalAI #LLM
5 tokens/sec on a brand new, just released GPU. I was hoping for more with the Intel Arc Pro B70. Luckily, it doesn't seem like it was a hardware issue!
Qwen 3.5 and Gemma 4 at Q4 got me about 20 t/s, so I figured something was wrong specifically with Q8. I tried different drivers, Vulkan, env vars. Nothing helped at first.
With @AnthropicAI Claude's help, I dug into Intel's IPEX-LLM and found their Q8 kernels actually hit 61% bandwidth. But their stack doesn't support modern models and crashes on this brand new GPU.
So we binary-patched their closed-source code to prove the fix was possible, reverse-engineered the optimization, and built it for upstream llama.cpp.
~200 lines. 3.1x speedup on Q8_0! Works with every model.
PR: github.com/ggml-org/llama.cp…
If you have an Arc A770/A750, test this and let me know if it helps!
@IntelGraphics @intel #llamacpp #IntelArc #Battlemage #LocalAI
