independent research node · battlemage host cell
Bare-metal Battlemage validation by Sergio Barrientos. Direct hardware telemetry, exact kernel/oneAPI configurations, and transparent reproducibility rules for every published tok/s.
flagship testbed · 32gb gddr6
Dedicated Battlemage BMG-31 testbed on Linux xe kernel driver. 32 Xe-cores, 256 XMX engines, 608 GB/s memory bandwidth, and 32,656 MiB visible VRAM. Used for living LLM recipe quantization, speculative decoding studies, and power scaling.
distributed cluster · 64gb total vram
Two-card workstation topology configured for MiniMax H3 FL2VA text-to-video generation with PyTorch 2.13.0+xpu. Dedicated role splitting: UNet + VAE computation pinned to GPU 0, CLIP text encoding pinned to GPU 1.
hardware workbench
Dedicated AI inference and kernel validation host
150 W sustained research cap · 18 W idle draw · 608 GB/s bus
Primary workstation driving continuous vLLM XPU and llama.cpp SYCL campaigns. Operates at a configured 150W energy-counter cap for sustainable thermal stability without throughput loss.
xe kernel driver (PCI ID 8086:e20b)1.3.36+ · Compute runtime 24.52.32224.28Native DPC++ compilation, Level Zero, and PyTorch XPU
Native Level Zero backends · upstream kernel integration
Standardized build pipelines targeting Intel DPC++ compiler, Level Zero driver interfaces, and upstream PyTorch XPU backends with custom community patches.
icx / dpcpp 2024.2.1 with -fsycl -O3 flags-DGGML_SYCL_F16=ON -DGGML_SYCL_TARGET=INTELONEAPI_DEVICE_SELECTOR=level_zero:gpuCoordinates first: model, quant, engine, shape, power
C1 vs Cn separated · cold TTFT vs warm KV cache hits
Every rate on this hub is backed by an explicit coordinate set: model shape, quantization artifact, container digest, concurrency stream count, and measured energy-counter watts.
Independent research with zero vendor sponsorship
2 public cookbooks · 20 indexed skills · 6 dossiers
XeCores is an independent community project. All recipes, container patches, and telemetry are published openly on GitHub for complete third-party reproducibility.
intel-arc-pro-b70-inference-cookbookintel-arc-pro-b70-videogen-cookbooklive telemetry & shell
sergio@b70-lab:~$ xpu-smi discovery --dump 1,2,18,20,22
# Intel XPU System Management Interface — Target Device Topology
| Device ID | Device Name | PCI BDF | Visible VRAM | Memory Bandwidth | Max Power (TBP) |
|---|---|---|---|---|---|
| 0 | Intel(R) Arc(TM) Pro B70 GPU | 0000:03:00.0 | 32,656 MiB | 608 GB/s (GDDR6) | 230 W (Cap: 150 W) |
| 1 | Intel(R) Arc(TM) Pro B70 GPU | 0000:07:00.0 | 32,656 MiB | 608 GB/s (GDDR6) | 230 W (VideoGen Node) |
Driver version: xe 6.12.0-intel-xpu · Level Zero Loader: 1.3.36 · Device State: Ready
sergio@b70-lab:~$ docker run --rm -it --device /dev/dri:/dev/dri --ipc=host \
-v /models:/models -e ONEAPI_DEVICE_SELECTOR=level_zero:0 \
vllm/vllm-openai-xpu@sha256:f01e24f6... \
--model /models/Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16 \
--kv-cache-dtype fp8 --gpu-memory-utilization 0.88 --max-model-len 100000
INFO 08-27 18:20:01 [xpu_executor.py:104] Initialized XPU backend on Intel Arc Pro B70 (BMG-31)
INFO 08-27 18:20:04 [speculative.py:214] MTP draft tensor overlay loaded · 4 speculative tokens
INFO 08-27 18:20:07 [server.py:82] Uvicorn running on http://0.0.0.0:8000 (VRAM resident: 18.2 GiB)
sergio@b70-lab:~$ ./llama-bench -m /models/Muse-Glimmer-30B-Q4_K_M.gguf \
-p 512,2048 -n 128 -t 8 -ngl 99 -fa 1 -ctk q8_0 -ctv q4_1 -r 5
# llama.cpp SYCL DPC++ build (-DGGML_SYCL_F16=ON)
| Model | Backend | Threads | Test | Prompt Tokens | Gen Tokens | Rate (tok/s) |
|---|---|---|---|---|---|---|
| Muse-Glimmer-30B | SYCL (XPU0) | 8 | pp512 | 512 | 0 | 3,240.5 ± 21.3 |
| Muse-Glimmer-30B | SYCL (XPU0) | 8 | tg128 | 512 | 128 | 42.8 ± 0.4 |
| Muse-Glimmer-30B | SYCL (XPU0) | 8 | pp2048 | 2048 | 0 | 2,890.1 ± 18.7 |
| Muse-Glimmer-30B | SYCL (XPU0) | 8 | tg128 | 2048 | 128 | 41.6 ± 0.3 |