
Intel has launched a significant software update for its Arc Pro “Project Battlematrix”initiative, introducing the LLM Scaler v1.0 with notable enhancements.
Intel Arc Pro GPUs Receive Important Software Upgrade with LLM Scaler v1.0
During Computex 2025, Intel introduced Project Battlematrix alongside its new Arc Pro GPUs. This initiative aims to provide a comprehensive solution for inference workstation platforms utilizing multiple Arc Pro GPUs. In its product roadmap, Intel committed to delivering the first container deployment with features such as vLLM staging and basic telemetry support in Q3, which has now materialized with the release of LLM Scaler v1.0.

The LLM Scaler container v1.0 incorporates a comprehensive set of features and optimizations:
- vLLM Enhancements:
- Up to 1.8x performance improvement for 40K sequence length on 32B KPI models and 4.2x for 70B KPI models
- Approximately 10% increase in output throughput for 8B-32B KPI models relative to the previous version
- By-layer online quantization for reduced GPU memory requirements
- Experimental support for pipeline parallelism (PP) in vLLM
- Experimental features including torch.compile and speculative decoding
- Support for embedding and rerank models
- Enhanced multi-modal model compatibility
- Automatic detection of maximum input length
- Data parallelism functionality
- OneCCL Benchmark Tool Activation
- XPU Manager Features:
- GPU Power Monitoring
- GPU Firmware Updates
- GPU Diagnostics
- GPU Memory Bandwidth Analysis
Intel emphasizes that the updated software stack is designed for user-friendliness and adheres to industry benchmarks. Tailored for Linux environments, this container boasts the potential for up to 80% performance gains through multi-GPU scaling and PCIe P2P data transfers. Additionally, it offers enterprise-grade reliability and management capabilities, including ECC, SR-IOV, telemetry, and remote firmware updates.
According to Intel’s roadmap, this release will soon be succeeded by a more robust container launch later this quarter, which will focus on enhanced performance and vLLM serving. Lastly, a comprehensive feature set is scheduled for rollout in Q4.
Leave a Reply