Intel: Accelerating LLM Inference on Intel Data Center GPUs using BigDL-LLM (ID 816988)

Vendor article (Dec 2024). Methodology fully disclosed: batch 1, greedy, 1024 input / 128 output tokens, 'next token latency' on INT4 and FP16 Llama-2-7B/13B + Mistral-7B. BigDL-LLM 2.5.0b20240111 (INT4) / 2.5.0b20240123 (FP16), PyTorch 2.1.0a0, IPEX 2.1.10+xpu, oneAPI 2024.0, Ubuntu 22.04.2, GPU driver 775_20, Xeon 8480+ host. Flex 170 on 2S 8480+ + 256GB DDR5-4800. Published values live in figures (PNG) not tables - so methodology is exact but the absolute tok/s must be read off the chart; only the RELATIVE FP16 vs INT4 ordering and the ~35% self-speculative-decoding latency improvement are stated in text.