DeepSeek-V4.1 Flash: Achieving 3.8x Faster Throughput on PCIe

Developers on the Dolphin Network implemented various inference optimizations for the DeepSeek-V4.1 Flash model running across four NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs connected via PCIe. By utilizing techniques such as a paired transport mechanism for tensor parallelism, bounded prefill processing, and fused expert kernels, they achieved 2.53 to 3.83 times the baseline output-token throughput in cached-prompt benchmarks. Open article →