We made the world’s first 1 PFLOPS Flash-Attention on a single H100 GPU providing the fastest AI image generation possible

Stream HPC partnered with startup HippoML to develop the world's first 1 PFLOPS Flash-Attention implementation on a single NVIDIA H100 GPU. By modifying Flash Attention to support int8 and fp8 quantized types and leveraging Hopper-specific hardware features, the team achieved record-breaking performance for AI image generation. NVIDIA subsequently acquired HippoML following the milestone. Open article →