AI Infrastructure at Periodic (periodic.com)

Periodic has announced significant advancements in AI infrastructure that empower the efficient training of specialized models, outperforming leading models like GPT-6 Astra and Claude Fable 5.1 in X-ray diffraction evaluations. Utilizing a peak of 1,300 H200 GPUs, the system enhances training throughput, inference generation speed, and scientific tool execution. Key innovations, such as improved GPU memory efficiency and optimized asynchronous operations, allow Periodic to maintain over 95% cluster utilization. These enhancements facilitate rapid exploration of novel superconductors and magnets, thus accelerating the research cycle from concept to experimental validation. Central to Periodic's success are optimizations derived from open-source frameworks like Megatron and SGLang, which have been modified to accommodate scientific reinforcement learning (RL) workloads. Innovations such as Delta router replay and prefill-decode disaggregation have notably increased training throughput by 4.1x and reduced inference times to 25 tokens per second for large models, showcasing the high scalability and efficiency of their infrastructure. By focusing on data parallelism and intelligent workload balancing, Periodic not only enhances research productivity but also reinforces its commitment to contributing improvements back to the open-source community.