According to @deepseek_ai, the newly introduced DeepGEMM library supports both dense and MoE GEMMs, achieving up to 1350+ FP8 TFLOPS on Hopper GPUs. This advancement is significant for V3/R1 training and inference, offering traders insights into potential hardware investments and performance efficiencies in AI-driven trading algorithms. The library is designed to be lightweight with no heavy dependencies, which is crucial for optimizing trading software infrastructure. Furthermore, its fully Just-In-Time compiled nature enhances performance, relevant for high-frequency trading applications.

According to DeepSeek, the new FlashMLA kernel for Hopper GPUs, optimized for variable-length sequences, is now in production, offering BF16 support and achieving 3000 GB/s memory-bound and 580 TFLOPS, which can significantly boost computational efficiency in crypto trading algorithms.