China Advances in AI Without “Cutting-Edge” Accelerators: DeepSeek’s “FlashMLA” Project Boosts TFLOPS Power 8x with NVIDIA H800 GPUs

China appears to be taking bold strides in overcoming limitations imposed by reduced-performance NVIDIA AI accelerators. Thanks to DeepSeek’s groundbreaking FlashMLA technology, they’ve achieved a significant boost, delivering a performance eight times higher than expected with the Hopper H800s AI accelerators.

### Elevating China’s AI Game: The Role of DeepSeek’s FlashMLA with NVIDIA’s Hopper GPUs

China is forging its path in hardware capabilities, with local companies like DeepSeek leading the charge by leveraging software innovations to maximize existing equipment. DeepSeek’s recent advancements are particularly noteworthy as they have managed to extract remarkably high performance from the pared-down NVIDIA Hopper H800 GPUs. The company has ingeniously managed memory expenditure and resource utilization across various inference requests to achieve this feat.

To offer a bit more context, DeepSeek has dedicated an “OpenSource” week to unveiling technologies and tools that will soon be easily accessible via Github repositories. The initiative kicked off with a bang with the introduction of FlashMLA, an optimized decoding kernel specifically designed for NVIDIA’s Hopper GPUs. Before delving into the mechanics, let’s explore the fascinating enhancements it brings to the table.

According to DeepSeek, they’ve achieved an impressive 580 TFLOPS for BF16 matrix operations on the Hopper H800. This output is nearly eight times what’s generally deemed the industry norm. FlashMLA doesn’t stop there; it also boasts a memory bandwidth peaking at 3000 GB/s, almost twice the theoretical maximum of the H800. All these gains have been realized through smart coding innovations rather than hardware upgrades.

Using “low-rank key-value compression,” FlashMLA essentially breaks down data into smaller, more manageable portions, speeding up processing and cutting memory use by 40%-60%. The technology also incorporates a dynamic block-based paging system that adjusts memory allocation based on task complexity instead of adopting a one-size-fits-all approach. This adaptability significantly boosts the processing capacity of models dealing with variable-length sequences.

The advancements DeepSeek brings with FlashMLA showcase the multifaceted nature of AI computing. It’s not confined to a single dimension, and this tool is a testament to that diversity. For now, FlashMLA is specific to Hopper GPUs, but there’s growing anticipation around what the H100 could achieve with this innovative technology in the future.