China Advances in AI Without “Cutting-Edge” Accelerators; DeepSeek’s Latest “FlashMLA” Project Achieves 8x TFLOPS Boost Using NVIDIA H800 GPUs

China is making strategic strides in bypassing the limitations set by NVIDIA’s downgraded AI accelerators, thanks to DeepSeek’s innovative FlashMLA project that dramatically boosts TFLOPS using the Hopper H800s AI accelerators.

Revolutionizing AI Performance: DeepSeek’s Advancements in Hopper GPUs

China is proving that it won’t be held back when it comes to scaling hardware capabilities. By leveraging clever software strategies, companies like DeepSeek are finding ways to maximize the potential of existing technology. The latest breakthrough from DeepSeek is particularly notable. The company has successfully extracted enhanced performance from NVIDIA’s limited Hopper H800 GPUs by ingeniously optimizing memory use and managing resources across AI inference tasks.

In a recent tweet, DeepSeek announced the kickoff of their OpenSource Week initiative, proudly showcasing FlashMLA—a specially crafted "decoding kernel" for NVIDIA’s Hopper GPUs. This initiative aims to democratize access to groundbreaking tools through GitHub repositories, and Day 1 has certainly made an impact. Let’s explore the innovative strides FlashMLA has introduced to the industry.

DeepSeek reports achieving a remarkable 580 TFLOPS for BF16 matrix calculations on the Hopper H800, which exponentially surpasses the typical industry standard. Additionally, through savvy memory management, FlashMLA reaches an astounding 3000 GB/s memory bandwidth—nearly double the H800’s supposed maximum. All these gains, mind you, are realized solely through software ingenuity, without any hardware modifications.

Further claims by DeepSeek, shared on social media by enthusiasts, highlight the almost surreal processing speed and memory efficiency achieved with these GPUs, making it clear just how game-changing FlashMLA could be.

To achieve these results, DeepSeek’s FlashMLA employs a technique known as "low-rank key-value compression." In simpler terms, this method breaks data into smaller, more manageable pieces, facilitating swifter processing and reducing memory demand by up to 60%. Moreover, it introduces a block-based memory allocation system that adjusts according to task demands, ensuring efficient handling of tasks with different lengths—and thus boosting processing efficiency overall.

This advancement from DeepSeek illustrates that AI computing isn’t propelled by a single factor; instead, it thrives on a complex web of diverse strategies, as shown with FlashMLA. For now, the tool is tuned specifically for Hopper GPUs, but one can’t help but wonder about the possibilities once it’s applied to newer models like the H100.