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

China’s tech scene is turning heads again as DeepSeek introduces a remarkable alternative to NVIDIA’s modified AI accelerators. Their latest creation, building on the Hopper H800s, boasts a performance that’s eight times more powerful than standard TFLOPS.

DeepSeek calls their new tool FlashMLA, which could be a game-changer for China’s AI sector. The company isn’t waiting around for others to catch up; instead, they are software-savvy, crafting innovative solutions that maximize the existing hardware. Their ability to extract immense power from NVIDIA’s “cut-down” Hopper H800 GPUs is a testament to their ingenuity. By tweaking memory use and carefully managing resources across tasks, they’ve made significant strides.

DeepSeek recently took to Twitter to showcase their advances during “OpenSource” week, a time when they’re sharing a treasure trove of their innovations with the world via Github. FlashMLA was introduced on day one, setting the stage for a week of revelations. It’s a decoding kernel made especially for NVIDIA Hopper GPUs, and it’s already making waves in the industry.

The performance stats are astonishing. DeepSeek reports achieving 580 TFLOPS for BF16 matrix operations on the Hopper H800, shattering the usual industry benchmarks. In terms of memory, FlashMLA has pushed the limits to 3000 GB/s—almost double the peak theoretical capacity of the H800. What’s more impressive is that this leap is accomplished purely through software, not by tweaking the hardware.

Visionary x AI also shared their excitement on Twitter, emphasizing how the FlashMLA achieves 580 TFLOPS, massively outperforming the industry norm of 73.5 TFLOPS, and how it manages memory speeds that outdo the H800’s 1681 GB/s peak.

The secret behind FlashMLA’s efficiency lies in its “low-rank key-value compression” technique, which breaks down data into smaller, more manageable bits. This process not only speeds up operations but also cuts memory use by 40-60%. Additionally, FlashMLA employs a block-based memory paging approach that adapts to the task at hand, ensuring models efficiently handle varying sequence lengths and boosting overall performance.

What DeepSeek has achieved is a reminder that AI computing isn’t tethered to just one factor; it’s a multifaceted field with room for creative breakthroughs like FlashMLA. Currently tailored for Hopper GPUs, we’re eagerly anticipating what it might achieve when applied to NVIDIA’s H100.