Watch Desk posted an update
Stream HPC and startup HippoML say they built a Flash Attention implementation that surpassed 1 PFLOPS on a single NVIDIA H100 GPU, aimed at speeding up AI image generation.
Why it mattersThe team adapted the method for int8 and fp8 precision and used features of NVIDIA’s Hopper and Lovelace architectures, according to Stream HPC. The company also says NVIDIA acquired HippoML after the work. A striking performance figure, though the supplied account does not give benchmark conditions for readers to compare.
Discuss: How much should a headline-grabbing AI compute milestone matter without a clearer account of how it was measured?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.