🖥️ One GPU cannot hold GLM-5.3, so Nvidia split the 753-billion-parameter model across two machines

Nvidia's new guide runs Z.ai's GLM-5.3, quantized to NVFP4, on two DGX Stations linked by ConnectX-8. The checkpoint is about 464 GB. One GB300 GPU has 252 GB of memory, so each machine keeps about 218 GiB of the weights and they answer through one API. On Nvidia's coding test, output peaked at 607.8 tokens a second with 16 requests at once, then fell to 385 when 32 ran together.

⭐ The fast number is a crowd result. One person still gets about 115 tokens a second.

#AI