AMD revealed its Helios architecture, aimed at enhancing AI model training and inference, during the Advancing AI event. This system features 72 MI400 GPUs that utilize Ultra Accelerator Link (UALink) to maximize efficiency. With a total of 432 GB of HBM4 memory and 40 petaFLOPS performance, Helios promises 2.9 exaFLOPS for inference and 1.4 exaFLOPS for training. The architecture may compete fiercely with Nvidia's upcoming systems, such as the Vera Rubin NVL144, providing a significant advancement in GPU-driven computation.
AMD's new Helios rack-scale architecture, featuring UALink interconnect and MI400 GPUs, is set to deliver exaFLOPS performance for large-scale AI deployments.
With 72 MI400s, Helios aims for 2.9 exaFLOPS of FP4 for inference and 1.4 exaFLOPS of FP8 for training, competing directly with Nvidia's Vera Rubin.
Collection
[
|
...
]