Those speedups when compared to older architecture from the competitor is nothing to write home about. AMD will bring in some price pressure to NVIDIA and getting little bit closer.
The performance seems to be determined by how much HBM memory and bandwidth they can package into 8-way systems. Then you need software architecture that keeps those processors busy at all times to get maximum output.
>We have been waiting to see some benchmarks for AI training, and AMD has finally given us two datapoints to work with, using the older Llama 2 models from Meta Platforms:
>The interesting bit here for us is to see a 10 percent performance advantage for the MI325X over the H200 when a single device is used, but that advantage disappears when moving to an eight-way GPU node.
>Where is the Llama 3.1 training data, AMD? This is what matters right now.
The performance seems to be determined by how much HBM memory and bandwidth they can package into 8-way systems. Then you need software architecture that keeps those processors busy at all times to get maximum output.
>We have been waiting to see some benchmarks for AI training, and AMD has finally given us two datapoints to work with, using the older Llama 2 models from Meta Platforms:
>The interesting bit here for us is to see a 10 percent performance advantage for the MI325X over the H200 when a single device is used, but that advantage disappears when moving to an eight-way GPU node.
>Where is the Llama 3.1 training data, AMD? This is what matters right now.