Announcement
NexTron is claimed to deliver up to 20.9× DeepSpeed’s SFT throughput on H100 GPUs
Physion Labs says its matched-hardware benchmarks also showed up to 3.1× end-to-end RL throughput versus Miles–Megatron.
TLDR
Physion Labs has shared the technology behind NexTron, its world-intelligence training engine. In matched-hardware benchmarks, it reports up to 20.9× SFT throughput versus DeepSpeed on H100 GPUs, up to 6.6× versus Megatron on H200 GPUs and up to 7.2× versus Megatron on B200 GPUs. It also reports up to 3.1× end-to-end RL throughput versus Miles–Megatron and says NexTron can fine-tune a 180B model on eight H100 GPUs.
Combined views
5.8K
2 Sources, first seen ago
