This blog presents the motivation, design principles, and key results behind RecScale, a system-aware approach to scaling Deep Learning Recommendation Models (DLRMs) that addresses critical memory and communication bottlenecks in distributed training.
Efficient full-parameter fine-tuning of GPT-OSS-20B & Qwen3-14B models on a single NVIDIA GH200 and Llama3-70B on four NVIDIA GH200 Superchips, while delivering up to 600 TFLOPS training throughput.
This blog presents a deep analysis of Alpha-Fold 3 (AF3) training pipelines, pinpointing their inefficiencies and introduces MegaFold: an end-to-end training system for AF3 that addresses the aforementioned issues.
This blog presents the motivation, insights, and key optimizations behind VoltanaLLM, our system for energy-efficient LLM inference. We’ll walk through why energy matters, how conventional GPU frequency scaling falls short, the surprising behaviors we uncovered when profiling LLM serving, how P/D disaggregated serving creates unique opportunities, and how VoltanaLLM’s co-design of frequency control and routing achieves up to 36.3% GPU energy savings while maintaining near-perfect Service Level Objective (SLO) attainment.
This blog presents the background and key optimizations behind X-MoE, along with our hands-on experience scaling MoE model training on Frontier, the AMD GPU supercomputer.