🧠 AI🟢 Bullish

Practical FP4 Training for Large-Scale MoE Models on Hopper GPUs

arXiv – CS AI|Wuyue Zhang, Chongdong Huang, Chunbo You, Cheng Gu, Fengjuan Wang, Mou Sun|March 4, 2026 at 05:00 AM|1 views

🤖AI Summary

Researchers developed a training method for large-scale Mixture-of-Experts (MoE) models using FP4 precision on Hopper GPUs without native 4-bit support. The technique achieves 14.8% memory reduction and 12.5% throughput improvement for 671B parameter models by using FP4 for activations while keeping core computations in FP8.

Key Takeaways

→New training recipe enables FP4 efficiency for MoE models on Hopper GPUs without native 4-bit computation support
→Direct FP8-to-FP4 quantization avoids costly precision round-trips between different number formats
→Method reduces peak activation memory by 14.8% (11.8 GB) for 671B parameter models
→Training throughput improves by 12.5% from 1157 to 1302 tokens per GPU per second
→Approach maintains convergence quality while achieving substantial memory and bandwidth savings

#fp4-training #mixture-of-experts #hopper-gpu #model-optimization #memory-efficiency #training-throughput #quantization #large-scale-models

Read Original →via arXiv – CS AI

Act on this with AI

Stay ahead of the market.

Connect your wallet to an AI agent. It reads balances, proposes swaps and bridges across 15 chains — you keep full control of your keys.

Connect Wallet to AI →How it works

AI1h ago

Warren Buffett complained for decades that boosting profits by excluding exec stock comp was ‘cynical’—Nvidia just surprised Wall Street and agreed

Nvidia surprised Wall Street by agreeing to include executive stock compensation in its profit calculations, addressing a decades-old complaint by Warren Buffett about excluding such costs. This accounting change will likely boost Nvidia's credibility with investors while potentially pressuring competitors to follow suit.

AI4h ago

A Neuropsychologically Grounded Evaluation of LLM Cognitive Abilities

Researchers introduced NeuroCognition, a new benchmark for evaluating LLMs based on neuropsychological tests, revealing that while models show unified capability across tasks, they struggle with foundational cognitive abilities. The study found LLMs perform well on text but degrade with images and complexity, suggesting current models lack core adaptive cognition compared to human intelligence.

AI4h ago

Estimating Visual Attribute Effects in Advertising from Observational Data: A Deepfake-Informed Double Machine Learning Approach

Researchers developed DICE-DML, a new framework that uses deepfake technology and machine learning to measure causal effects of visual attributes in digital advertising. The method addresses bias issues in standard approaches when analyzing how image elements like skin tone affect consumer engagement on social media platforms.