High Bandwidth Memory (HBM)
الذاكرة عالية النطاق
الذاكرة الرئيسية للمعالج الرسومي، كبيرة السعة (40–80 غيغابايت) لكن بطيئة نسبياً (~2 تيرابايت/ثانية). الانتباه المعياري يكتب مصفوفة N×N هنا وهو ما يسبّب عنق الزجاجة.
The GPU's main memory, large in capacity (40–80 GB) but relatively slow (~2 TB/s). Standard attention writes the full N×N matrix here, which is the bottleneck FlashAttention eliminates.
Also translated asذاكرة النطاق العريض، الذاكرة الرئيسية للمعالج الرسومي
First appears in this corpus in: FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness (2022)
Appears in these papers
- FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness2022in the sky ✦
- FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness2022in the sky ✦
- GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints2023in the sky ✦
- Mamba: Linear-Time Sequence Modeling with Selective State Spaces2023in the sky ✦