Ce Zhang, Postdoctoral Researcher, Stanford University
summary
Built a Caffe-compatible engine Caffe con Troll (CcT) that applies standard batching and CPU optimizations to get about 4.5x throughput versus Caffe and sustain roughly 80% of peak FLOPS on modern CPUs, making multicore CPU hosts competitive with a single GPU and enabling simple hybrid CPU–GPU scheduling (per-host strategy changes give 3–4x speed differences on AlexNet). A DeepDive pipeline processes ~300,000 papers to extract ~2 million relations and matches expert-curated paleontology database quality in blind evaluations.
AI-written summary · about
chapters (4)
- 0:00Start
- 6:09Challenge of Manual KBC
- 9:56Validation on Real Applications
- 16:33One of the four shallow ideas
related talks
13mFrom Silicon to Scale: Training at Scale
16mRevisiting Pipeline Parallelism for LLM Serving
26mDeepSpeed: Efficient Training Scalability for Deep Learning
40mOptimize Your AI Cloud Infrastructure: A Hardware Perspective
14mFine-tuning giant neural networks on commodity hardware with automatic pipeline...
34mDeep Customized Kubernetes for Machine Learning in Tencent
34mPanel Discussion: Training and Inference at Planet Scale
36mAccelerating High-Performance Machine Learning at Scale i...