Architecture Effects When Interacting with GPU Technologies
summary
Benchmarks compare NVIDIA V100/P100 GPUs on Power9 AC922 (NVLink) vs x86 PCIe and DGX2, running an owl CNN, the Casa GPU aligner, plankton CNN and a grass‑seed classifier. Datasets: ~250 TB from forest recorders, ~100 TB/week ocean video. Results: identical code yields 2–3x faster end‑to‑end, 3x I/O throughput, segmentation/classification mixes give 5–8x, training drops from ~1 month to ~7 days; a live website demo races machines side‑by‑side.
AI-written summary · about
chapters (17)
- 0:00Intro
- 0:23Background
- 1:18GRB Model
- 2:49OSU OSL
- 3:48Agenda
- 5:20Ocean of Data
- 7:10Casa
- 9:47Predicting Ocean Health
- 13:02Grass Seed Separation
- 14:00Training Data
- 16:41Demo
- 28:03DGX
- 29:56CPU vs GPU
- 31:28Benchmarks
- 35:38Compile
- 37:03HandCompile
- 38:42Results
related talks
43mAI APPLICATIONS ON POWER9 SYSTEM
50mThe democratization of tomorrow’s technology
40mOptimize Your AI Cloud Infrastructure: A Hardware Perspective
40mFueling the AI Revolution with Gaming
19mJosh Patterson, Voltron Data | Nvidia GTC 2024
42mDifferential Software Solutions on HPC and IBM Power Architecture
13mFrom Silicon to Scale: Training at Scale
21mAnalyst Notebook