Preview: Hardware and Offloads
summary
Four systems claim concrete wins: a prediction-based scheduler integrates ML into Azure VM placement and shows twice the oversubscription capacity without hurting critical VMs; an integrated proactive ABR for 360° video allocates power per function and reduces energy oscillations; a surveillance pipeline uses spatial-skew multi-pass operators and 100x more video processed in real time with a zero-streaming front end; kernel and hardware pieces introduce a guarded page table and mountable Merkle trees to scale enclaves, and a nanoPU-style path that gives single-cycle packet access to cut tail latency.
AI-written summary · about
chapters (10)
- 0:00Intro
- 0:38Background
- 3:35Predictions
- 9:51Memory Management
- 15:43Logs
- 16:33NUMA
- 17:01Locality Aware Memory
- 17:34Complex Error Prone
- 18:33NROS
- 19:51NanoPU
related talks
23mKeynote: Launchpad
02mHighlights
13mAkeana : Breaking Performance Barriers
12mKeynote: 100G Intelligent Edge Innovation with RISC-V (The Future is Now)
20mSpaceExit: Enabling Efficient Adaptive Computing in Space with Early Exits
14mFreedom from memory interference
12mCost-effective Hardware Accelerator Recommendation for Edge Computing
28mML on the Edge for Industry 4.0 with Arm Ethos N-78 Neural Processing Unit