High Speed Query Execution with Accelerators and C++
summary
A hardware SoC for big‑data queries claims at least x60 speedup versus CPU by running a novel massively parallel non‑von‑Neumann MFC plus dedicated accelerators on an ASIC (FPGA prototypes exist), exposed as a Spark plugin that compiles Spark plans via a pigeon query compiler and LLVM into DFG kernels, generated from C++/Python metaprogramming (C++17/20). Memory uses streaming buffers, 48 LSUs and a Banyan interconnect, and device runs as a PCIe card.
AI-written summary · about
chapters (10)
- 0:00Start
- 4:29Von Neumann Architecture
- 5:34Simple CPU Diagram
- 13:02Simple GPU Diagram
- 13:47GPU System Architecture
- 14:46Dataflow Architectures
- 20:55Memory Sub-System
- 22:37DOGE- Dataflow Oriented Graph Execution
- 29:33PIDGIN : Kernels
- 35:38PIDGIN: Complex Case kernels
related talks
13mHardware/Software Co-Programmable Framework for Computational SSDs to Accelerate Deep...
19mSpark vs. Accelerated Compute: A Balanced Look at Data Processing Strategies
09mAccelerating compute with software defined hardware (FPGA's)
01h 06mBodo: Automatic HPC Performance and Scaling for Data Processing in Python
16mTowards High-throughput and Low-latency Billion-scale Vector Search via CPU/GPU...
22mCXL-ANNS: Software-Hardware Collaborative Memory Disaggregation and Computation...
39mAccelerating Data Science with RAPIDS
01h 52mNeural Network Accelerator Design