Enabling SQL to FPGA Using HLS for Filtering/Stream Data Transformations
summary
A compiler automatically transpiles SQL and a table schema to HLS C++ that becomes a streaming FPGA filter, using the Fletcher framework for Arrow buffers and CAPI on POWER to avoid host copies. Tests run from C simulation to cloud FPGA images and show per-instance throughput matching simulation (~1,000 expected, ~950 measured) with about 30 MB/s per 32-bit bus bit. Resource usage stays tiny so SQL→HLS C++ transpiler, Fletcher + CAPI on POWER, ~950 throughput, <0.5% FPGA per core.
AI-written summary · about
chapters (11)
- 0:00Introduction to FPGA and SQL
- 2:53Fletcher and Arrow overview
- 3:58Compiler design and goals
- 7:11Front-end and AST generation
- 9:00Middleware transformations
- 11:07Back-end HLS code generation
- 12:32Testing and verification stages
- 14:47Hardware throughput results
- 17:40Resource usage and scalability
- 18:42Future optimizations and potential
- 20:32Outlook and FPGA adoption
related talks
17mHardware Accelerated Big Data Processing
06mAchieve predictable performance
27mFPGA-Accelerated Compactions for LSM based Key-Value Store
23mBetter, Faster, Stronger Streaming: Your First Dive into Flink SQL
38mKeynote | Data Warehouse, Data Lakes, What's Next?
48mAnalytics Architectures on Microsoft Fabric
29mSQL on FHIR 2.0
18mThe "New Normal" or An Evolution