Keynote: When Hardware and Databases Collide
summary
Argues databases gain most by harnessing GPUs for massive parallel index lookups and OLAP-style batching (academics report 100x improvement in some queries) while acknowledging HBM limits to a few hundred GB and high cost. Describes HBM as stacked DRAM on an interposer and details Compute Express Link as a load/store remote-memory layer (three protocols) that can expose terabytes of pooled memory to solve stranded-memory problems in hyperscalers. Suggests pre-warming or offloading indexing to GPUs; transactional tiny-TPC workloads are unlikely to benefit.
AI-written summary · about
chapters (17)
- 0:00Introduction and keynote welcome
- 2:51Databases on diverse hardware
- 7:25GPUs as database accelerators
- 13:50GPU memory bandwidth and HBM
- 15:13CXL for expanding memory reach
- 22:47CXL potential and use cases
- 24:55Leveraging GPUs in databases
- 26:21Diversity in tech community
- 28:07GPUs for transactional vs analytics workloads
- 30:35CXL vs RDMA comparison
- 31:38Hardware interfaces and standardization
- 34:51GPUs for streaming data workloads
- 36:16Persistent memory limitations
- 37:40PG ST and GPU integration
- 38:21Berkeley DB and PostgreSQL history
- 42:35CXL accessibility concerns
- 45:05Q&A session
related talks
39mUnlock the potential of your data: Optimize your data store architecture...
37mKEYNOTE: Data Architecture Turned Upside Down
18mThe "New Normal" or An Evolution
38mKeynote | Data Warehouse, Data Lakes, What's Next?
52mByte-Addressable Persistent Memory and DBMS
33mBuilding smarter ML and AI models and making them more accurate
51mPl/cuda the fusion of HPC grade power
53mKeynote: From RAP to Snowflake - A Look at 50 Years of SQL DB Scalability