IPU SQL Acceleration & Poplar.NET

The IPU Query Acceleration suite offloads relational aggregations and vector similarity operations from database engines directly to Graphcore Colossus IPU hardware.

Subsystem Components

1. PostgreSQL C Extension: Intercepts aggregation query trees and vector operators (<->, <=>), packaging batch data into gRPC plan envelopes.
2. SQL Server CLR Bridge: Managed .NET assembly linking SQL Server user-defined aggregates and table functions to the Accelerator daemon.
3. Poplar.NET Driver Shim: Managed C# wrapper providing safe native handles to Graphcore Poplar/PopART graph compilation and engine sessions.

Benchmark & Driver Invocations

1. Run IPU Query Acceleration Benchmark

dotnet run --project src/HFLabs/Products/QueryAccel/QueryAccel.csproj -c Release -- --benchmark

2. Test Poplar.NET Driver Lifecycle

dotnet run --project src/HFLabs/Hardware/Poplar.NET/PoplarDriver.csproj -c Release

Offloaded Operations

• Relational Aggregations: COUNT, SUM, AVG, MIN, MAX executed in parallel across thousands of IPU SRAM tiles.
• Vector Distance Metrics: Cosine similarity, Euclidean (L2) distance, and inner product for dense embedding indices.