Hardware Accelerators

Streaming Graph Accelerator — JetStream

JetStream evaluates queries over streaming graphs where edges are constantly inserted, removed, or updated. By adopting an event-driven execution model, the accelerator limits computation to the affected neighborhood, reuses intermediate state, and reorganizes memory access patterns for higher bandwidth utilization. JetStream reduces cold-start recomputation by up to 90% and achieves 18× speedups over optimized CPU frameworks at comparable batch sizes.

Evolving Graph Accelerator — MEGA

MEGA targets evolving graphs that require evaluating a query across a sequence of snapshots. It builds on the CommonGraph software model and introduces Batch-Oriented Execution so snapshots can be processed concurrently. Combined with incremental event streaming and tailored memory hierarchies, MEGA sustains high throughput while avoiding the overhead of handling deletions explicitly.

Design Principles

• Event-driven scheduling that matches graph update semantics.
• Custom data paths and scratchpad hierarchies to maximize locality.
• Hardware–software co-design that surfaces accelerator capabilities through compiler support and APIs.