Cyfra: Scala on GPU
summary
Cyfra compiles Scala expressions down to GPU IR as a normal library (no compiler plugin or macros) and runs via Vulkan, claiming cross-vendor support for Nvidia, AMD, Apple and Intel. A local Scala inference engine stitches ~200 GPU programs for a 1B model and ~500 for 3B into one G execution, keeping data on-GPU to avoid RAM roundtrips. Performance sits near llama.cpp with room to optimize. Highlights: library-as-DSL, abstraction elevator.
AI-written summary · about
chapters (15)
- 0:00Cyfra and Scala on GPU
- 1:28Why GPU programming matters
- 3:56Challenges of GPU programming
- 6:02Cyfra bridges GPU tooling gap
- 7:51First Cyfra GPU example
- 9:17Cyfra’s DSL features
- 12:04Limitations of simple GPU functions
- 13:27Introducing G programs
- 16:01Composing pipelines with G
- 17:47Cyfra vs JVM GPU tools
- 21:04Recipe for GPU libraries
- 23:11Cyfra’s architecture
- 25:38Real-world: Tifra Lama
- 29:10LLM inference in Scala
- 30:58Performance and roadmap
related talks
48mMassively Parallel Distributed Scala Compilation And You
57mScalan: a reasonably typed meta programming framework in Scala
43mLeveraging Scala to Build Hardware at Scale
35mA Deep Dive into Scala Native Internals
36mHands-on Scala-Native
15mKeynote: Reimagining the Future of High Performance Computing Catalysed by RISC-V
06mInterview
32mData Science in Scala with ScalaPy