Beethoven

Hardware acceleration for FPGA and ASIC, simplified

What is Beethoven?​

Beethoven is a complete framework that makes hardware acceleration accessible. Write your accelerator in Chisel, and Beethoven handles the rest: automatic C++ bindings, multi-platform deployment, memory management, and runtime infrastructure.

Design once. Deploy everywhere. From simulation to AWS F2 to Xilinx Kria to custom ASICs.

Why Beethoven?​

Traditional FPGA/ASIC development is painful. You write RTL, manually create register maps, wire up memory controllers, build platform-specific shells, write custom drivers, and pray it all works. Beethoven eliminates this drudgery.

Quick Start​

1. Clone the template:

git clone https://github.com/Composer-Team/beethoven-template
cd beethoven-template

2. Define your accelerator:

class MyAccelerator(implicit p: Parameters) extends AcceleratorCore {
  val io = BeethovenIO(
    new AccelCommand("process") { val addr = Address() },
    EmptyAccelResponse()
  )
  val reader = getReaderModule("input")
  val writer = getWriterModule("output")
  // Your logic here
}

3. Build and simulate:

sbt run  # Generate hardware
cd Beethoven-Runtime && make sim_icarus
cd ../testbench && make && ./my_test

4. Deploy to FPGA: Change one line (platform = new KriaPlatform) and rebuild. Same design now runs on real hardware.

→ Full Getting Started Guide

Key Features​

Automatic Software Integration​

Your BeethovenIO interface becomes a type-safe C++ function. No manual command encoding, no register maps, no guesswork.

// Hardware (Chisel)
BeethovenIO(new AccelCommand("matmul") {
  val a_addr = Address()
  val b_addr = Address()
  val n = UInt(32.W)
})

// Generated C++ (automatic)
namespace MyCore {
  response_handle<bool> matmul(
    uint16_t core_id,
    remote_ptr a_addr,
    remote_ptr b_addr,
    uint32_t n
  );
}

Multi-Platform Deployment​

Write once, deploy anywhere. Beethoven abstracts platform details:

• AWS F2/F1: 3-die cloud FPGAs with automatic AFI generation
• Xilinx Kria: Embedded Zynq UltraScale+ boards
• Xilinx U200: Data center accelerator cards
• Simulation: Verilator, VCS, Icarus Verilog
• Custom Platforms: Define your own (ASIC, custom FPGA)

Memory Made Simple​

Forget AXI4 protocol specs. Request memory interfaces by name:

val reader = getReaderModule("input_data")
val writer = getWriterModule("output_data")

reader.requestChannel.bits.addr := my_address
reader.requestChannel.bits.len := num_bytes
// Data automatically streams on reader.dataChannel

Beethoven generates DMA engines, handles protocol conversion, and manages physical memory channels.

Multi-Core Architectures​

Build complex heterogeneous systems with multiple accelerator types:

• Automatic resource allocation across cores
• Core-to-core communication topology
• Platform-aware placement and floorplanning

Example: 23-core transformer attention accelerator deployed on AWS F2.

Full Hardware Control​

Beethoven doesn't hide hardware complexity - it manages it. You still write Chisel (or Verilog via blackboxes), with full access to:

• Custom datapaths and pipelines
• Scratchpads and on-chip memory
• Multi-die floorplanning (SLR partitioning)
• Clock domain crossings
• Custom protocols

Proven at Scale​

Beethoven has been used to build and deploy:

• 23-core transformer attention accelerator on AWS F2
• Multi-die floorplanned designs across 3 SLRs
• High-throughput DMA engines saturating 512-bit memory interfaces
• Heterogeneous accelerator topologies with core-to-core communication

Performance comparable to hand-written RTL. Development time measured in days, not months.

How It Works​

1. Define your hardware in Chisel using AcceleratorCore
2. Specify interfaces with BeethovenIO (host commands) and memory channels
3. Configure your build with target platform and build mode
4. Beethoven generates:
   • Synthesizable Verilog RTL
   • Type-safe C++ bindings matching your interface
   • Memory controllers and DMA engines
   • Platform-specific integration (shells, constraints)
5. Write testbench using generated C++ API
6. Simulate with Verilator/VCS/Icarus
7. Deploy to FPGA or ASIC with platform-specific flow

→ See Complete Example

Who Should Use Beethoven?​

✅ You want hardware-level performance but don't want to spend weeks on integration boilerplate

✅ You need multi-platform support (cloud FPGAs, embedded boards, custom ASICs)

✅ You're building custom accelerators where HLS doesn't give you enough control

✅ You value fast iteration with co-simulation and automatic software generation

❌ You need a GUI-based design tool (Beethoven is code-first)

❌ Your team can't learn Chisel/Scala (though we provide templates and examples)

Compare Beethoven​

→ Detailed Comparison

Get Started Now​

New to Beethoven?​

Start installing Beethoven using Beethoven CLI Tools and see demo examples at Beethoven Zoo.

Specific Use Case?​

• Multi-die FPGAs - SLR partitioning and floorplanning
• AWS F2 Deployment - Cloud FPGA acceleration
• Embedded Systems - Xilinx Kria/Zynq boards
• Debugging - Simulation and FPGA debugging

Tutorial Slide​

Follow our ISCA26 Tutorial slides for more references!

CheatSheet​

A Beethoven-related-command cheatsheet is provided!

Learn More​

Conference Tutorials​

Hands-on workshops and tutorials from conferences where Beethoven has been presented.

• Conference Tutorials Page - Workshop materials and presentations

Research & Publications​

Read peer-reviewed research on Beethoven's architecture and performance.

• ISPASS 2025 Paper - Framework design and evaluation
• All Publications - Conference papers and technical reports

Community & Support​

• Issues & Bugs: GitHub Issues
• Questions: Team Contacts
• Template: Beethoven Accelerator Template

Hardware acceleration shouldn't be this hard. Beethoven makes it simple.