FireSim bring-up on AWS F1 (Rocket + Gemmini)

Status: planning, M5+ cloud-burst path Owner: board/fpga Platform decision: see docs/board/fpga/platform-selection.md

Purpose

FireSim is the cloud-burst alternative to owning a VCU118. It compiles the same Chipyard-generated Rocket+Gemmini target to AWS F1 (f1.2xlarge, f1.4xlarge, f1.16xlarge) or F2 once it is generally available, and runs it at hardware speed for long benchmark sweeps that would tie up an on-prem board for days.

This document is the bring-up runbook. It assumes the program has decided to use FireSim per the platform-selection decision; if VCU118 is the chosen path, follow board/fpga/vcu118/README.md instead.

AWS prerequisites

1. AWS account with EC2 F1 quota in us-east-1 or us-west-2. F1 quota is off by default; request Running On-Demand F instances >= 16 vCPU two weeks before first build.
2. S3 bucket for AGFI (Amazon FPGA Image) artifacts in the same region.
3. IAM role permitting ec2:*, s3:* on the bucket, and fpga:* for AGFI creation.
4. SSH key pair for the FireSim manager instance.

Toolchain layout

Vivado is required on the build instance even in the cloud path; FireSim wraps it but does not replace it.

Build phases

1. Local metasim (Verilator)

Before any cloud spend, run the same target in metasim on the manager (or a developer laptop) to confirm functional correctness:

cd $CHIPYARD/sims/firesim
./scripts/build-setup.sh
source sourceme-f1-manager.sh
firesim setupexample
firesim runworkload --runworkload-overrides metasim

Expected: Linux boots in metasim in roughly 10-30 minutes wall-clock per simulated boot, depending on host core count.

2. F1 AGFI build

firesim buildbitstream --hwdb-entry rocket_gemmini_f1

Wall-clock budget: 6-10 hours for a first-time XCVU9P build (Vivado synth + implementation + AFI ingestion by AWS). Subsequent rebuilds with the same shell are 4-6 hours.

3. On-cloud run

firesim launchrunfarm
firesim infrasetup
firesim runworkload --hwdb-entry rocket_gemmini_f1 \
                    --workload-name linux-uniform

Expected wall-clock for Linux boot to userland shell on F1: 30-90 seconds at the simulated SoC clock of ~80-100 MHz. This is the headline win versus metasim and the reason F1 is worth the cost.

Cost guard-rails

• f1.2xlarge on-demand: ~$1.65/h. Always set a max-runtime guard on the workload runner.
• Forgetting to firesim terminaterunfarm is the most common cost surprise. Add a CloudWatch alarm on EC2 hours-per-day for the FireSim tag.
• AGFI builds incur build-instance hours plus a one-time AFI ingestion fee (a few dollars per image).

What FireSim does not do

• It does not eliminate Vivado. The AGFI build runs Vivado under the hood.
• It does not validate physical timing for ASIC PD; only the FPGA shell.
• It does not provide JTAG to a real RISC-V debug module on F1 (use the FireSim TSI/dmi bridge instead).
• It does not replace VCU118 if reproducible per-cycle wall-clock matters -- AWS instance variance is real.

References

• docs/board/fpga/platform-selection.md
• board/fpga/vcu118/README.md
• docs/generators/chipyard/README.md
• FireSim docs: https://docs.fires.im