PPABench

A benchmark that verifies agentic chip design. A prompt becomes a core hardened into the fixed OpenFrame / chipignite chassis, then graded on both halves of a real chip: harden (PPA — area, power, fmax, DRC, LVS) and a hidden testbench (functional correctness + energy). Cycles are counted externally at the wrapper boundary and inputs are randomized per run, so results cannot be faked or precomputed.

PDK sky130 HD chassis OpenFrame / chipignite anti-cheat external cycles + random input v 0.1 · pre-MVP

MCU A Fixed RV32I — CRC32

graded on IPC + energy/instruction · ranked by EDP. Lower energy ranks higher.
#SubmissionFuncCyclesIPCCPI pJ/cycleEnergy nJpJ/instrArea µm²fmax MHzEDP nJ·µs
1 picosoc · reference baseline 1,433,9970.012083.23 3.545,076295 3,81088.51.46e8
Reference baseline (picorv32). Functional + cycles from live Verilator/cocotb (cycles counted externally, start→done). Energy = P·cycles/f @ 50 MHz. See MCU A details →

MCU B Open ISA — AES-128-CBC (random input)

verify-by-result over random key/plaintext · ranked by energy/byte.
#SubmissionFuncCyclesCyc/block pJ/cycleEnergy nJpJ/byteArea µm²fmax MHzTrials
1 picosoc · reference baseline 1,088,825272,206 3.543,85460,226 3,81088.53/3
3 fresh random seeds, all exact-match vs software golden. Held-out variant (SipHash-2-4) used off-leaderboard for contamination checks. See MCU B details →

JPEG Streaming baseline JPEG compressor

pixel stream in → JPEG bytes out · graded on PSNR / bitrate + PPA + throughput.
No submissions yet — reference run pending.
Read the JPEG benchmark spec →

PPABench v0.1 · reference baseline picosoc/picorv32 on sky130 HD · grading: harden (OpenROAD/Magic/netgen) + hidden cocotb TB · github.com/balbekov