Every number on this page is reproducible. No simulated hardware, no fabricated results.
Reproducibility guarantee: All Python results can be verified by running python -m atomik_core benchmark on your own machine. FPGA results are from Gowin/Vivado synthesis reports and hardware test suites (80/80 parallel bank tests, 9/9 ATOMiK core tests, 6/6 display pipeline tests). Synthesis reports and test logs are in the source repository.
Python 3.12 / Kubuntu 24.04
| Throughput (ACCUM) | ~5 M ops/s |
| Rollback (10k undo) | ~140x faster |
| Change detection | ~330,000x less memory |
| Bandwidth (1 MB state) | 131,072x reduction |
Methodology: python -m atomik_core benchmark
Developer's machine. Run the same command on your hardware to compare.
Single bank, N=1, 81 MHz core clock
| Throughput | 94.5 Mops/s |
| LOAD latency | 64 cycles |
| ACCUM latency | 70 cycles |
| READ latency | 99 cycles |
| LUT usage | 477 (5.5%) |
Methodology: Synthesis reports + 80/80 hardware tests
Production SoC: PicoRV32 @ 25.2 MHz + ATOMiK @ 81 MHz. Deterministic latency (stdev < 0.5 cycles).
N=16 parallel banks, 66 MHz
| Throughput | 1,056 Mops/s |
| LUT usage | 1,779 (20.6%) |
| LUT scaling | 3.7x for 16x throughput |
Methodology: Synthesis sweep + 80/80 hardware tests
Sub-linear LUT scaling: ~65 LUT + 64 FF per additional bank.
N=512 parallel banks, 135.6 MHz
| Throughput | 69.7 Gops/s |
| LUT usage | 23,542 (44.3%) |
| Fmax (N=1) | 444.4 MHz |
| Marginal LUT/bank | ~34 |
Methodology: Vivado synthesis ceiling characterization (6 configs, 4 strategies each)
Synthesis-validated. Board integration pending (AX7020 in hand). N=512 is the XC7Z020 placement limit.
Same algebra at every tier. Code once, accelerate anywhere.
pip install atomik-corepython -m atomik_core benchmarkpython -m atomik_core benchmark --jsonThe benchmark suite runs five tests: rollback, change detection, multi-node convergence, bandwidth, and raw throughput. All operations are O(1) -- constant time regardless of state history.
pip install atomik-core && python -m atomik_core benchmark