Results that did not originate with us: an Intel developer article, two conference talks by users, a public benchmark suite maintained by a competitor.
A technical article on Intel's developer site, written for Intel's audience, on the record-once replay-many approach and what it does to a Monte Carlo on Intel CPUs.
Read on intel.com →Stephan Bosch, quantitative developer at ING, presented his own comparison of AAD approaches, including AADC, at the WBS Quantitative Finance Conference in Valencia.
Slides (PDF) →Svetlana Borovkova's talk on AAD for tail risk, model risk and stress testing, with AADC as the engine.
Slides (PDF) →A benchmark suite published by the developers of XAD, with Heston Monte Carlo, SABR calibration, XVA CVA and a LIBOR swaption. We ran it unchanged on 17 September 2026 with AADC added as an engine, same machine, same run, 10,000 paths. AADC was the fastest engine on all four, and 2.0 to 3.1 times faster than XAD's own code-generation mode. The numbers below are from that run; the suite's published tables do not include an AADC column.
| benchmark | finite differences | XAD, tape | XAD, code generation | AADC |
|---|---|---|---|---|
| Heston Monte Carlo, 8 sensitivities | 101.1 ms | 56.0 ms | 37.0 ms | 12.2 ms |
| SABR calibration, 15 | 59.0 ms | 13.1 ms | 8.1 ms | 3.7 ms |
| XVA CVA, 40 | 42.5 s | 3,556 ms | 1,253 ms | 400 ms |
| LIBOR swaption, 161 | 36.4 s | 1,387 ms | 695 ms | 346 ms |
Xeon Platinum 8280L, gcc 12.2, -O3 -mavx2 -mfma, AADC classic engine. XAD code-generation times are from the suite's own published results.csv, whose finite-difference times agree with ours to within 1 to 12 per cent.
An analyst ranking, not a benchmark, but the only independent comparison of AAD vendors that is published annually. 12th overall in the 2026 Quantitative Analytics 50 with five category awards.
Awards and rankings →