Benchmark results
Everything below comes from the latest full run: an interactive cross-card comparison, with each card’s full tables behind the links below. Card identity, ATR, CPLC, free memory and the capability probes are on the card signals page; the method and capability limits are on the methodology page.
Cross-card comparison
Pick one card as the baseline; it is always shown in the first column. Every other card shows its per-operation timing change relative to it. A green ↓ means faster than the baseline, a red ↑ means slower; the percentage is the change relative to the baseline.
| Test | NXP P71D600D600 · contactless · applet v10baseline | NXP P71D321D321 · contactless · applet v10 |
|---|---|---|
| Transport cost | ||
| 0 B | 4.06 ms | 3.33 ms↓ 18.0% |
| 8 B | 5.81 ms | 3.85 ms↓ 33.8% |
| 32 B | 9.98 ms | 4.91 ms↓ 50.8% |
| 64 B | 15.7 ms | 6.54 ms↓ 58.3% |
| Key pair generation | ||
| P-256 key pair generation | 133.5 ms | 142.0 ms↑ 6.4% |
| RSA-2048 key pair generation | 2.19 s | 7.03 s↑ 220.5% |
| Computation | ||
| 16-bit addition | 33.6 µs | 35.5 µs↑ 5.8% |
| 16-bit multiply-add | 47.9 µs | 50.0 µs↑ 4.6% |
| 16-bit division | 45.8 µs | 47.3 µs↑ 3.2% |
| byte array read | 96.9 µs | 86.8 µs↓ 10.4% |
| byte array write | 91.5 µs | 82.4 µs↓ 9.9% |
| method call | 66.2 µs | 66.5 µs↑ 0.5% |
| short array read | 96.8 µs | 88.3 µs↓ 8.8% |
| short array write | 92.6 µs | 83.4 µs↓ 9.8% |
| Memory | ||
| persistent byte write | 2.84 ms | 2.91 ms↑ 2.6% |
| persistent short write | 2.84 ms | 2.92 ms↑ 2.5% |
| persistent byte read | 87.1 µs | 84.6 µs↓ 2.9% |
| persistent short read | 87.0 µs | 86.5 µs↓ 0.5% |
| transient byte write | 98.4 µs | 89.3 µs↓ 9.3% |
| transient byte read | 98.9 µs | 86.2 µs↓ 12.9% |
| atomic arrayCopy, 64 B | 3.25 ms | 3.34 ms↑ 2.7% |
| non-atomic arrayCopy, 64 B | 3.26 ms | 3.29 ms↑ 1.1% |
| transaction per byte write | 3.13 ms | 3.19 ms↑ 2.0% |
| allocate 32-byte array | 9.21 ms | 9.43 ms↑ 2.4% |
| allocate 32 B + requestObjectDeletion | 22.1 ms † | 17.1 ms †↓ 22.9% |
| persistent block write, 16 B, one transaction | 6.07 ms | 6.90 ms↑ 13.7% |
| persistent block write, 32 B, one transaction | 8.88 ms | 10.5 ms↑ 18.2% |
| persistent block write, 64 B, one transaction | 14.6 ms | 18.0 ms↑ 23.4% |
| non-atomic arrayCopy, 16 B | 2.98 ms | 3.07 ms↑ 3.0% |
| non-atomic arrayCopy, 32 B | 3.09 ms | 3.14 ms↑ 1.5% |
| Cryptography | ||
| AES-128 ECB encrypt | 1.66 ms | 1.64 ms↓ 1.7% |
| AES-128 ECB decrypt | 1.74 ms | 1.65 ms↓ 4.9% |
| AES-192 ECB encrypt | unsupported | unsupported |
| AES-256 ECB encrypt | unsupported | unsupported |
| AES-128 CBC encrypt (zero IV) | 1.75 ms | 1.67 ms↓ 4.9% |
| AES-128 ECB init + encrypt | 3.11 ms | 7.97 ms↑ 156.2% |
| DES ECB encrypt | 1.66 ms | 1.61 ms↓ 2.8% |
| 3DES ECB encrypt | 1.73 ms | 1.65 ms↓ 4.5% |
| AES-128 CTR encrypt (zero IV) | 1.63 ms | 1.57 ms↓ 3.3% |
| SHA-256, 32 B message | 2.37 ms | 2.38 ms↑ 0.2% |
| SHA-256, 1 KiB message | 26.0 ms | 26.0 ms↓ 0.1% |
| SHA-1, 32 B message | 1.91 ms | 1.88 ms↓ 1.9% |
| SHA-512, 32 B message | 3.58 ms | 3.71 ms↑ 3.6% |
| HMAC-SHA-256 | 36.9 ms | unsupported |
| TRNG, 32 B | 912.7 µs | 757.7 µs↓ 17.0% |
| ECDSA P-256 SHA-256 sign | 46.8 ms | 53.5 ms↑ 14.2% |
| ECDSA P-256 SHA-256 verify | not reported | not reported |
| ECDH P-256, peer = G | 30.8 ms | 38.1 ms↑ 23.7% |
| RSA-2048 SHA-256 sign | 90.2 ms | 105.8 ms↑ 17.2% |
| RSA-2048 SHA-256 verify | not reported | not reported |
| RSA-2048 PKCS#1 encrypt | 6.86 ms | 8.13 ms↑ 18.4% |
| RSA-2048 PKCS#1 decrypt | 86.8 ms | 102.7 ms↑ 18.4% |
Transport is per command, key generation is per key pair, and the rest are per operation. † marks R² < 0.99; RSA key generation is dominated by the prime search and can vary by seconds.
Full data per card
- NXP P71D600 ·
contactless-d600— contactless · Measured at: 2026-11-03 03:51 UTC · Applet: v10 · Open full data - NXP P71D321 ·
contactless-d321— contactless · Measured at: 2026-11-03 04:01 UTC · Applet: v10 · Open full data
How to read the numbers
Section titled “How to read the numbers”- Per operation is the fitted slope divided by the number of operations in one iteration. The CPU tests perform 1000 operations per unit of
n, so those rows are the cost of one arithmetic or array operation; each crypto iteration is one operation. - APDU overhead is the fitted intercept and is independent of the operation: contactless round trip, command dispatch, plus the extra
GET RESPONSEwhen the applet returns data. That is why a single command’s wall-clock time is not a per-operation time. - The cross-card comparison lets you pick any card as the baseline, which is always shown in the first column; every other column shows the per-operation timing change relative to it: a green
↓is faster, a red↑is slower. Both cards ran the same applet version (v10), the same host code and the same reader, so the difference comes from the card. Runs with different applet versions should not be compared directly. - Repeatability across applet versions: the same D600 card was measured with applet v4 and then v5 (v5 only adds the capability-probe command; the measurement loops are unchanged). Across the 37 valid timing rows the median per-operation deviation was 0.5%, transport round trips differed by under 1.5%, and only the high-variance RSA key generation and allocate+delete rows moved substantially. Going from v6 to v7 changed only the write tests: the 31 non-write rows moved by a median of 0.32%, while the four write-related rows rose by two orders of magnitude (see below). Going from v7 to v8 only adds the block-write commands: the 36 rows the two versions share moved by a median of 0.24%, and transport round trips by under 1%. Going from v8 to v10 only changes how HMAC keys are built and probed: the 41 shared rows moved by a median of 0.29% on D600 and 0.16% on D321, and transport round trips by under 1.1% and 3.2% respectively.
- How a persistent write is written matters: the card skips programming bits that are already in the target state, so the write tests must write different data on every command. Otherwise they measure the interpreter loop (~84 µs) rather than programming (~2.8–2.9 ms). The write tests now derive their values from a persistent seed and echo it so the host can check that two identical commands wrote different data. For the same 16/32/64 bytes, one transaction and a single
arrayCopydiffer by more than five times (details on the methodology page). - Rows with R² below 0.99 (
†) should not be extrapolated: today only the allocate+delete test is not linear inn, so its slope is an average over the sampled range. - Not reported covers standalone RSA and ECDSA verify. On both cards
Signature.verify()returns false for a key pair generated in an earlier command, while sign + verify inside one command succeeds, so the data would not be trustworthy. Details are on the methodology page. - Unsupported covers AES-192 and AES-256 on both cards, and HMAC-SHA-256 on
D321. They are not slow; the cards do not implement them.D600HMAC-SHA-256 was also listed as unsupported before v10, but that came from the test building HMAC keys in the wrong unit; with that fixed it is supported and timed (see the methodology page).
Run-to-run variation
Section titled “Run-to-run variation”The same card and code still vary slightly between runs. CPU and symmetric crypto results are usually stable within 1%; two groups need care:
- RSA-2048 key pair generation is decided by the prime search: in the latest (v10) run 1.7–3.6 s per pair on the
D600card (median 2.2 s) and 2.1–8.9 s on theD321card (median 7.0 s), against 5.4–13.8 s and 4.7–20.1 s in the previous (v8) run. The table gives the median plus min/max. - Persistent writes differ by about 30× between a same-value rewrite and real programming, which is why every write now carries fresh data. The “~1.4 ms slow state” seen earlier was a blend of the two; those rows now fit at
R² = 1.0000. The harness still re-measures whenR² < 0.99and keeps the flag if it stays non-linear.
When quoting a result, include the card ID and the run time rather than a bare number.