These results are for version 4.5.70 of the Device Detection C library and version 4.5.34 of the IP Intelligence C library. The machines and how we tested, including what InMemory and Balanced mean, follow the results.
Device Detection🔗
InMemory, detections per second🔗
| Machine | IsMobile | PlatformName | BrowserName | IsCrawler | One per component | Every property |
| Dell server | 11,171,641 | 15,064,331 | 10,290,677 | 9,151,766 | 2,935,237 | 1,615,576 |
| AWS c7g.2xlarge | 1,362,952 | 1,701,221 | 1,172,524 | 1,174,748 | 306,275 | 166,181 |
| AWS c6i.2xlarge | 1,002,638 | 1,232,228 | 873,894 | 868,417 | 235,993 | 108,482 |
| Mac mini (2020) | 1,244,922 | 1,448,673 | 1,077,340 | 993,003 | 285,818 | 164,508 |
| Mac mini (2018) | 748,856 | 1,039,969 | 724,112 | 635,516 | 181,778 | 93,795 |
| Raspberry Pi 4 Model B | 91,206 | 113,152 | 83,799 | 76,767 | 22,397 | 9,678 |
| Desktop PC from 2008 | 66,240 | 104,594 | 75,791 | 61,373 | 16,148 | 9,278 |
Measured on 31 July 2026.
Balanced, detections per second🔗
| Machine | IsMobile | PlatformName | BrowserName | IsCrawler | One per component | Every property |
| Dell server | 241,159 | 210,165 | 150,522 | 189,234 | 56,606 | 638 |
| AWS c7g.2xlarge | 350,858 | 350,747 | 184,068 | 327,593 | 60,601 | 3,374 |
| AWS c6i.2xlarge | 236,870 | 233,612 | 143,369 | 223,668 | 46,170 | 2,750 |
| Mac mini (2020) | 347,849 | 332,328 | 203,284 | 208,838 | 64,038 | 2,747 |
| Mac mini (2018) | 145,968 | 134,552 | 84,041 | 140,662 | 27,676 | 1,568 |
| Raspberry Pi 4 Model B | 35,871 | 42,699 | 21,204 | 40,980 | 6,760 | 306 |
| Desktop PC from 2008 | 75,363 | 83,635 | 47,220 | 62,049 | 14,276 | 472 |
Measured on 31 July 2026.
The every-property column for Balanced comes from about 20,000 records rather than one million. Balanced is very slow when every property is requested, so if you need many properties, keep the whole data file in memory. Performance Options also recommends keeping the data file in memory when many threads run at once.
On the 2008 desktop, Balanced beat InMemory for IsMobile and IsCrawler. Don't read much into that, as Balanced timed only the first one million records and InMemory all six million.
Startup and memory🔗
| Machine | InMemory startup (ms) | Balanced startup (ms) | InMemory dataset memory (MB) | Balanced dataset memory (MB) |
| Dell server | 197 | 130 | 97.0 | 12.0 |
| AWS c7g.2xlarge | 142 | 84 | 95.7 | 7.6 |
| AWS c6i.2xlarge | 148 | 58 | 95.7 | 7.6 |
| Mac mini (2020) | 76 | 30 | n/a | 7.5 |
| Mac mini (2018) | 163 | 63 | 95.6 | 7.5 |
| Raspberry Pi 4 Model B | 1,067 | 743 | 95.8 | 7.4 |
| Desktop PC from 2008 | 2,900 | 376 | n/a | 7.6 |
Measured on 31 July 2026.
IP Intelligence🔗
InMemory, lookups per second🔗
| Machine | RegisteredName | AsnName | Longitude | All three | Every property |
| Dell server | 25,449,915 | 32,571,687 | 10,104,818 | 2,435,107 | 1,974,117 |
| AWS c7g.2xlarge | 4,363,143 | 5,688,929 | 2,077,427 | 1,131,583 | 270,740 |
| AWS c6i.2xlarge | 2,518,674 | 3,053,785 | 1,507,127 | 761,703 | 178,844 |
| Mac mini (2020) | 2,352,278 | 4,806,364 | 884,235 | 559,059 | 209,876 |
| Mac mini (2018) | 1,194,046 | 2,512,458 | 518,619 | 321,427 | 133,013 |
Measured on 31 July 2026.
Balanced, lookups per second🔗
| Machine | RegisteredName | AsnName | Longitude | All three | Every property |
| Dell server | 613,047 | 519,802 | 553,565 | 292,703 | 4,763 |
| AWS c7g.2xlarge | 1,511,950 | 1,865,780 | 1,001,305 | 334,432 | 10,887 |
| AWS c6i.2xlarge | 1,230,586 | 1,512,510 | 789,145 | 282,252 | 9,713 |
| Mac mini (2020) | 1,496,085 | 1,508,657 | 522,488 | 268,159 | 15,914 |
| Mac mini (2018) | 890,080 | 1,093,366 | 404,271 | 135,130 | 5,465 |
Measured on 31 July 2026.
As with Device Detection, the Balanced every-property result comes from about 20,000 records.
Startup and memory🔗
| Machine | InMemory startup (ms) | Balanced startup (ms) | InMemory dataset memory (MB) | Balanced dataset memory (MB) |
| Dell server | 2,754 | 1,139 | 5,838.0 | 2,667.0 |
| AWS c7g.2xlarge | 1,615 | 743 | 5,838.5 | 2,663.1 |
| AWS c6i.2xlarge | 2,402 | 1,089 | 5,838.5 | 2,663.1 |
| Mac mini (2020) | 2,893 | 726 | n/a | 2,413.1 |
| Mac mini (2018) | 4,706 | 1,594 | n/a | 2,457.2 |
Measured on 31 July 2026.
Machines🔗
| Machine | Processor | Cores (physical/logical) | Memory (GB) | Operating system | Compiler |
| Dell server | Two AMD EPYC 7742, 64 cores each, one thread per core | 128/128 | 1,259 | Ubuntu 24.04, in a Docker container | GCC 13.3 |
| AWS c7g.2xlarge | AWS Graviton3 (Arm Neoverse-V1), 2.6 GHz | 8/8 | 16 | Amazon Linux 2023 | GCC 11.5 |
| AWS c6i.2xlarge | Intel Xeon Platinum 8375C (Ice Lake), 3.5 GHz all-core turbo, two threads per core | 4/8 | 16 | Amazon Linux 2023 | GCC 11.5 |
| Mac mini (2020) | Apple M1, four performance and four efficiency cores | 8/8 | 8 | macOS 14.4 | Apple Clang 15.0 |
| Mac mini (2018) | Intel Core i5-8500B, 3.0 GHz | 6/6 | 8 | macOS 15.7 | Apple Clang 17.0 |
| Raspberry Pi 4 Model B | Arm Cortex-A72 | 4/4 | 3.7 | Debian 12 | GCC 12.2 |
| Desktop PC from 2008 | Intel Core 2 Duo E8400, 3.0 GHz | 2/2 | 3.6 | Ubuntu 24.04 | GCC 13.3 |
Memory is the amount the operating system reported, in GB of 1,073,741,824 bytes, except for the two AWS instances, which show the size of the instance.
How we tested🔗
We ran the Device Detection C library on seven machines, from a 128-core server down to a desktop PC built in 2008, and the IP Intelligence C library on five of them. We recorded throughput, startup time, and memory use on each. Every machine used the same library versions, data files, and evidence, and we ran each test once.
Libraries and data🔗
We built the C libraries from the public repositories, at version 4.5.70 for Device Detection and 4.5.34 for IP Intelligence. Each machine compiled its own Release build with -O3. The build script also adds -march=native, dropping it if the compiler rejects it. The machines table lists each compiler.
The data files were TAC-HashV41.hash (99,667,840 bytes) for Device Detection and 51Degrees-IPIV4EnterpriseIpiV41.ipi (6,121,965,416 bytes) for IP Intelligence, both published on 30 July 2026.
Configurations🔗
Each library ran in two configurations:
- InMemory loads the whole data file into memory. It is the configuration that the
MaxPerformanceprofile selects. For these runs we also compiled the library with the memory-only option (FIFTYONE_DEGREES_MEMORY_ONLY), which leaves out the code for reading from disk. - Balanced, the default profile, reads most of the data file through caches that keep the most used data in memory and leave the rest on disk. For IP Intelligence, Balanced still loads some parts of the data file fully into memory, including the text strings and the IP address graphs described in IP Intelligence Algorithm. Each Balanced run made an untimed warm-up pass over the same records it then timed, so the caches had already seen that traffic. We set the expected concurrency to the number of threads, as Concurrent Processing describes. On the Dell server, with 128 threads, that raised every cache smaller than 16,384 entries to 16,384.
Properties🔗
The IsMobile, PlatformName, BrowserName, and IsCrawler columns each request that property alone. In that order, they come from the hardware, operating system, browser, and crawler components. "One per component" requests all four properties together, and "Every property" requests all 279. For IP Intelligence, the RegisteredName, AsnName, and Longitude columns each request that property alone. "All three" requests them together, and "Every property" requests all 49.
Requesting fewer properties is faster because each detection has fewer values to read, and for Device Detection the library evaluates only the components that the requested properties belong to (see Restrict properties).
Evidence🔗
Device Detection used six million sets of HTTP request headers from real web traffic. Each set has a User-Agent, plus User-Agent Client Hints where the browser sent them. Popular devices and browsers appear many times.
IP Intelligence used six million IP addresses from real web traffic, with no address repeated in the set.
Each machine ran one thread per logical processor, and each thread took its own share of the records. In the timed pass every record was processed once, and no results cache was used. The InMemory runs timed all six million records and the Balanced runs timed the first one million. The one exception is the Balanced every-property run, which was limited to 20,000 records because reading every property with Balanced is very slow. Each thread's share of that limit was rounded up to a whole number of records, which made the totals 20,096 on the Dell server and 20,004 on the 2018 Mac mini.
Reading the tables🔗
Detections per second is the total number of records that all the threads processed, divided by the wall-clock time from the first thread starting to the last one finishing. A detection here means taking one record, finding the result, and reading every requested property value. The IP Intelligence tables count lookups per second in the same way, where a lookup finds the properties for one IP address. Loading the evidence into memory happened before the clock started and isn't included.
Startup is how long it took to load the data file and be ready to detect, with every property requested.
Dataset memory is how much the process's resident memory grew when the data file was loaded, before any detections ran. It doesn't include the evidence, and for Balanced it doesn't include the caches as they fill up. MB means 1,048,576 bytes. Four of the InMemory results showed no growth at all, which can't be right when the whole file is loaded, and the tables show n/a for those. Only the memory reading is in doubt, because checksums of the detection results from those runs match those from the other machines.
To work out how long one detection takes with every thread busy, divide the number of threads by the detections per second, because each thread handles its records one after another. On the Graviton3 instance with InMemory and IsMobile, eight threads at 1,362,952 detections per second works out at about 5.9 microseconds per detection.
What we didn't measure🔗
- Which machines kept
-march=native. - A standard build using
MaxPerformance, which may be slower than the InMemory figures here. - Memory use once the Balanced caches had filled.
- How much results vary from one run to the next, because each test ran once.
- What effect the intermittent low-voltage warnings we saw during the Raspberry Pi run had on its results.
- IP Intelligence on the Raspberry Pi and the 2008 desktop. The IP Intelligence InMemory run needs 5,838 MB, which is more memory than either machine has, so we left the whole product out on both. That includes the Balanced run, even though Balanced needs less memory.
- Why the 128-core server was slower than the eight-core Graviton3 instance in every Balanced column.