GMPbench results. We have run the benchmark on the highest-frequency CPU of each type to which we have convenient access. Scaling to lower or higher frequencies should work well, since GMP mainly works off the caches.

The GMPbench suite is available for download. To run the benchmarks, you also need to compile gexpr.c and put it somewhere in your path.

GMP exercises a particular set of processor capabilities, widening integer multiplication being the most important one. Processors with poor integer multiply support get worse scores on GMPbench than on other benchmarks.

GMPbench measures multiplication of same size and different size operands, squaring, division, gcd, and extended gcd, and then RSA encryption and π computation. That's it. GMP typically performs a whole lot more operations which are not measured at all.

Let us make it absolutely clear that GMP exercises just integer instructions, and that GMPbench thus is an integer benchmark. We often see claims to the contrary, i.e. that it is a floating-point benchmark.

GMPbench is a single-threaded benchmark. Most CPUs below can execute multiple threads in parallel using some separate and some shared hardware resources. Most operations in GMP up to quite large operand sizes would scale very well over multi-core systems, but poorly with SMT. A hypothetical multi-threaded GMPbench would therefore give about 8 times the current Ryzen 2700X score but only about 4 times the current Xeon E3-1270v5 score.

The results below have been measured with software workarounds for the various hardware bugs (Meltdown, Spectre, L1TF, MDS, etc) at various stages of completeness. BIOS were at the newest available version, but the latest vendor provided microcode was not provided unless provided by the BIOS. We don't have an estimate of the slowdown GMP will suffer when workarounds are final.

Preliminary measurements for Intel Skylake (and thus Kaby Lake, Coffee Lake, Whiskey Lake, Amber Lake, Fool Ake, etc which are all just Sky Lake reiterations) indicate significant slowdown. We see no slowdown for AMD CPUs. There might be newer Intel CPUs which have been bug fixed, but last time we checked (in mid 2020) the bugs were discovered faster than they were fixed. Performance results which are inflated by corners being cut resulting in a security disaster are not of great interest. We have therefore crossed over Intel results.

GMP repo [measured at different times, therefore unfair] GMPbench 0.2 results

GMP 5.0.1 GMPbench 0.2 results

GMP 4.3.x GMPbench 0.2 results

GMP 4.3.x GMPbench 0.1 results

1. The clock frequencies for the above measures are not the same as for GMP 4.2, since we didn't have access to the same hardware. However, we have remeasured some of the 4.2 numbers and updated the table below.
2. The last column, "Optimal", is an estimate of what could be attained by writing optimised assembly code for this processor.

GMP 4.2.x GMPbench 0.1 results

GMP 4.1.x results

• There was no assembly loop support for Opteron/Athlon64 in GMP 4.1.4. We therefore include two results above for Opteron, 32-bit results using the Athlon32 assembly loops, and 64-bit results using plain C with inline assembly.
• The performance for the Pentium 4 EM64T processors disappoints. There are performance problems with many instructions GMP depends on: 64-bit multiply, integer right shift, and conditional moves and set-on-condition instructions all need around 10 cycles, and are not fully pipelined. See also Instruction latencies and throughput for AMD and Intel x86 processors
• The 32-bit 90nm Pentium 4 processors (Prescott) run GMP applications slower than older Pentium 4 processors. The reason is that Prescott has longer latencies for SSE2 instructions and memory loads.
• UltraSPARC 3's poor scores are a result of its inadequate integer multiply support (shortcomings in both ISA and implementation).