Whetstone

Whetstone is one of the oldest computer benchmarks, originating in 1972 and named after the English village where it was developed. It targets floating-point and transcendental math performance, complementing the integer-focused Dhrystone. Its unit, MWIPS (millions of Whetstone instructions per second), was a standard scientific-computing figure for many years and still appears occasionally alongside Dhrystone on embedded datasheets.

What It Measures

Whetstone exercises a mix of operations representative of numerical and engineering code: floating-point addition and multiplication, array indexing, conditional branches, and library calls to functions such as sine, cosine, square root, exponential, and logarithm. The result, MWIPS, summarizes how quickly the processor and its math library execute this blend. It is essentially a floating-point throughput estimate, with a heavy contribution from the speed of the transcendental-function library, which often dominates the total time.

Methodology

The benchmark runs a fixed set of modules in a loop, weighting each operation type to mirror observed scientific workloads, and times the total. As with Dhrystone, enough iterations are needed to swamp timer granularity, and compilers must not eliminate or precompute the math. Transcendental-function performance depends heavily on the math library, so results disclose compiler, flags, and library version. Valid comparisons hold those constant, because a faster libm can change MWIPS substantially without any hardware change. Run rules also caution against vectorizing the loop in ways the original code did not intend, which would break comparability with historical numbers.

How to Interpret Results

Higher MWIPS means faster floating-point and library math. Use it as a coarse comparison of scientific-style arithmetic between cores or toolchains. Because much of the score comes from transcendental functions, a faster math library can raise MWIPS without any hardware change, so always note the library and flags before comparing. Whetstone pairs naturally with Dhrystone for a rough integer-plus-floating-point picture of a small core, but neither should drive a serious purchasing or architecture decision on its own.

Limitations

Like Dhrystone, Whetstone is dated and cache-resident, so it ignores memory bandwidth and large working sets. It predates SIMD and modern vector units and does not reflect how current numerical code is vectorized or offloaded to accelerators. Library-dependent results make cross-platform comparison fragile, and small loop sizes mean the benchmark never stresses the memory hierarchy. For serious floating-point evaluation use LINPACK or the SPEC CPU floating-point suite; reserve Whetstone for historical context, smoke tests, and constrained embedded cores where nothing heavier is practical.