CODEWITHSUNDEEP
floating-pointperformance-testingarbitrary-precision
zell
zell

Reputation: 10204

Performance benchmark of float128 type?

Has anyone experienced with float128 type?

I would like to know about its performance compared with double, and other high precision types such as boost::multiprecision:cpp_dec_float? Is there any benchmark already done?

Upvotes: 3

Views: 1422

Answers (1)

Ecir Hana
Ecir Hana

Reputation: 11508

Apparently, float128 is much slower than double. Like up to 100x slower but the precise slowdown would depend on the operations, of course.

Multiplying two 256x256 matrices is 100x slower.

The paper Benchmark of an MPFR emulation of Binary128 arithmetic cites 250x slowdown for summation, 120x slowdown for product, 400x slowdown for dot product.

Another paper Twofolds in C and C++ says __float128 summation is 150x (25x times 6x) slower than double summation.

Upvotes: 5

Related Questions