Most of the current benchmarks being run on Intel’s many-integrated core (MIC) technology exercise it on scientific or engineering algorithms. But the massively parallelXeon Phi co-processor is also a shoo-in for financial predictions, according to new benchmarks recently published by Xcelerit Computing Ltd. (Dublin, Ireland).
Xcelerit ran its financial benchmark on a Xeon Phi 5110P coprocessor installed in a HP ProLiant SL250 server housing twin Xeon E5-2670 main processors with 8 cores each. As such the comparison was between the 16 Xeon E5 cores running at 2.5-GHz compared to the 60 Xeon Phi cores running at just over 1-GHz.
How They Benchmarked
The benchmark used was the Monte-Carlo LIBOR Swaption Portfolio Pricing algorithm. The Monte-Carlo aspect rolls-the-dice with a random number generator that simulates possible future swaption paths. The London Interbank Offered Rate (LIBOR) is the short term interest rate charged when banks borrow from other banks around the world to cover end-of-day shortfalls in on-hand cash.
A swaption is a financial option that grants its owner the right, but not the obligation, to enter into an interest rate swap. Two types of swaptions exist: a payer-swaption that gives its owner the right to swap a fixed rate for a floating rate; whereas a receiver-swaption gives the owners the right to swap a floating rate for a fixed rate.
Like a chess-playing algorithm that looks ahead to all possible moves, the Monte-Carlo LIBOR Swaption Portfolio Pricing algorithm looks ahead to all possible swaption portfolios and their associated payoffs and sensitivity to floating interest rate changes.
The Intel Xeon Phi co-processor (red and blue) ran up to three-times faster than the traditional Xeon E5s (black) for parallel processing paths over 128k, offering up to three-times the performance for parallel paths as high as 1024k. SOURCE: Xcelerit
For the comparison, both executables were also run on the twin octal-core Xeon E5s. The results (see graph) shows that the 60 slower Xeon Phi cores outperformed the 16 faster Xeon E5 cores for parallel processing paths exceeding 128k (131,072).
For just a few paths, say 32k (32,768), the slower clock speed of the Xeon Phi cores resulted in performance three-times slower than the faster-clock-speed twin octal-core Xeon E5s.
And for maxed out simulations comparing as many as 1024K (1,048,576) paths, the massively parallel 60-core Xeon Phi ran almost three times faster than the Xeon E5s. Also of note is that the gap widened further when double-precision calculations were compared.