Importance of Itanium Floating Point Performance

When working with mathematical equations, such as the kind found in high-energy physics, fluid dynamics, and weather simulations, to say that the equations become complex is to seriously understate the case. Luckily for Itanium-based systems users, the registers and other elements of the chip were designed to optimize the Itanium processor family's ability to perform floating-point calculations.

In complex simulations such as weather or fluid dynamics, there are so many variables that few if any organizations outside of government research centers have really tried to pursue an effective, highly accurate simulation. This is the domain of the people who practice HPTC, and is where floating-point performance is most useful.

'Floating point' calculations are so called because they deal with equations where the decimal 'floats' around. For example, multiplying an integer that extends to 10 decimal places by another integer that is extended out to the 10th power.

Floating point performance is concerned with the places that go beyond the decimal. This capability is absolutely necessary when pinpoint accuracy is needed, such as in delicate scientific experiments. There are two critical elements involving how many decimal places of precision the computer can deal with in order to give you the most exact answer possible.

Extending the Decimal

The first important element needed to get the precision needed in many of these calculations is to be able to handle the largest number of decimal places possible. The more levels of decimal point accuracy available, the more precise the answer. For example, a 64-bit wide number that an Itanium-based system can handle obviously holds many more levels of precision than a 32-bit wide number used by RISC machines.

Reducing Rounding Error

The second element deals with the small, incremental errors that are inevitable when limitations are placed on the number of digits that are available on the machine performing the calculations. While incremental errors by themselves can be microscopically small, taken cumulatively they can throw the off the final solution.

A common example is where you want to multiply two decimal numbers with long strings of digits. You'll get what's called 'round off' error. Let's say that you're working with one number that runs out to five digits beyond the decimal point and a second number that has eight or nine digits beyond the decimal point. However, your computer may only be able to hold a number that goes out to six decimal places. In order to arrive at a solution, the computer 'rounds' off the back end of the number.

When doing equations that require precision, you don't want to 'drop' those extra digits. If you're performing millions and millions of calculations, you'll end up with a cumulatively great error. And of course, you'll come to a wrong conclusion from the erroneous data given back to you.

Built into the Intel Itanium processor is a math function that can help with these types of calculations. This function actually provides a couple of extra bits within the function to increase the precision of the final answer. Many HPTC applications take advantage of the fact that this instruction is there, as this gives better accuracy and the performance to execute this kind of math.