equilvalence of ARM versus x86 for numerical computation?
 

As ARM is making inroads into the cluster market, I am curious
how easy it will be to switch to this platform.

In the noughties I worked for a company using clusters of Opterons.
They looked at using Teslas for parallel grunting. However these had
different rounding errors. Since many of their algorithms involved
successive approximations, it took six programmer-months to tweak code
and prove they got the same results as on x86.

equilvalence of ARM versus x86 for numerical computation?
 

On 24/05/2014 10:02 PM, wrote:
As ARM is making inroads into the cluster market, I am curious
how easy it will be to switch to this platform.

In the noughties I worked for a company using clusters of Opterons.
They looked at using Teslas for parallel grunting. However these had
different rounding errors. Since many of their algorithms involved
successive approximations, it took six programmer-months to tweak code
and prove they got the same results as on x86.

FP math might always have some rounding errors from platform to
platform, even within some platforms. Integer math on the other hand
should be entirely consistent across the board.

Yousuf Khan