Hi Gaurav,
I will definitely find it useful. I have some 1+1D codes for evolving
spherically harmonic decomposed modes of fields in black hole spacetimes
using the Discontinuous Galerkin method. Here I find that, in some cases
even for moderate l-values, the accuracy is limited by double precision
round-off error, rather than truncation error.
In these cases, I definitely need higher precision in order to take full
advantage of the exponential convergence with DG-order.
Cheers,
Peter
On Friday 2016-09-02 07:27, Gaurav Khanna wrote:
Date: Fri, 2 Sep 2016 07:27:24
From: Gaurav Khanna <[email protected]>
To: [email protected]
Subject: Re: [Developers] High numerical precision processor from IBM
Hello Folks —
Just an update on this (rather old) message (below). Hardware support for
quadruple-precision floating-point (128-bit) will be available on the Power9
processor nextyear:
http://forwardthinking.pcmag.com/show-reports/347472-amd-ibm-and-inte
l-point-the-way-to-new-processors
If any of you find this functionality useful, or even generally support the
idea, feel free to drop me a line. I will forward it to the team at IBM that
fought for this to be included. Thanks.
Cheers
Gaurav
---------------------------------------------------------------------------
-------------
GAURAV KHANNA, (508) 910 6605, http://gravity.phy.umassd.edu
Professor, Physics Department, College of Engineering
Assoc. Director, Center for Scientific Computing & Visualization Research
Graduate Program Director, Engg & Appl. Sci. Ph.D. Program
University of Massachusetts Dartmouth
"Black holes are where God divided by zero." - Steven Wright
On Jul 3, 2012, at 4:06 PM, Gaurav Khanna <[email protected]>
wrote:
Dear Colleague --
I hope you are doing well. I'm requesting some feedback here on the
potential need for high-precision numerics in your research work in
the next few years. By high-precision, I'm referring to higher than
double floating-point precision (64-bit or ~14 decimal digits) i.e.
what is traditionally known as quadruple-precision (128-bit or ~30
decimal digits) and perhaps even higher. Today, one can easily emulate
such a high level of precision through software libraries, but these
often perform an order-of-magnitude slower than full
hardware-supported, double-precision computations.
I know that in some subareas of our research community high-precision
numerics are going to be necessary and are already in use in some
projects. But there may be other areas as well. Long duration
simulations or computations utilizing higher-order methods
(pseudo-spectral etc.) are likely to benefit from high-precision
numerics. There may also be some benefits in the context of studying
some borderline ill-conditioned problems; and not necessarily only
cases where very high accuracy is desirable. Please take a somewhat
farsighted view and consider this question in the context of your own
research.
The main reason I'm inquiring is that I'm engaged with IBM R&D on this
issue, and they are considering developing a processor that has
hardware-level support for quadruple floating-point precision.
Currently, they are developing a POWER based server with an
FPGA-accelerator that is tightly-coupled to the main CPU. The FPGA
could have the option of serving as a high-precision numerics
accelerator at the hardware level. In fact, such a system is already
operating in a number of IBM labs.
At this stage, it is crucial to provide IBM with some feedback on the
"market" for such a product, so that it can actually advance past the
research development stage. Therefore, I'm requesting some feedback
from you on this matter. Do you see yourself using quadruple-precision
numerics in the next few years? If so, could you briefly explain why?
Would you be interested in such a server / HPC? Do you know others who
could be interested? Please feel free to forward this note to others.
Thank you for your time.
Best wishes,
Gaurav
-----------------------------------------------
GAURAV KHANNA
UMass Dartmouth, Physics
(508) 910 6605
http://gravity.phy.umassd.edu/
"Black holes are where God divided by zero." - Steven Wright
_______________________________________________
Developers mailing list
[email protected]
http://www.cactuscode.org/mailman/listinfo/developers
