Double precision GPUs coming very soon

dsouza123 · 2007-10-13, 13:56

Double precision hardware support is coming soon from both ATI and nVidia GPUs

RV670 aka the Radeon HD 2950: November 19 (formerly November 15)
G92 aka the GeForce 8800 GT: October 29 (formerly November 12)

We learned that G92 and RV670 are sporting dual-precision units
(128, 112 or 96 to be available in G92, RV670 will have 320 units),
albeit it is unknown what type of Binary Floating-Point formats are these units supporting.
Who is supporting IEEE 754 format, 32-bit or 64-bit - time will tell.

http://www.theinquirer.net/gb/inquir...gpgpu-monsters

dsouza123 · 2007-10-13, 18:59

The following link gives more general information about the ATI RV670 cards
including an image of an AMD slide with a bullet point listing double precision floating point.

http://www.xbitlabs.com/news/video/d...930223300.html

paulunderwood · 2007-10-13, 22:39

This is simulated double precision.

Ati has not released their CTM (close to metal) compiler. It is vaporware.

Dresdenboy · 2007-10-14, 20:39

My assumption is, that they implemented DP in a way similar to what has been done in Cell's SPEs. Full featured DP with the same throughput as SP calculations or even half the throughput (filling the registers with 2 doubles instead of 4 singles) would cost a few hundred million transistors more and need a lot of power.

A software emulation via driver/CTM compiler would probably be far from being useful (since this could have been done already).

dsouza123 · 2007-10-15, 02:20

CUDA 1.0 ( C for NVidia GPUs) now documents the PTX assembly language.

http://www.behardware.com/articles/6...ical-uses.html

----------------
The CUDA Developer SDK provides examples with source code to help you get started with CUDA.
Examples include:
Parallel bitonic sort
Matrix multiplication
Matrix transpose
Performance profiling using timers
Parallel prefix sum (scan) of large arrays
Image convolution
1D DWT using Haar wavelet <-- is this like what Prime95 uses ?
OpenGL and Direct3D graphics interoperation examples
CUDA BLAS and FFT library usage examples
CPU-GPU C and C++ code integration

http://developer.nvidia.com/object/cuda.html

NVidia is also releasing a Tesla line of GPU cards,
calculation accelerators without video connections.

The base model C870 has 1.5 GB memory and a price to match $1299.
---------------------------------------------------------
Correction: The G92 GPU will be in the GeForce 9800 cards.

2007-10-13, 13:56	#1
dsouza123 Sep 2002 662₁₀ Posts	Double precision GPUs coming very soon Double precision hardware support is coming soon from both ATI and nVidia GPUs RV670 aka the Radeon HD 2950: November 19 (formerly November 15) G92 aka the GeForce 8800 GT: October 29 (formerly November 12) We learned that G92 and RV670 are sporting dual-precision units (128, 112 or 96 to be available in G92, RV670 will have 320 units), albeit it is unknown what type of Binary Floating-Point formats are these units supporting. Who is supporting IEEE 754 format, 32-bit or 64-bit - time will tell. http://www.theinquirer.net/gb/inquir...gpgpu-monsters

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
What percentage of CPUs/GPUs have done a double check?	Mark Rose	Data	4	2016-06-17 14:38
x.265 half the size, double the computation; so if you double again? 1/4th?	jasong	jasong	7	2015-08-17 10:56
translating double to single precision?	ixfd64	Hardware	5	2012-09-12 05:10
Fast double precision Division	__HRB__	Programming	21	2012-01-10 02:10
double precision in LL tests	drew	Software	4	2006-08-08 04:08

2007-10-13, 18:59	#2
dsouza123 Sep 2002 2×331 Posts	The following link gives more general information about the ATI RV670 cards including an image of an AMD slide with a bullet point listing double precision floating point. http://www.xbitlabs.com/news/video/d...930223300.html

2007-10-13, 22:39	#3
paulunderwood Sep 2002 Database er0rr 7233₈ Posts	This is simulated double precision. Ati has not released their CTM (close to metal) compiler. It is vaporware.

2007-10-14, 20:39	#4
Dresdenboy Apr 2003 Berlin, Germany 169₁₆ Posts	My assumption is, that they implemented DP in a way similar to what has been done in Cell's SPEs. Full featured DP with the same throughput as SP calculations or even half the throughput (filling the registers with 2 doubles instead of 4 singles) would cost a few hundred million transistors more and need a lot of power. A software emulation via driver/CTM compiler would probably be far from being useful (since this could have been done already).

2007-10-15, 02:20	#5
dsouza123 Sep 2002 2·331 Posts	CUDA 1.0 ( C for NVidia GPUs) now documents the PTX assembly language. http://www.behardware.com/articles/6...ical-uses.html ---------------- The CUDA Developer SDK provides examples with source code to help you get started with CUDA. Examples include: Parallel bitonic sort Matrix multiplication Matrix transpose Performance profiling using timers Parallel prefix sum (scan) of large arrays Image convolution 1D DWT using Haar wavelet <-- is this like what Prime95 uses ? OpenGL and Direct3D graphics interoperation examples CUDA BLAS and FFT library usage examples CPU-GPU C and C++ code integration http://developer.nvidia.com/object/cuda.html NVidia is also releasing a Tesla line of GPU cards, calculation accelerators without video connections. The base model C870 has 1.5 GB memory and a price to match $1299. --------------------------------------------------------- Correction: The G92 GPU will be in the GeForce 9800 cards.