» Search forum


Index :: BOINCstats Cafe :: Shoutbox History

Moved from "Invites"

Pages: 1 ... 2111 2112 2113 2114 [2115] 2116 2117 2118 2119 ... 2483
DRESI
BAM!ID: 29387
Joined: 2007-06-27
Posts: 131
Credits: 116,944,570
World-rank: 10,605

2012-11-03 22:44:28

@JC: At my experiences is memory bandwith less important than # of CUDA Cores and Core Clock... and a fast CPU to feed them.
mmstick
BAM!ID: 77704
Joined: 2009-11-23
Posts: 10
Credits: 110,786,720
World-rank: 11,015

2012-11-04 00:45:55

@Dresi, don't leave out compute units for all those OpenCL projects. The need for a fast CPU is not so important in most.
7.62
BAM!ID: 101974
Joined: 2011-06-26
Posts: 156
Credits: 1,994,355,728
World-rank: 1,426

2012-11-04 02:06:32

mmstick is right about certain projects needing more CPU/RAM... I can't feed both cards POEM simultaneously.
7.62
BAM!ID: 101974
Joined: 2011-06-26
Posts: 156
Credits: 1,994,355,728
World-rank: 1,426

2012-11-04 02:08:19

Even with 30nm Samsung @ 2133 Mhz Dual Channel, i5 @ 4.5 GHz. 6 WU wide on ONE card is all in POEM.
7.62
BAM!ID: 101974
Joined: 2011-06-26
Posts: 156
Credits: 1,994,355,728
World-rank: 1,426

2012-11-04 02:12:10

I've seen great things from 660 Tis given their cost. Even with the crippled FP64 etc. No replacement for displacement!
7.62
BAM!ID: 101974
Joined: 2011-06-26
Posts: 156
Credits: 1,994,355,728
World-rank: 1,426

2012-11-04 02:17:02

A 192-bit bus on a 660 Ti is not a huge concern given its GDDR5 clockspeeds. A 670 will just address more pixels.
7.62
BAM!ID: 101974
Joined: 2011-06-26
Posts: 156
Credits: 1,994,355,728
World-rank: 1,426

2012-11-04 02:30:42

The bandwidth of a 192-bit 660Ti is comparable to a 256-bit 7850/70 or last gen 320-bit GTX 570.
JCMarsh
 
BAM!ID: 116564
Joined: 2012-02-07
Posts: 208
Credits: 28,030,979
World-rank: 27,876

2012-11-04 03:43:59

Cool. Thanks for your input. Now to convince the breadwinner Santa I really NEED a new GPU more than anything else
mmstick
BAM!ID: 77704
Joined: 2009-11-23
Posts: 10
Credits: 110,786,720
World-rank: 11,015

2012-11-04 23:36:06

Radeon HD 8000 series is right around the corner, and will bringing a lot of GPGPU computing enhancements.
mmstick
BAM!ID: 77704
Joined: 2009-11-23
Posts: 10
Credits: 110,786,720
World-rank: 11,015

2012-11-04 23:36:50

NVIDIA is well out of the race now, the GTX 600 series was the final nail in their coffin concerning GPGPU performance.
mmstick
BAM!ID: 77704
Joined: 2009-11-23
Posts: 10
Credits: 110,786,720
World-rank: 11,015

2012-11-04 23:37:17

All projects at BOINC from here on are shifting to OpenCL, where NVIDIAs drivers fail at.
Zingo
BAM!ID: 114124
Joined: 2011-12-30
Posts: 25
Credits: 268,799,696
World-rank: 6,152

2012-11-05 02:01:26

Is there any info on when OpenCL will be availalbe so i can crunch on my i5 and i7?
7.62
BAM!ID: 101974
Joined: 2011-06-26
Posts: 156
Credits: 1,994,355,728
World-rank: 1,426

2012-11-05 07:09:21

FP64 compute on mainstream cards? Yes, NV is out. They sell Teslas for that .
7.62
BAM!ID: 101974
Joined: 2011-06-26
Posts: 156
Credits: 1,994,355,728
World-rank: 1,426

2012-11-05 07:15:22

Some people still insist on NV. Single-precision is GTX Kepler's ace, and many popular projects (GPUGRID) only need SP
7.62
BAM!ID: 101974
Joined: 2011-06-26
Posts: 156
Credits: 1,994,355,728
World-rank: 1,426

2012-11-05 07:28:07

Zingo, check Intel's Win8 drivers for HD3000/HD4000. OpenCL 1.1 is listed. Such a nice standard language.
mmstick
BAM!ID: 77704
Joined: 2009-11-23
Posts: 10
Credits: 110,786,720
World-rank: 11,015

2012-11-05 08:39:24

GPGPU is not really Kepler's ace, even GTX 500 series has faster GPGPU capabilities. SP on AMD though, 5 TFLOPS is nice.
Zingo
BAM!ID: 114124
Joined: 2011-12-30
Posts: 25
Credits: 268,799,696
World-rank: 6,152

2012-11-05 10:56:29

7.62* I'm well aware of that, but afaik BOINC cant use it yet.
Dirk Broer
 
BAM!ID: 117974
Joined: 2012-03-01
Posts: 343
Credits: 6,564,229,963
World-rank: 574

2012-11-05 11:59:31

Nobody has written an application yet that can utilize the Intel HD4000/3000/2500/2000/1000s on Sandy/Ivy Bridge
Dirk Broer
 
BAM!ID: 117974
Joined: 2012-03-01
Posts: 343
Credits: 6,564,229,963
World-rank: 574

2012-11-05 14:15:54

@mmstick: WCG's HCC performs good on nVidia though, and is an OpenCL application.
7.62
BAM!ID: 101974
Joined: 2011-06-26
Posts: 156
Credits: 1,994,355,728
World-rank: 1,426

2012-11-06 05:12:19

I thought that's what you meant. Projects don't see benefit in coding for the Intel iGPU just yet.
7.62
BAM!ID: 101974
Joined: 2011-06-26
Posts: 156
Credits: 1,994,355,728
World-rank: 1,426

2012-11-06 05:22:31

They see performance per chip. Not the number of chips. 10 HD 4000s almost equal an HD7850 in SP.
7.62
BAM!ID: 101974
Joined: 2011-06-26
Posts: 156
Credits: 1,994,355,728
World-rank: 1,426

2012-11-06 05:34:02

GTX 5xx FP64 are 1/8 of SP capabilities. GTX 6xx, 1/24. SP is up almost 2x from GTX 5xx to GTX 6xx.
7.62
BAM!ID: 101974
Joined: 2011-06-26
Posts: 156
Credits: 1,994,355,728
World-rank: 1,426

2012-11-06 05:36:47

Only real handicap is FP64, and performance in projects is more situational than anything. Yes, GCN is the better GPGPU arch.
mmstick
BAM!ID: 77704
Joined: 2009-11-23
Posts: 10
Credits: 110,786,720
World-rank: 11,015

2012-11-06 06:19:32

@Dirk Broer NVIDIA's performance on HCC is very abysmal, It consumes more CPU and is significantly slower.
mmstick
BAM!ID: 77704
Joined: 2009-11-23
Posts: 10
Credits: 110,786,720
World-rank: 11,015

2012-11-06 06:26:37

My 7950 has a kernel time of 18 seconds with HCC, now try and find a NVIDIA GPU that is at least 1/4th that fast.
Pages: 1 ... 2111 2112 2113 2114 [2115] 2116 2117 2118 2119 ... 2483

Index :: BOINCstats Cafe :: Shoutbox History
Reason: