ATI said that they made the decision not to include double precision in their lower end 57xx series in order to save die space and because they didn't feel there was a demand for it in those cards. The only real change was that the 4770 did include double precision, but that was in many ways a unique product in that it was made to test the viability and design of their new 40nm process. All the 48xx, 58xx and 59xx series consumer cards support fully enabled double precision and do it very efficiently compared to anything else on the market.

GDF, you're right: ATI does dp using the same shaders as sp. However, to couple them for dp some clever control logic is needed. That's what they're saving by not including it in the mainstream chips. It's an economic decision as well as a marketing decision to drive crunchers to the faster cards. Which are more energy efficient at this job anyway.

Fractal, there's no need for a "conspiracy theory" here. To begin with, if anything is labeld like that it's considered wrong by definition.. which is not the case here ;)
But on a more serious note: it's just a business decision made by nVidia, without technical reasons or advantages. Their official marketing stance on it is pure BS and the real reason is pushing Tesla sales. There's no conspiracy here, they're free to cripple their product as much as they want. It's just that we crunchers think it's a bad move, short and long term.

Regarding the issue of increased reliability due to slower clock speeds: sp and dp are done by the same shaders. And the GTX 4x0 chips don't have dp disabled, it's just that less of their shaders are allowed to run in dp mode at any time. So which ever clock speed penalty might exist in dp mode, the Geforce models also receive it.
But there's more. As I explained here chips do degrade over time. Seldomly they suddely break, but they're always affected by continous decay, i.e. the maximum clock speed they can reach drops. With this knowledge it's easy to see how nVidia increases reliability by using lower clock speeds: if you take similar chips and run them under similar conditions (same voltage & temperature), they'll decay in the same way. Assume your chips can do 1.4 GHz and lose 0.1 GHz per year on average (*), then you could run them at 1.2 GHz for 2 years until you start to receive major RMAs because the cards start to fail. If you run the chips at 1.1 GHz you'll get another year, so you can either extend your warrenty or reduce the return rate considerably.

MrS

(*) This number would be different for every single chip and the decay rate wouldn't be constant over time, so this is just a Gedankenexperiment.

---

Scanning for our furry friends since Jan 2002

Here's a GTX 470 at Collatz (integer math):

http://boinc.thesonntags.com/collatz/show_host_detail.php?hostid=20977

Looks like it's about 55% of the speed of the HD 5870 (around the speed of a $150 HD 5770). Someone else with a GTX 470 confirmed the same results and also tried a GTX 480, OCed it's about 90% of the speed of a stock HD 5850. Here's the thread:

http://boinc.thesonntags.com/collatz/forum_thread.php?id=378&nowrap=true#7162

Edit: Here's a GTX 480 on GPUGRID, so far it's running a little slower (and less reliably) than my GTX 260:

http://www.gpugrid.net/show_host_detail.php?hostid=35174

Edit 2: Two guys are trying to get their GTX 470 cards to run at MilkyWay (double precision), so far with no luck. Will post an update when they get them working.

Horses for courses!
A speedboat is fast on the water and an F1 car is fast on a track. They don’t swap well, and Fermi is just out of the garage for its first spin.
It will work well here when the team releases more ACEMD (V1) tasks or writes, tests and releases an app for Fermi cards. We already know ACEMD Ver 2 task will not work and why they will not work (The app is specifically/rigidly written for earlier GPU core architecture), and we already know ACEMD tasks will work (as this is not the case; albeit at the expense of some speed improvements introduced with V2). So there is the short term and long term solutions in a nut shell.

I have now added the cuda3 beta application again for Fermi based cards only.

gdf

I have tried for 30 minutes every 8 seconds make connection with gpugrid for downloading any beta3 WU.
The reply is - no work available!!!

Can i do anything about it?

thx for helping out!

Can i have the same problems with my new GTX 470 card, which i want to install tomorrow? Should i wait with installing??

Ton (ftpd)

---

Ton (ftpd) Netherlands

The only application that you have some chances to see working on Fermi is the beta application. You should select in the preferences to just crunch for that at the moment.

Until we get a Fermi here we cannot really debug it.

Do you have the latest driver installed?

I have submitted 50 beta WUs but there all out now, so just wait to get one.

gdf

I installed driver 197.41

I will try tomorrowmorning again!

---

Ton (ftpd) Netherlands

just downloaded 1 beta-WU (cuda30) - # 1359859.
Cancelled after 15 seconds.

---

Ton (ftpd) Netherlands

I wonder if some1 who got fermi, can put here compute capability (in TFLOPS).

---

I wonder if some1 who got fermi, can put here compute capability (in TFLOPS).

They've already been posted in the next-door thread:

NVIDIA GPU 0: GeForce GTX 470 (driver version 19741, CUDA version 3000, compute capability 2.0, 1280MB, 1089 GFLOPS peak)
NVIDIA GPU 0: GeForce GTX 480 (driver version 19741, CUDA version 3000, compute capability 2.0, 1536MB, 1345 GFLOPS peak)

I wonder if some1 who got fermi, can put here compute capability (in TFLOPS).

They've already been posted in the next-door thread:

NVIDIA GPU 0: GeForce GTX 470 (driver version 19741, CUDA version 3000, compute capability 2.0, 1280MB, 1089 GFLOPS peak)
NVIDIA GPU 0: GeForce GTX 480 (driver version 19741, CUDA version 3000, compute capability 2.0, 1536MB, 1345 GFLOPS peak)

The flops value reported by BOINC means close to nothing. They are indicative only within the same GPU version (G200, Fermi). Across GPU core versions and compared to ATI cards they don't mean anything.

gdf

We have ordered one of these hosts without GPUs.

gdf

The above ships with 4 Teslas and costs as much as a new car.

You would be better off just buying a good PSU and getting a couple of Fermi cards (or a new sysetem, or two, or three...).
Two GTX480 cards will probably do as much work as 4 or 5 GTX295s, ie more than 4 Teslas!

Have you considered the cost effectiveness of using 470s instead 480s?

I know having the the best is going to return the fastest results but at US price ratios you can get 8 x 470s for the price of 5 1/2 x 480s.

---

Thanks - Steve

Once paid for the host, there is no point of saving into the GPUs.
Plus this is to go as fast as possible and also the cooling seems better to me on the GTX480.

GDF

I wonder if some1 who got fermi, can put here compute capability (in TFLOPS).

They've already been posted in the next-door thread:

NVIDIA GPU 0: GeForce GTX 470 (driver version 19741, CUDA version 3000, compute capability 2.0, 1280MB, 1089 GFLOPS peak)
NVIDIA GPU 0: GeForce GTX 480 (driver version 19741, CUDA version 3000, compute capability 2.0, 1536MB, 1345 GFLOPS peak)

THX :-)

---

For BOINC this is a really bad move. There are projects which could really benefit from wide spread fast double precision capabilities. But for the existing projects it's not that bad:

GPU-Grid...

Your list forgot to include DNETC@Home which supports both ATI and Nvidia.

That's probably because I don't find their science very compelling ;)
Do you know if they use integer or single or double fp?

MrS

---

Scanning for our furry friends since Jan 2002

That's probably because I don't find their science very compelling ;)
Do you know if they use integer or single or double fp?

MrS

The CUDA variant uses integer instructions. There should be no need for floating point calculations in a rc5-72 known plaintext attack.

To prevent some misconceptions: A known plaintext attack doesn't mean that the message is known.

That's probably because I don't find their science very compelling ;)
Do you know if they use integer or single or double fp?

DNETC are Single Precision

Regards
Zy