Anyone got experience of the nVidia (Fermi) 400 series GPUs for Boinc crunching?

How good compared to the 200-series?

Is this problem with slow glReadPixels() (1/2 to 1/4 speed of 200-series) of any concern?

Go for a 1 GByte or 512 MByte version?


Happy crunchin',
Martin

---

See new freedom: Mageia Linux
Take a look for yourself: Linux Format
The Future is what We all make IT (GPLv3)

My GTX465 is exactly as fast as my GTX280.

And needs the same power. :D

Helli

---

A loooong time ago: First Credits after SETI@home Restart

I use the gtx470 & gtx480 with stock application.
Works OK. About 8 minutes for one WU.
Also use GTX295.
See my profile.

Ton (ftpd) Netherlands

---

If you're on a budget I found the GT240 in any variant (512 or 1GB, DDR3 or DDR5 RAM) is the best solution if you don't want to spend extra also having to increase the power of your PSU (if it's a standard ~300W PSU).

My understanding is you can go for a Fermi GTX400-series but won't get the full benefit until Lunatics release a version of their s/w which fully utilises the features of the Fermi architecture. Am I right?

My understanding is you can go for a Fermi GTX400-series but won't get the full benefit until Lunatics release a version of their s/w which fully utilises the features of the Fermi architecture. Am I right?

Pretty much :)

Anyone got experience of the nVidia (Fermi) 400 series GPUs for Boinc crunching?

How good compared to the 200-series?

With current code, everything at GTX 260 or above is proving fairly memory bound, with both on-Card GPU memory, Host Memory, and PCIe Bus transfers of memory. The effect tends to 'level' the performance of all these cards toward an artificial ceiling. with refinement 200 series still have legroom, but tending to cap out sooner, 400 series are hinting they have a lot more potential to unlock.

Hiding & minimising these accesses & transfers is becoming a priority in development, and in unit tests tends to open up a little legroom with 200 series, and 'quite a lot' with Fermi architecture.

Is this problem with slow glReadPixels() (1/2 to 1/4 speed of 200-series) of any concern?

No, There are faster alternatives on Fermi to using Texture memory, which was a common 200 series optimisation.

Go for a 1 GByte or 512 MByte version?

1536 ;)

Jason

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

My GTX465 is exactly as fast as my GTX280.

And needs the same power. :D

WHAT?!!!

352 cores vs 128 cores give the same performance?!!!

No wonder nVidia don't list the FLOPs rating for the GTX465!


Is the 400 series artificially crippled or a design disaster?


Happy crunchin',
Martin

---

See new freedom: Mageia Linux
Take a look for yourself: Linux Format
The Future is what We all make IT (GPLv3)

My understanding is you can go for a Fermi GTX400-series but won't get the full benefit until Lunatics release a version of their s/w which fully utilises the features of the Fermi architecture. Am I right?

Pretty much :)

Ahhhh...

So any clues as to how well the 400 series can perform?


Happy crunchin',
Martin

---

See new freedom: Mageia Linux
Take a look for yourself: Linux Format
The Future is what We all make IT (GPLv3)

Anyone got experience of the nVidia (Fermi) 400 series GPUs for Boinc crunching?

How good compared to the 200-series?

Is this problem with slow glReadPixels() (1/2 to 1/4 speed of 200-series) of any concern?

Go for a 1 GByte or 512 MByte version?


Happy crunchin',
Martin

Hi Martin,

Mine are roughly on a par as well. I have a GTX460 from EVGA (slightly overclocked) and a GTX260 (slightly overclocked from Gigabyte's overclock) and they still post roughly the same time per wu (Windows 7-64).

The 460 looks like it isn't doing much and doesn't get anywhere near as hot as the 260. It also downclocks much better when not used - it runs at 825/1650/1900 (core/shader/memory) when running full tilt, and reduces to 405/810/162 at medium setting and then reduces to 50/101/67.5 at lowest setting. So it can reduce it's power consumption quite a bit when only browsing web, etc. and other low power scenarios.

Smallest memory is 768Mb and other is normally 1Gb (on the 460). Gamers find the 768Mb a bit limiting, the smaller memory slows the card down in games. Not sure about the other models, I don't have any.

The 470's commonly have 1280Mb and the 480's commonly have 1536Mb. Not sure about the 450's.

I find it a good mid-range card, but obviously all the prices will come down now the 570's and 580's are out. It won't be long before they knock out a 560 and 550 too, I bet.

Hope that helps a bit.

Giz.

---

A proud GPU User Server Donor!

My understanding is you can go for a Fermi GTX400-series but won't get the full benefit until Lunatics release a version of their s/w which fully utilises the features of the Fermi architecture. Am I right?

Pretty much :)

Ahhhh...

So any clues as to how well the 400 series can perform?


Happy crunchin',
Martin

One thing I can say is that I wouldn't trade my two 480's for anything except four 480's. I totally love them. Occaisionally I try a game when my limited time allows, and they will burn through anything out there with ease.

With running two WU's at a time, I topped out around 47,000 RAC. Of course there was some water cooling and overclocking going on....

Steve

---

Warning, addicted to SETI crunching!
Crunching as a member of GPU Users Group.
GPUUG Website

So any clues as to how well the 400 series can perform?

Well my 480 does 3 mid angle range tasks at once in ~18 minutes, so 6 minutes per task. (with x32f build which is mostly cleaned up stock code to reduce error rate a little, with some very minor optimisations)

at ~100 credits each I make that about 240*100 about 24000 credits per day attributable to this card, which more or less gels with my peak RAC before the servers went down ( ~26k , dual core core2 CPU )

Unclear as to how much further they can be pushed, but general (non-fermi specific) cases seem to extract around ~30% throughput improvement on 200 series for worst case workloads , while similar workload/optimisations measure ~60% upwards for Fermi (greater optimisation return for generic optimisations)

That's without Fermi specific optimisations yet, so time will tell.

Jason

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

My 480 runs through a work unit every 7 minutes on lunatics current release. My 250 takes about 19 minutes with the same optimized apps. I love my 480 wouldn't trade it for nothing, well right now.

---

Traveling through space at ~67,000mph!

I have to agree I love my GTX480's. Both for Seti and for gaming. I have after market coolers on each of them and overclocked them. Running 3 units per GPU takes 18mn (6mn per unit) and about 850M ram out of 1536M total with temps hitting a max of 62C. With 2 cards running and the i7 920 running at a hair over 4G my power draw hovers right around 700 watts.

---

Evga Gtx 460 768 version overclocked to 877/1754/1970 Core/Shader/Mem Clock
pretty much limit of card right now as my PS isnt big enough..

Crunching in 9 min flat!

Anyone got experience of the nVidia (Fermi) 400 series GPUs for Boinc crunching?

How good compared to the 200-series?

Is this problem with slow glReadPixels() (1/2 to 1/4 speed of 200-series) of any concern?

Go for a 1 GByte or 512 MByte version?


Happy crunchin',
Martin

I just got an eVGA GTS 450 Superclocked. 1GB DDR5 RAM.

As far as bang for the buck goes, this card is really hard to beat. It blows through WU's in 12 minutes or less!

For gaming it's really great. DX11 performance is really, really good and I'm able to play most all games at max or very close to max settings.

Cost me $129. Also, power consumption is very low for a card this powerful and it's super quiet.

My first rig with a 480 just came online.

triggerfish

I can say my 480 runs REAL hot. Under load with Lunatics latest - the GPUs are running
about 88C (about 190F) underload. Room is 75F. I have 1 card running in a very well ventilated
case and the rest of the rig is water cooled. Fan on card is at 65% 2730RPM this card
gets so warm I can't touch the connector plate on the back of the rig. Wonder if others
see the same that have 480s (I should have sprung for the waterblock version of the card).

---

I've got a GTX 460, I've modified the app info file and set it to do 3 tasks at once...it does the longer CUDA wu's in about 32 minutes max (I've also installed the Lunatics Opti-Apps v.37). I've got the MSI TWIN FROZR II, so my GPU is set at 850 core 1900 shaders and doesn't hit 50C...I keep my fan on 72%

Give Fermi a few months so that the boys over at Lunatics get the Fermi code better worked out, these Fermi cards are going to be screamers!

Bang for buck, you can't go wrong with a well-cooled GTX 460, and if you can hold out i expect the GF114 GTX560 to be excellent after it's initial hype dies out and prices start to moderate. I think Q1 2011 shall be quite interesting for us GPU users.

---

With water cooling my two 480's I am at the following settings.

Core 860 MHz
Shaders 1720 MHz
Memory 1975 MHz
Temps 46 to 50°C for each card.
Running 3 wu per card.

There is a financial cost with setting up water cooling, but the above is an easy example of how well it works.

Steve

---

Warning, addicted to SETI crunching!
Crunching as a member of GPU Users Group.
GPUUG Website

After due deliberation... ;-)


Looks like I'll be jumping in to get a:

Palit 1GB GeForce GTS 450 Sonic DDR5 NVIDIA Graphics Card

1GB Palit GTS 450 Sonic, 40nm, 3900MHz GDDR5, GPU 880MHz, Shader 1760MHz, 192 Cores, 2x DVI-I/ HDMI


That looks to be the best balance of performance vs temperature vs cost for my usage and to add some compute power to the big crunch.


Regarding the unexpectedly slow glReadPixels() ...

It looks like the 'consumer' nVidia Fermis are throttled back to 1/4 the speed of what the Fermis are capable of. Has the glReadPixels() function been clobbered by that?

From comments elsewhere on the web:

... FrameBuffer<-->MainMemory transfers ... is where the problem seems to be. Reading the citations however, it looks like you can workaround the problem by using a pixel buffer object transfer instead, which is a DMA transfer (as opposed to a CPU mediated transfer). ...

Is that any help for s@h crunching?


But then also, the 'consumer' Fermis are a lot cheaper than the 'Tesla' marketed 'GPGPU' cards. At first glance, the deliberate performance throttling might look like an outrage and a heinous anathema to all the optimisers amongst us... But then again, it makes very good business sense to nVidia to have just the one design and one chip to go through fab that can then span their entire market. It means we can get chips designed for the 'pro' market yet that run cooler and are very much cheaper than 'pro' pricing and heat.

Who knows... Some optimisers may find how to 'unlock' the full performance... (However, I suspect that the onboard PSU and cooling for the consumer cards wouldn't be able to handle the extra (x4) supply current needed...)


Happy crunchin',
Martin

---

See new freedom: Mageia Linux
Take a look for yourself: Linux Format
The Future is what We all make IT (GPLv3)

Martin,

Were you intending to use the GeForce GTS 450 under Linux or Windows?, as the there isn't a Fermi compatible Seti Cuda app for Linux yet,

Claggy

Martin,

Were intending to use the GeForce GTS 450 under Linux or Windows?, as the there isn't a Fermi compatible Seti Cuda app for Linux yet,

Claggy

The card could be used under Linux at GPUGrid, and probably other BOINC projects too.