New GPU-Client with NV support: 6.12 beta

Does this work along side the SMP client?

 

Theoretically it should - creating around 10% CPU load on my C2D. But then, I'm running Vista x64 - and the SMP-client doesn't like that atm.

edit:
fixed - both are humming along nicely together. The SMP-client does not seem overly impaired by having the CPU feeding the GPU for the other client and still is supplying about 1.700 ppd.

 

OK. I'm getting the same error on a 4 core Vista Home Premium (32bit) box. Any ideas?

 

Tried the modified 177.39 drivers which were leaked a few days ago?

 

Yep.

 

Dunno about 177.39. I'm using 177.35 with a hacked .inf that guru3d did to support G80 and Quadros.

 

I upgraded to the modfied drivers and the UNSTABLE_MACHINE error has gone away. Now it's being replaced by Folding@home Core Shutdown: EARLY_UNIT_END. That and FILE_IO_ERROR. Is this normal for the beta?

And it's running fine along side the SMP client with no noticable sow downs.

Also why does there appear to be 2 instances of FahCore_11 running?

One last thing, will the folding client ever end up running on a PhysX card? Since it's basically a stream processor like a GPU?

 

Try running it as "run as administrator" if you're on Vista. That sounds like an authority problem.

 

Seems to be working fine now, no changes. 6 units done... faster than the PS3 but I guess you get less points for these?

 

I've just looked at the extreme overclocking site and realised that the drop in my output coincided with me playing MGS4. Freaky stats reading!!

 

A lot of the time you can get rid of these sorts of issues by just adjusting the permissions in the program folder so that it's no longer "read only". Dunno if that'll work with folding at home, though.

 

Great perf numbers for GPU client on G92 also with SMP client on dual-core 4GHz Penryn. Really juicy stuff here.

4GHz Penryn (E8400 @ 445x9)
4GB RAM @ 1GHz
8800 GT @ 685/1688/1940 on 177.35s - XP

4850 PPD for GPU client on its own. Combined with CPU SMP client (3600 GPU, 2300 CPU) 5900 PPD. Tweaked affinity and priorities brings this to 6450 total PPD (4550 GPU, 1900 CPU).

CPU SMP client all-out without GPU client running yields approximately 3000 PPD, so clearly a mix and match of the two clients yields the best results.

Temps don't really seem a concern. I have 6x 120mm case fans, a Xigmatek 964 HSF in push-pull config (dual 92mms full-out i.e. all thermal controls off), and the 88gt gets the TT duorb treatment with fans @ 70% via rivatuner.

 

The NV GPUs slows down with larger proteins.

Some selected quotes:

---

by DanEnsign on Mon Sep 08, 2008 6:56 pm

Hi,

I'm warning you, you will start to see a new series of projects coming out this morning. The proteins we're studying here are larger and in some sense more relevant to biology than anything we've run on GPU2 to date.

However, due to the larger size of the protein, you may see a PPD drop putting NV cards more in line with ATI production. This isn't due to different benchmarking or a different way to assign points, but rather we anticipate NV to be slightly less efficient with larger proteins than it has been with the smaller ones.

---

by John Naylor on Mon Sep 08, 2008 9:11 pm

CoreX5 wrote:Why if these projects are larger, are they worth less? Also why is the benchmark machine an ati 3850 when the same projects are not being run on nvidia as the ones being run on Ati clients? Just seeking some clarification as this could look like it was changed just to appease ati and their clients.
​

They are worth less because despite being bigger the benchmark machine processes them quicker than the old units (I think the simulations are shorter, just the proteins are bigger). The benchmark machine has always been a 3850 so nothing has changed.

And as an aside, the project staff have repeatedly said that nvidia's huge lead in the PPD stakes may change. It now has (I'm an nvidia folder too so I would be as miffed as anyone were it not for the fact that I don't give two hoots about points), so you can't say we weren't warned

---

by toTOW on Mon Sep 08, 2008 10:32 pm

There is no point about calling it favoritism : we have always said that the first batch of WU (with ~550 atoms) was doing very well on the NV hardware and that might not be the case with future WU (see the "Preliminary results" in the PPD thread title).

It's just a matter of architecture : NV chose to use a few amount of complex shaders which are fast, but few. ATI chose simpler shader, which are a bit less efficient, but a lot.

On NV architecture, more atoms (and thus complexity and parallelism) means that there are no free SP to use, as they were already used hence the PPD drop.

On ATI hardware, there are still plenty SP available, so more complexity means better performance, as there were SP available.

We still see performance difference between two architectures because of SP efficiency, but both hardware tend to join their performances : NV slows a bit, and ATI gets a bit faster.


---

by slegrand on Mon Sep 08, 2008 10:50 pm

No worries, there are plenty of NV optimizations in the pipe as well, but I need to validate them before turning them loose and right now, I'm swamped with some other F@H-related work. As things stand, the NV algorithm is more efficient but the ATI algorithm is simpler. Hence around lambda-sized molecules, ATI gets a perf spike while NV is still warming up and bogged down by the housekeeping incurred for its more efficient algorithm. For Fip35-sized molecules, NV is underutilized on GTX280, but ATI is doubly so on 4870 and that's why the perf differential there is just crazy.

In the long run in this GPU generation, I *expect* SP for SP that an NV GPU will be about 2.5x faster than an ATI SP. This is of course then modified by SP count and processor clock. If Mike finds a structural optimization to his code, then I will revisit that ratio, but that's my prediction for now. And the game resets with the next hw generation anyway so it's an ongoing arms race.

 

Yup, all my Nvidia cores have received non-supervillins now and PPD is down across the board. Some rough preliminary PPD results from my FahMon;
GTX280 down from 6.8k to 5.7k
G92 GTS went from 4.8k to 3.4k
G80 GTS from 3.4k to 2.1k
G84 GT from 1.6k to 0.8k

But who could complain? The Nvidias have benefitted from excellent scores for a long time and those are still some great numbers. And after all it's all about the science

 

Anyone else with this problem? Launching Process Explorer while this client is running crashes the system with an nvdisp4.dll error.

I am running the 182.05 betas

 

Correct me if I'm wrong, but is there not a newer client available?

 

Oh, yes. I downloaded the latest NV GPU client 6.23. It is with it I have the problem.