GPU2 4850 PPD Numbers

[EvilAlchemist]

System:

Q6600 @ 3.4 Ghz (9 x 378 Mhz)
Abit IP35-Pro
2GB DDR2 @ 378 Mhz
5 x 750GB Hard Drives
2 X 4850 GPU's
444 Watts pulled 100% Full GPU and CPU Load


GPU Client 1 - 1442.50 PPD

[02:26:39] Project: 4726 (Run 3, Clone 1, Gen 49)
*SNIP*
[02:26:51] Completed 19%
[02:28:52] Completed 20%
[02:30:49] Completed 21%
[02:32:45] Completed 22%

GPU Client 2 - 1437.52 PPD

[02:26:40] Project: 4731 (Run 2, Clone 27, Gen 20)
*SNIP*
[02:26:52] Completed 11%
[02:28:53] Completed 12%
[02:30:51] Completed 13%
[02:32:49] Completed 14%

SMP Client - 1544.58 PPD

[02:27:04] Project: 2665 (Run 3, Clone 439, Gen 16)
*SNIP*
[02:45:41] Writing local files
[02:45:41] Completed 190000 out of 250000 steps (76 percent)
[03:03:35] Writing local files
[03:03:35] Completed 192500 out of 250000 steps (77 percent)

Lost about 200 PPD on the SMP client
But gained about 1400 PPD from the GPU Clients

As has been stated in the F@H Forum, the 4850 is not using all of it's performance as of yet. This is due to the 2.5 x increase in shaders from the 38XX series.

I am sure they will get this fix in the near future.

 

[EvilAlchemist]

After Some Overclocking on the these new ATI Cards ....

GPU Clock Speed is what is important.
Memory Clock seems to have no impact.

I Overclocked the GPU to 700Mhz and lowered the memory to 750Mhz.
(Stock Settings - GPU 625 - Memory 1000Mhz)

Increased PPD by 100 per client.
(-200 from SMP / +200 from OC'ed GPU)

No change in GPU Temps

Boxen Totals = 1520PPD (GPU1) + 1496PPD (GPU2) +1534PPD (SMP) = 4550 PPD

 

[metallicafan]

Hmm interesting. If BB would have honored their sale and I would compare performance with you. But no such luck. Thanks for the info.

How warm do they run?

 

[EvilAlchemist]

How warm do they run?

Don't let this scare you - 82C Full Load ...

But this seems about right due to all the reviews I have seen.

One thing I did notice is that even at idle they are around 70C.

But again ... this seems to be common and the card has been designed to run at this temp.

The Fan runs at a constant 35% ... which is again the norm ppl are reporting.

I did just read an article over that XtremeSystems .. and they stated that using some AC Silver will drop temps about 20C .... I will look into this!

 

[EvilAlchemist]

Update:

Once I left my system alone for a while, the SMP went back up to 1981.94 PPD on WU 2665..

 

[APOLLO]

So, you are running two clients in one system without issues. BTW, did you purchase the 4850s because of gaming or strictly folding?

 

[metallicafan]

Don't let this scare you - 82C Full Load ...

But this seems about right due to all the reviews I have seen.

One thing I did notice is that even at idle they are around 70C.

But again ... this seems to be common and the card has been designed to run at this temp.

The Fan runs at a constant 35% ... which is again the norm ppl are reporting.

I did just read an article over that XtremeSystems .. and they stated that using some AC Silver will drop temps about 20C .... I will look into this!

Yeah this is about what my 3870 runs with the fan at stock speeds. I use rivatuner to control the fan speed to about 50 - 60% and then it runs at a cool 55-60°C. I would bet that Rivatuner might not be able to control the fan speeds of the 4800s yet but it might be worth a shot.

 

[metallicafan]

So, you are running two clients in one system without issues. BTW, did you purchase the 4850s because of gaming or strictly folding?

Hopefully he bought them for a little of both but at $150 a piece it really doesn't matter. These things game similar than(if not better than) a 9800GTX and I think that given some bigger WUs these things will be folding beasts!

 

[digital_exhaust]

Interesting.....looking forward to seeing what the future will hold for these cards.....

Nice increase in production though!!

How did you go about OC'ing the cards?

 

[OhSoCheesy]

Apollo, I run 2 console instances on my work laptop with core 2 duo proc with no problems. Sometimes if I am going to have a lot open (photoshop, other proprietary programs, etc) I will close one instance just to keep temps down. The Dell Lat D620 has terrible cooling, so I have to be careful of heat.
EA, have you had time to game with these cards? Are they what you expected? I've been in the market to build some new boxen and so i've waited to see how the video cards would play out. So I can afford either 2 AMD cards or 1 nVidia card. What do you think is the all around best solution?

 

[schenks]

should be interesting to see if ATI can bring their driver performance up to match the similar nVidia cards, because these should be hitting at least closer to the 4000-5000 PPD mark that the 8800 GT and higher are hitting.

 

[APOLLO]

should be interesting to see if ATI can bring their driver performance up to match the similar nVidia cards, because these should be hitting at least closer to the 4000-5000 PPD mark that the 8800 GT and higher are hitting.

Yes, and I'm curious about the CPU utilization as well. Should be very interesting to see whatever the results are.

 

[EvilAlchemist]

So, you are running two clients in one system without issues. BTW, did you purchase the 4850s because of gaming or strictly folding?

100% for Folding .... but i was not planning to buy any ....
Since I got them for $150 each .. seems like a good time to upgrade.

I did have a 1950XTX and a X1650 in the system.

 

[OhSoCheesy]

what are you going to do with those cards EA?

 

[EvilAlchemist]

Are they what you expected? So I can afford either 2 AMD cards or 1 nVidia card. What do you think is the all around best solution?

To be honest, I am not sure what the best solution is as of yet.
It really all depends on the progress of the ATI drives and the scaling on the cards.

I have not had a chance to game on them. Once I make sure the clients are stable, I will let you know, but as of right now, the nVidia client - esp the new 280's are kicking out 8K ppd)

what are you going to do with those cards EA?

My old 1950 XTX is for sale on Ebay.
I am going to keep the X1650 .. always good to have an extra video card for another boxen later down the road.

 

[OhSoCheesy]

Evil, where did you find those cards for $150? The cheapest I've seen is $170 after MIR. I also think I'll wait another week until the curtain is lifted on the 4870.

 

[BinarySynapse]

Evil, where did you find those cards for $150? The cheapest I've seen is $170 after MIR. I also think I'll wait another week until the curtain is lifted on the 4870.

I think he got in on the Best Buy 25% off deal that ended last weekend.


So the Nvidia cards are producing more points than the ATI ones it seems. Will this be possible to match up with drivers alone? Or is there some other reason for it?

 

[Xilikon]

There are some discussion on FCF about why ATI is doing less than NVIDIA despite the difference in the number of shaders and so far, people think it's due to the difference in the way shaders is configured : NVIDIA have 128 shaders (as example) but it's multipurpose so it mean it can do any operation and folding will take advantage of all of them. ATI OTOH have much more, 320 shaders as example for a equivalent card, but only 1 out of 5 is all purpose while the rest is only for integer additions (make sense from a graphics view where most of the action is 3d objects shifting, where those shaders will be useful). Only 64 out of 320 is all purpose so folding is only being able to tap into them.

It show clearly in the folding performance difference between a 8800GT and 3870 with the former being double the speed than the latter with F@H. The 4850 have 800 shaders but if we use the same 1/5 rule, there is 160 actual shaders. It should perform better than a 8800GT but it's currently not fully used and future tweaks will make it better for sure, along with bigger atoms. A GTX 260/GTX 280 will still kill them with 192/240 shaders.

Keep in mind it's just speculations because ATI and NVIDIA F@H developers doesn't want to confirm or deny this. However, it's very plausible since the programming datasheet of ATI make mention of the 1/5 shaders thing.

If you are interested to read more and discuss, go to http://foldingforum.org/viewtopic.php?f=43&t=3413

 

[mhouston]

All of our SPs can do FP MADs and most other FP ops. 1/5th of our units can do trig and other extended functions. You can feel free to take a look at our ISA since we actually publish the docs.

The difference is in the level of performance tuning vs. stability tuning. I think we are pretty good on stability and with each core update we increase performance. Nvidia has the fastest client with the current WUs, for now, but we are chipping away at their performance lead. Scott and Nvidia have done an amazing job, but don't forget that they are riding on 3 years of AMD code development as well and didn't start from scratch. We shall see what 4XXX tuning yields as there are lots of new features on the boards for compute and lots of headroom in the architecture as the current small proteins and code isn't scaling across all 800 SPs. When the code was originally worked on and tuning began, Folding@Home was targeting lambda and not villain, with lambda being 2X the size of villain.

It will be interesting in a few months to see where things stand for performance and stability. The good news for Folding@Home is that there are now 2 vendors battling it out, and unlike gaming benchmarks, performance here benefits science and help in the understanding of several really nasty diseases. My grandfather had Parkinson's disease, which is one of the diseases studied by Folding@Home, so I take Folding@Home seriously and personally.

 

[Xilikon]

Thanks Mike for the clarification (I was just merely passing what I read on FCF with theories based on those docs). I'm sure that with lots of tweaks, AMD can be very competitive and I'm looking forward to them

 

[kaleb_zero]

I'm also impressed with the number of FLOPS comming out of all of the GPUs now.. Looking at over half a petaflop and growing fast... It is like a HPC dream come true!

 

[Sunin]

What do you think the 4850 is capable of PPD wise? In the future that is with full utilization?

 

[jebo_4jc]

What do you think the 4850 is capable of PPD wise? In the future that is with full utilization?

It might be safe to assume roughly equal to the 9800GTX since those cards are comparable in gaming performance

 

[mhouston]

What do you think the 4850 is capable of PPD wise? In the future that is with full utilization?

Honestly, it's hard to say as the client and core on the Folding@Home side are changing, as are our drivers and compilers. The main issue for the smaller proteins is that with high-end GPUs, *any* serial portions of the code, be it host side calculations or driver overheads, can quickly limit the speedup. This has been our main concentration thus far before we have even really started that much tuning of the GPU kernel code because the CPU side of things was the limiting factor for us. Just with CPU side tuning, on a 3870 we've gone from 140ns/day to 180ns/day with the 1.03 core on a q6600. The 1.04 core has more tweaks, and we have a version getting 265ns/day now in the lab (lots of QA needed before that goes out). I think there is still headroom on the 3870, but much of that will now come from kernel tuning.

The 4800 has many more tuning options from the extended compute capability. The biggest first issue is just getting all the SPs fired up and working on the heavier kernels.

 

[jebo_4jc]

Honestly, it's hard to say as the client and core on the Folding@Home side are changing, as are our drivers and compilers. The main issue for the smaller proteins is that with high-end GPUs, *any* serial portions of the code, be it host side calculations or driver overheads, can quickly limit the speedup. This has been our main concentration thus far before we have even really started that much tuning of the GPU kernel code because the CPU side of things was the limiting factor for us. Just with CPU side tuning, on a 3870 we've gone from 140ns/day to 180ns/day with the 1.03 core on a q6600. The 1.04 core has more tweaks, and we have a version getting 265ns/day now in the lab (lots of QA needed before that goes out). I think there is still headroom on the 3870, but much of that will now come from kernel tuning.

The 4800 has many more tuning options from the extended compute capability. The biggest first issue is just getting all the SPs fired up and working on the heavier kernels.

we wait anxiously

 

[EvilAlchemist]

Well, little update for ya ...

I had to take the second card out of my machine..
There was either a problem with running two clients:

Here are the possibilities:
1) Two clients are conflicting ... (2nd client EUE over and over ... about 1100 of them actually)
2) PCI-E bus overload .. causes crash
3) Driver Problems ..

So .. I dropped it down to one GPU2 + 1 SMP

I am getting 1725 PPD on GPU2 now .. which makes me think the PCI-E bus was getting loaded...

I am gonna put this second card in another machine and go from there.

It does kinda suck that my SMP and GPU clients both get the about the Same PPD right now ...

But I do look forward to improvments

 

[SmokeRngs]

Well, little update for ya ...

I had to take the second card out of my machine..
There was either a problem with running two clients:

Here are the possibilities:
1) Two clients are conflicting ... (2nd client EUE over and over ... about 1100 of them actually)
2) PCI-E bus overload .. causes crash
3) Driver Problems ..

So .. I dropped it down to one GPU2 + 1 SMP

I am getting 1725 PPD on GPU2 now .. which makes me think the PCI-E bus was getting loaded...

I am gonna put this second card in another machine and go from there.

It does kinda suck that my SMP and GPU clients both get the about the Same PPD right now ...

But I do look forward to improvments

/me uses Jedi powers.

"You want to give me one of your 4850s."

/me finishes using Jedi powers.

 

[mavalpha]

Keep in mind that P35 only allows a 16/4 split. I can easily imagine, looking at Crossfire (gaming) results, that you are right about a lack of bandwidth. Conflicts between the two cards are unlikely, as the cards are operating independently of one another- P35 does not support peer-to-peer writes between the slots, so unless the clients are talking to each other (why would they?), they should be unaware of each other's existence.

(Down, but not out. My production will be back up as soon as I finish my project in the next week.)

 

[APOLLO]

Keep in mind that P35 only allows a 16/4 split. I can easily imagine, looking at Crossfire (gaming) results, that you are right about a lack of bandwidth. Conflicts between the two cards are unlikely, as the cards are operating independently of one another- P35 does not support peer-to-peer writes between the slots, so unless the clients are talking to each other (why would they?), they should be unaware of each other's existence.

What you're stating if true, then even the upcoming multi-GPU client won't help in that regard because it won't change anything as far as data moving through the PCIE-E bus is concerned.

 

[EvilAlchemist]

I am leaning toward a client conflict ... but the pci-e bus did not help matters ..

now that I only have 1 x 4850 all is running smooth..

I put my x1600 in the second pci-e slot to run my third monitor ...

I did see on Folding Forums .. we some users of multi-Gui were having problems ..

I am sure it will get fixed soon ....


I did notice on the new client about the core lock ....
can anyone explain?

 

[Xilikon]

Actually, I saw people running with a 2x pci-e slot without too much issues so it's more like client conflict (there is no easy way to set machine id for each client, as Adam Beberg or Mike Houston mentionned).