• Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
    Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.

Core x17 - WU 8900

A

AndyE

Limp Gawd
Joined
May 13, 2013
Messages
276
Pande Group has started to distribute a new core 17 work unit - 8900
http://foldingforum.org/viewtopic.php?f=66&t=24337&sid=aa7c54bfcb8c86eafb6182b7faa96cd1

All my 4 Titans picked up this new WU today. I don't have permission to post in the beta forum of foldingforum.org, so I will share my experiences here.

The 8900 WU seems to be better computationally optimized vs. 7663.
Power consumption is approximately 25 watt higher per GPU (50 watt for the system)
Temperature is also 5-8 degrees higher than with 7663 work units.
I had to reduce clock speed down to 1162 MHz to avoid speed bumps (throttling). One GPU which had initially a higher frequency throttled down to 836 MHz - had to reboot the system to get the GPU back to full speed. This work unit seems to put more stress on the GPU than 7663. Overclockers beware ;)

TPF is on my systems is 2min35sec

The bonus calculator has not been updated yet for 8900 to check ppd

The web control shows 400k ppd with 2 Titans (the first 400k PPD forecast on my dual GPU systems)
original.jpg


More details to the WU
original.jpg


rgds,
Andy
 
AndyE,

psummaryC has been updated with all the information. Looks like your Titans are pulling ahead of the 780 now... I'm getting TPF of 2:55 with PPD @ 169k.

Curiously, my 690s/680 haven't gotten a single one of these WUs, might only before for GK110?
 
Thanks for the psummaryC pointer.

btw, TPF 2:55 vs. 2:35 is the shader ratio of the 780 and Titan (12 SMX to 14 SMX) ....

All my 4 Titans got 8900s, the two 780s are working on 7663s.
 
Lucky you! Looks like a nice little healthy boost for us 600 series and beyond folk. I'm curious about the GPU being demanded even further. Since the 7663 already hammers them pretty good (2 x 660's myself) and require a single core per GPU, is the load a little less on each core being held by each GPU? I know they can't fix the issue with bonding a single core to a GPU, but I'm hoping they can tweak the usage on them. The 25% load per GPU is making it very difficult these days lol. Perhaps that is slightly offset with the 8900 or no?
 
Nope, still using a whole core per GPU. The GPU however is pushed further, temps and utilization are up from the 7663.
 
Liger88 said:
Lucky you! Looks like a nice little healthy boost for us 600 series and beyond folk. I'm curious about the GPU being demanded even further. Since the 7663 already hammers them pretty good (2 x 660's myself) and require a single core per GPU, is the load a little less on each core being held by each GPU? I know they can't fix the issue with bonding a single core to a GPU, but I'm hoping they can tweak the usage on them. The 25% load per GPU is making it very difficult these days lol. Perhaps that is slightly offset with the 8900 or no?
Click to expand...
These are 2 different things:

1) The CPU load:
This is a driver issue with NVidia drivers which keeps the CPU at 100% (but basically doing nothing, just polling the GPU, if it is ready to accept the next command). AMD has in this regards better drivers, as the CPU is basically kept in idle mode. I am not using the CPU for folding (I used deliberately the cheapest low power CPU - a Celeron G1610). If I would fold on the CPU, the way teh NVidia driver consumes one CPU core, would reduce effectively my available CPU cores for folding by one/GPU in the system.

2) The GPU load:
There are many factors impacting the so called "computational density". Similarily to a CPU, where you could use a simple loop with one addition or multiplication, your CPU would report "100% busy". But only a few transistors are actually doing something. Well tuned applications like Intel's LinPack keep most of the transistors of the chip simultaneously busy, which is the key reason why those apps are more demanding to cool and need more energy.

The same is true for GPUs. Pande Group switched over to a new programming model (OpenCL) to drive the gPUs and I assume that consecutive refinements of these cores will gradually increase the "computational intensity" on the GPU. This is what I see here happening. WU 7663 were the first core17 apps, and now WU 8900 with a modified core17 is a bit more optimized.

All of this is usually independent of the capabilities of the CPU (exceptions apply)

rgds,
Andy
 
Ahhh thanks for clearing that up Andy, hopefully Nvidia gets their act together especially with the Open CL 1.2 support so they don't fall too far behind AMD.

As long as Core_17 will keep eating up a full core I suppose I'll take all the optimizations that they can bring to the GPU side of things to increase the PPD. Those are some really beautiful numbers.
 
You can force core_17 to use less than a full core (for Nvidia), but how much less (before you degrade performance) will depend on your cpu and gpu. Core_17 actually does a lot more than previous cores. There is still debug code, and it writes a checkpoint every even %. How much work core_17 demands of your cpu will be a function of how fast the GPU is (since accuracy checks and checkpoints are indexed to progress), so a fast GPU will require more resources.

If you have some combination of a relatively slow GPU, a relatively slow WU, and a relatively fast CPU, you can probably dedicate less resources to it. With QRB you want to be careful though. For mid-range and up GPUs, ppd potential is higher than for smp now, so if you're going to err on one side or the other, better to err on the side of giving the GPU too many resources...
 
Core 0x17
WU 8900 (26,0,8)
TPF 15m4s
est. PPD ~14.5k

GTX 460 DirectCU
GPU Core 675MHz
Mem 900 MHz
Shader 1350MHz
GPU Load 96~100%
Memory controller load 16~18%
forceware 311.06
CPU Q6600 @ 3GHz
 
2 x 660 SC's are pushing a total of 83K PPD, on the 7663's they were doing a solid 78K PPD. Not much of an improvement and far bigger/longer WU's, which I tend not to like in case errors come up. Definitely seems like they are meant for the beefier cards which can handle them faster.

Right now TPF is 7m22s and ~ 12 hours = 21K

I'd rather do just 7663's though and the 8900's definitely do seem to use more of the GPU than others have on the 600 series for very little in return. I thought 3 hour WU's were enough, then those 4 hour, 8 hour and now 12 hour ones having moved up the GPU latter from a GTX 260 > 660 are just harder to swallow. Especially if you're like me and had some serious WU errors in the past lol. Those things just compound after awhile if you're not careful.
 
Liger, without being apart of the Beta Team I can't know for certain but 7663 should be released to -advanced soon enough and which point just have -advanced instead of -beta and you'll dodge the 8900s
 
Thanassos said:
Liger, without being apart of the Beta Team I can't know for certain but 7663 should be released to -advanced soon enough and which point just have -advanced instead of -beta and you'll dodge the 8900s
Click to expand...

That being said, the 8900s spent more time in closed beta once stable vs the 7663 project so that may not follow, but I don't know for certain. I expect *something* will need to move soon to make space for the next closed beta. [Disclaimer, I am on the open and closed beta team]
 
as long as its putting out similar results to 7663, i'll take the increase in power. The real killer are the Core 16 units I still get every day or two.. they absolutely decimate my 7970's PPD (107k -> 6k)
 
HD7770 - tpf 11:25, est ppd 21,545. I'm folding smp on this CPU as well and impact seems minimal (27-34K ppd)
 
I picked up a 7970 the other day just to test out Core 17. The first unit it picked up was one of these 8900s. The card is a Gigabyte OC version, not a GHz Edition, but still clocked at 1GHz out of the box.

I'm getting a TPF of around 4:04 except I'm noticing something really wierd happening. After every second frame, the card clocks down to idle for about 5 seconds or so and then afterwards goes back to full utilization which then increases the TPF for the next frame to 4:20 or something. The frame after that, it returns to 4:04.

If the TPF would stay at 4:04 I would have over 100K PPD, but with that little bit of pause after every second frame, it drops my PPD down to 97K.

Client is version 7.3.6, folding power is set to full. Drivers are Catalyst 13.4.
CPU is a 2700K (@ stock) even though it sounds as though AMD doesn't require any extra CPU power to feed the GPU... so that shouldn't be a factor either way.

Anyone have any ideas how to get rid of that pause after every second frame?
 
The client does something like a backup or error check etc. every second frame from what I remember.
 
I see the pause too, but I don't seem to notice my client changing the TPF after it happens. I figured it was just the checkpointing system and didn't really think twice because it didn't have any kind of effect on the reported PPD
 
At first I thought it was the work unit itself that had some sort of issue with it. Then after posting here I checked my other system with my pair of GTX 670's and they're both apparently doing the same thing with the 8900's they're working on... though I didn't notice if they were doing this with 7663's. Hmmm.

After leaving the client folding for awhile, the average reported TPF does settle down to about 4:11 or 4:12, but from looking at the logs, the actual TPF still jumps between 4:04 and 4:22 or something.

At any rate, I'm gonna mess with the clocks on my 7970 to see if I can get that one up beyond 100K even with the pause factored in.
 
Thanassos said:
Liger, without being apart of the Beta Team I can't know for certain but 7663 should be released to -advanced soon enough and which point just have -advanced instead of -beta and you'll dodge the 8900s
Click to expand...


Yeah I'm hoping something comes up soon. My fear already hit with these 8900 WU's. With 60% done on the 12.5 hour beast WU, it suddenly became a bad WU, crashed, and restarted a new one from 0.00%. Have yet to see what the PPD damage is, but it's annoying doing these extremely long ones taking that risk and taking a huge PPD hit if something goes wrong. Wish they could limit those to the most powerful video cards that can do them in a reasonable time.

Just barely got used to doing these 4 hour ones and had some major problems the previous week that destroyed my 80% > rating taking 4 days to build up. Apparently other places are having 50/50 reports about 8900 giving them either less points than normal 7663's or being bad WU's and failing constantly.
 
In case you use overclocked gpu, take them back to normal. I had to do this with all my gpus. 8900 is a bit more demanding
 
AndyE said:
In case you use overclocked gpu, take them back to normal. I had to do this with all my gpus. 8900 is a bit more demanding
Click to expand...


Yup, already underclocked the GPU core by 50Mhz and memory by 100Mhz to get it stable from last weeks OC fail fiasco to get it running Core_15/17 (7663 units) stable and all was well. Have done 6 8900 WU's with no problems, this was the only one I ran into a problem with today. We'll see how that plays out. Trying to finish the 8900 WU on one gpu and hopefully when I restart it it'll grab a 7663 lol :)

I'll leave the 8900's to the big boys (680's, 780's, Titans, and 7970/50's).
 
Liger88 said:
Yeah I'm hoping something comes up soon. My fear already hit with these 8900 WU's. With 60% done on the 12.5 hour beast WU, it suddenly became a bad WU, crashed, and restarted a new one from 0.00%. Have yet to see what the PPD damage is, but it's annoying doing these extremely long ones taking that risk and taking a huge PPD hit if something goes wrong. Wish they could limit those to the most powerful video cards that can do them in a reasonable time.

Just barely got used to doing these 4 hour ones and had some major problems the previous week that destroyed my 80% > rating taking 4 days to build up. Apparently other places are having 50/50 reports about 8900 giving them either less points than normal 7663's or being bad WU's and failing constantly.
Click to expand...

It is in beta, so problems are to be expected. That being said, a vast majority of failures are OC related. 8900 is very hard on GPUs, and any OC (especially on Nvidia) is potentially too much (even factory overclocks). If you're running Nvidia, I'd recommend dropping to reference clocks, at least as long as these are in beta.

P.S. 7663 and 8900 are on different servers, with 8900 set to a much higher priority. So, as long as both are in beta, you may not see any more 7663s. Proteneer indicated 7663 has scientific value, but a significant part of its importance was as a proof of concept. So, it may be that 7663 won't be advanced, time will tell.

EDIT: 8900 is a big deal at Stanford. There are apparently several researchers preparing similar projects for core_17. While this doesn't mean there won't be small core_17 WUs in the future, I also wouldn't be surprised if 8900-like projects become the norm for core_17, rather than the exception.
 
I had <max-packet-size v='big'/> set.

I removed this, and now I am getting 4 8900 units for 340000 PPD.
 
AndyE said:
In case you use overclocked gpu, take them back to normal. I had to do this with all my gpus. 8900 is a bit more demanding
Click to expand...

Maybe but ymmv. I tried a 50 MHz bump for my HD7770 and its folding 8900's just fine so far. Granted its a measly 50 MHz boost up to 1050MHz on a XFX Ghost Edition HD7770, but with tpf's of 11+ min I might try for a little more. I used Afterburner with no voltage added (yet) . Its currently getting more ppd than my GTX460/550Ti /650Ti's with less power draw so it was a decent budget buy (<$100)
 
Toconator said:
Maybe but ymmv. I tried a 50 MHz bump for my HD7770 and its folding 8900's just fine so far. Granted its a measly 50 MHz boost up to 1050MHz on a XFX Ghost Edition HD7770, but with tpf's of 11+ min I might try for a little more. I used Afterburner with no voltage added (yet) . Its currently getting more ppd than my GTX460/550Ti /650Ti's with less power draw so it was a decent budget buy (<$100)
Click to expand...

AMD seems more tolerant of OCs on 8900, but it's still finicky.
 
I agree that overclocking might be a YMMV thing.

My pair of GTX 670's have pretty significant overclocks on them on top of their factory overclocks... so far, no failures and they've been doing 8900's for a couple days already.

I'm still keeping an eye on my 7970. I just finished bumping the core up another 100MHz from its already factory overclocked speed and it also seems to be fine so far.
 
Seems like, for some reason, it takes an extremely long time for this WU to restart after pausing. On multiple times it just appears to hang before the core actually engages the GPU, but it looks like it can take 3-5 minutes just to restart from pause
 
d4f867f2-7c92-4310-b379-c1247eb88f47_zps4f8cf078.jpg


Quad Titan with a mild 150Mhz overclock on the core, memory at stock.

This is the rebuild of my quad GTX470 rig.

Intel 970X @ 4.2Ghz
12GB Ram
Watercooled.
 
You must log in or register to reply here.
Back
Top