Hi Bigtuna,

I've checked your PC & compared to the one Beyond uses. He has 2GB VS your 1GB RAM, but you've had lots of failed WU's, too few to make estimates (My personal opinion), Beyond has had lots of successful WU's.

But that's beside the point. What I wanted to know, was if the GTX470 which gpugrid has had more time to play with, also had a nice boost running XP 64bit compared to running XP 32bit.

The only way I'd guess how to make a comparison, was if someone who ran XP 32bit with a GTX470 installed XP 64bit w/o modifying his PC & shared that knowledge.

---

Hi Bigtuna,

I've checked your PC & compared to the one Beyond uses. He has 2GB VS your 1GB RAM, but you've had lots of failed WU's, too few to make estimates (My personal opinion), Beyond has had lots of successful WU's.

But that's beside the point. What I wanted to know, was if the GTX470 which gpugrid has had more time to play with, also had a nice boost running XP 64bit compared to running XP 32bit.

The only way I'd guess how to make a comparison, was if someone who ran XP 32bit with a GTX470 installed XP 64bit w/o modifying his PC & shared that knowledge.

Failed work units?

So far the only failed work units have been the incompatible CUDA 3.0 work units that got sent during the beta testing. Full sized CUDA 3.1 tasks have been rock solid AFAIK. Of course they have only been running a day or two so that isn't much of a test.

Are you also using the swan_sync environmental variable, and do you have a core free?

That computer is running swan_sync=0 currently. Turned it "on" after the first couple full sized tasks. You can tell by the CPU time. With swan_sync "off" CPU time is minimal and with swan_sync=0 CPU time is about the same as the the GPU time, and yes there is a free core available on that box.

Did we ever decide exactly what swan_sync was?

The card came with a factory OC to 763/1526.

So far I'm impressed and disappointed at the same time. The 460 runs cool and quiet which is good but the performance is only about 2.5 x GT-240 but the price is 4 x GT-240.

The good news is I think I have figured out my gtx460 problems; bad news is I have one defective card that is prone to crashing. Running them both 1 at a time trouble shoot the issue.

Did we ever decide exactly what swan_sync was?

It is an environmental (system) variable used to synchronise the GPUGrid application to the CPU. I originally thought it linked the app specifically to a core but I now think it forces the operating system to continuously poll the app, allowing the app to have immediate CPU resources when required (removing a bottleneck). I expect the number zero causes continuous polling, whereas if a value of 5 was used for example it would only poll every 5 sec.

So far I'm impressed and disappointed at the same time. The 460 runs cool and quiet which is good but the performance is only about 2.5 x GT-240 but the price is 4 x GT-240.

This is the first app that actually works for the new GTX460 cards. There has been no time to look at the GF104 architecture in greater detail, and then refine the apps and tasks to suit the card. It may still have significant unrealised potential. The latest application v6.11 also improved performance for the first Fermi's (GF100), so hopefully an improved app will do the same for the GF104 cards in due course. The 6.11 app & latest driver combo also improves the Vista/W7 performance dramatically. Vista and Win 7 still perform slower than XP and Linux, but now do so to a similar extent, across all cards (they are equally slower, roughly speaking). Prior to the latest app the performance of Fermi's on Win7 and Vista was almost half what it is on XP. It's now only about 10 to 15% slower, on properly configured systems. As for price, that will drop over time in the same way the GF100 Fermi's dropped in price: My GTX470 cost £320, now it can be picked up for £277, and there are cheaper cards. You are also comparing the new GTX460 (2 weeks old) to a GT240 card that has dropped by some 40% since its release.

People should not be too surprised about the present performance given the shader layout; it's totally different than for the GF100 cards. Basically its performing as if it has 2/3rds the number of shaders, due to the way these are accessed. I doubt that the present app can tap into these last 33% of shaders efficiently, if at all. So perhaps there is good room for improvement.

trn-xs, take the faulty card back and get a refund or a replacement.

Did we ever decide exactly what swan_sync was?

It controls the method that ACEMD tells the CUDA runtime to use for polling for GPU work completion. The default method is for the application to block until completion; this keeps CPU load to a minimum but introduces latency that slows the program down.

Setting SWAN_SYNC=0 will cause ACEMD to poll for kernel completion, which minimises latency at the cost of CPU.

0 is the only valid value for SWAN_SYNC - anything else will cause undefined behaviour, so don't do it!

The 460 runs cool and quiet which is good but the performance is only about 2.5 x GT-240 but the price is 4 x GT-240.

We will be turning our attention to improving the performance on GF104 cards after the summer vacations. We know what needs to be done.

MJH

Well mine arrived today. I have swapped out a GTX275 for it and its off and running.

Interestingly GPU-Z 0.4.4 says it only has 224 shaders. Maybe they know something we don't :-)

---

BOINC blog

Well initial performance is actually worse than the GTX275. As MJH said they need to work on the app. I'd say GPU-Z needs an update too.

Thats the price for rushing out and buying the latest & greatest toy. We just need to wait a while for the s/w to catch up.

---

BOINC blog

Its a bit early to clearly tell here, but it looks like my 460 is out performing my 280 so far. Its not a comparison because my 280 is running old drivers and the 460 is OC'd to 750/1500.

The best comparison I can draw so far is:
gtx460 completes a 6,800 credit WU in 17,400 seconds
gtx280 completes a 6,800 credit WU in 18,800 seconds

Stock vs stock with both on Cuda 3.1 drivers might be a pretty even match. Both computers crunch wcg on all cores and i do not use swan_sync.

Kill-a-Watt power info for the GTX-460. I've got the 768 MB version with a factory OC to 763/1526 MHz.

At idle the AMD 4000+ system draws 93 watts.

Running GPUGRID with Swan_Sync=0 the system draws between 195 and 210 watts.

I think it would be generically more accurate to compare the GTX460 768MB with the GTX260-192, and compare the GTX460 1GB with the GTX260-216:
The 192 version of the GTX260 was about 10% slower than the 216 version (at reference, for most applications and games that it worked on; not here).
The 768MB version of the GTX460 is about 10% slower than the 1GB version.

If the GTX460 cards are only using 224 of their 336 shaders then they square up fairly evenly in terms of shader performance to the GTX260, with the 1GB version perhaps being about 5% better than the GTX260-216 at reference (on XP or Linux). Obviously if you get a good overclocked 768MB version it could outperform a GTX280, but equally a good GTX260-216 with a high clock would outperform a native 1GB version of the GTX460.

The present GTX460 benefits for crunching here are that they are quieter, run cooler, use less electric, and overclock more.
The problems are that they are still slightly too expensive, there are not enough 1GB card available, the drivers just really got the card working and need to mature, as does the app (after the holidays).

Come the autumn (and the GTX475 with 8cores and 384shaders) there will likely be new drivers and a new app, which should improve performances. The CC1.3 cards are almost 2 years old, so the drivers and apps are well refined at this stage; the present 6.05 app is well over twice as fast as the original.

It is likely that the power usage will rise slightly once these cards are better utilized, and the performance will increase somewhat more, so it’s a bit too early to do an accurate crunching performance per watt comparison here. The good news is that the architecture is set to remain for the GTX475, so when that turns up, the scientists will be better prepared for it.

Well initial performance is actually worse than the GTX275. As MJH said they need to work on the app. I'd say GPU-Z needs an update too.

Worse performance also than the GTX 260. I bought the GTX 460 specifically for GPUGRID but can't justify using it here until the performance improves. The good news is that it runs extremely well in Collatz and DNETC.

@bigtuna: What was the percentage of GPU usage for your Kill-A-Watt readings?

@sk: Wow, just when it seemed "silliest and most irrelevant post of the month" was wrapped up. It would be nice to stick to reality instead of speculation based on fantasy and misinformation. BTW, the 768MB and 1GB models of the GTX 460 perform the same in Folding, Collatz and DNETC. Collatz is an app that is very much affected by memory speed. I think we know the GTX 460 isn't being utilized well by GPUGRID. Hopefully that will improve.
_______________________

Tried running the GTX 460 in XP64 with swan_sync set and 1 core reserved for the GPU (GX780 quad core system). The result was no significant speed increase in GPUGRID (1 minute or so) and loss of 1 CPU core in another project. In addition the system became very slow to respond. Without swan_sync the system response was fine.

The shader speed is more important here than the RAM speed. Besides, the actual speed of the GDDR5 is not the issue with the 768MB card versions, the issue is the memory interface!

"The more expensive GTX 460 1GB version is based on the full-fat GF104 GPU and so has the maximum 256-bit memory interface. This is because the 1GB version of the GTX 460 has more memory links, and therefore more memory bandwidth. While the GTX 460 768MB has only six links to memory, so has only six 128MB memory chips on the card (6 x 128MB = 768MB), the GTX 460 1GB GPU has the full eight connections (8 x 128MB = 1GB). As each memory connection is 32 bits wide, the GTX 460 768MB has a 192-bit memory interface (6 x 32 bits = 192 bits) while the GTX 460 1GB has a 256-bit bus". Bit-tech, by James Baker.
Many games also observe this 10% reduction in performance between the 1GB and 768MB versions.

I see the GTX460 as performing about the same as a GTX260, with the 768MB version falling slightly behind the GTX260 and the 1GB version slightly outperforming a reference GTX260. There are plenty of results to go by.

@bigtuna: What was the percentage of GPU usage for your Kill-A-Watt readings

GPUZ reports between 75 and 90 percent GPU load (it varies and so does the power).

Many games also observe this 10% reduction in performance between the 1GB and 768MB versions.

Exaggeration and far too many exclamation points doesn't provide credibility. Even your own cherry-picked article doesn't show anywhere near 10% improvement in games. As you can see by the Anandtech test that you posted here, there is no difference between the 2 models at all in Folding:

http://www.gpugrid.net/forum_thread.php?id=2227&nowrap=true#17956

As I'm sure you know the 1GB card also uses a higher stock voltage and thus uses more power, runs at higher temps and is generally less overclockable compared to the 768MB model even at it's lower stock voltage (again, shown in the Anandtech article where you grabbed the Folding chart above).

The truth is results vary depending on the application. Most DC apps are not memory constricted. A few are. Exclaiming something else doesn't make it true.

Did we ever decide exactly what swan_sync was?

It controls the method that ACEMD tells the CUDA runtime to use for polling for GPU work completion. The default method is for the application to block until completion; this keeps CPU load to a minimum but introduces latency that slows the program down.

Setting SWAN_SYNC=0 will cause ACEMD to poll for kernel completion, which minimises latency at the cost of CPU.

0 is the only valid value for SWAN_SYNC - anything else will cause undefined behaviour, so don't do it!

MJH

Awesome, thanks for the info.

SWAN_SYNC is working great for me:

The first work unit I ran was with SWAN_SYNC "off". It took 17,150 seconds for a 6,803 point task.

With SWAN_SYNC "on" (SWAN_SYNC=0) similar tasks are taking about 15,850 seconds which is about an 8% difference.

The difference is significant enough that I'm leaving SWAN_SYNC "on".

Different operating systems and different hardware could react differently.

Running those 6803 point tasks (larger tasks have been a bit slower) the GTX-460 is good for about 37,083 points per day with SWAN_SYNC=0.

With SWAN_SYNC "off" the unit would pull down 34,272 points per day so the one sacrificed CPU core is good for about 2.8k points per day.

2.8k/day is more points that that one core would pull down running Rosetta.

NOTE: Edited to have SWAN_SYNC in caps which seems to be required for at least some Linux distros.

I don't think the 1Gb cards are going to provide any better bang for the buck.

According to GPU-Z the memory controller load when running GPUgrid 6.11 app and Swan_Sync=0 is around 15 to 17% on my 768Mb card. Its hardly under pressure.

---

BOINC blog

2.8k/day is more points that that one core would pull down running Rosetta.

But it's less than 1 core can get in Aqua@home hence why i keep my gtx 260 without swan_sync.

2.8k/day is more points that that one core would pull down running Rosetta.

But it's less than 1 core can get in Aqua@home hence why i keep my gtx 260 without swan_sync.

Same here, and CPU credits aren't exactly comparable to GPU credits. Some projects can only be run on the CPU and a core on those projects means a lot.

On the other hand lets compare credits between the GTX 260 and GTX 460. My GTX 260 does around 40,000 credits/day in GPUGRID and about the same or even a bit less in Collatz. On the other hand my GTX 460 does around 35,000 credits/day in GPUGRID and about 103,000 credits/day in Collatz without sacrificing a CPU core. Quite a difference...

IMO the only way as not to discriminate or favoritise projects, is if BOINC Projects can agree to give credits based on CPU/GPU performance per watt & the CPU/GPU time used on WU's x deadline bonus. Is this even possible though???

If effective CPU/GPU's were encouraged & ineffective CPU/GPU's were discouraged. BOINC could send the message that wasting watts is a waste of time.

---

2.8k/day is more points that that one core would pull down running Rosetta.

But it's less than 1 core can get in Aqua@home hence why i keep my gtx 260 without swan_sync.

Same here, and CPU credits aren't exactly comparable to GPU credits. Some projects can only be run on the CPU and a core on those projects means a lot.

On the other hand lets compare credits between the GTX 260 and GTX 460. My GTX 260 does around 40,000 credits/day in GPUGRID and about the same or even a bit less in Collatz. On the other hand my GTX 460 does around 35,000 credits/day in GPUGRID and about 103,000 credits/day in Collatz without sacrificing a CPU core. Quite a difference...

We are working on optimizing the application even further. In September some news.

gdf