GPU3 Client ATI Question.

[kpe1950]

I am curious, this is not the only reason why i went with a 4890 but when I seen the news a while back about the gpu3 might increase performance on the ati cards, that was one of the reason's I went with ati.

In a guesstimation if any, what do the Big Power Folders here think that a 4890 will push in PPD on the up coming GPU3 Client?

On a Guesstimation, not on record. plz?

[toTOW]

It's too early to guess anything ...

[7im]

It has been said before that the theoretical performance of the two GPU makers is about equal. However, we have seen where theoretical and actual are quite different, and OpenCL will fall somewhere in between. WAAAY too early to start guessing at anything. Any further disccussion at this point is mere folly.

p.s. although folly can be fun as well...

[toTOW]

I've found another distributed project where the 4870 is more than 3 times faster than the 9800 GTX+ (project is pure ALU/integer operations) ... I guess it varies depending on the kind of computations run on the GPU.

[WickedPixie]

It has been said before that the theoretical performance of the two GPU makers is about equal. However, we have seen where theoretical and actual are quite different, and OpenCL will fall somewhere in between. WAAAY too early to start guessing at anything. Any further disccussion at this point is mere folly.

[size=85]p.s. although folly can be fun as well...[/size]

Speaking of folly...

Are GPU3 clients internally tested first, or are they in the wild already?
There is a claim that one is running GPU3 client already using Fermi ES cards and running an instance of F@H per shader cluster...

Because of the new MIMD architecture (they have 32 clusters of 16 shaders) i was not able to load them at 100% in any other way but to launch 1 F@H client per cluster and per card. Every client is GPU3 core Beta (Open MM library). I supose it is much more efficient then previous GPU2.

http://www.evga.com/forums/tm.asp?m=100960237&mpage=1

[toTOW]

No, GPU3 are definitely not in the wild ... so either this guy is lying, or it's a leak ...

I also think his story about cluster of 16 shaders is quite obscure ... I also don't see how only 16 shaders could produce more PPD than a single 9800 GTX+ ...

[7im]

Obviously, this guy cobbled something together, and is running something completely unsupported, is probably turning in bad data, and if not, broke several NDAs and EULAs with that post. Nothing good comes from any of that.

He said 32 clusters of 16 shaders, and running 1 client per cluster, per card, i.e. 32 clients per card, on 6 cards. Somehow, that just doesn't add up when you are limited to 16 Machine IDs per computer.

[deekey777]

Fermi is 16 SMs which contain 32 "CUDA cores": http://www.realworldtech.com/page.cfm?A ... 110932&p=4
MIMD? Yes, every SM can do his own stuff. But: http://forum.beyond3d.com/showpost.php? ... tcount=757

The guy needs some attention.

[WickedPixie]

Thanks for all the reply.

Isn't folly so much fun?

Obviously, this guy cobbled something together, and is running something completely unsupported, is probably turning in bad data, and if not, broke several NDAs and EULAs with that post. Nothing good comes from any of that.

He said 32 clusters of 16 shaders, and running 1 client per cluster, per card, i.e. 32 clients per card, on 6 cards. Somehow, that just doesn't add up when you are limited to 16 Machine IDs per computer.

Although it does worry me that where one claims to run a number of Fermi ES cards on a GPU client yet to released, whether it is true or not, the manufacturer fails to nip it in the bud due to possible NDA and EULA violations (whether it is from Stanford or Nvidia), let alone possibly turning in bad data if true. Wouldn't the Mfr of said card know where they have sent their ES cards to... again if true?

Makes one wonder why let a rumour perpetuate in their forums...


My apologies to the OP for the hijack.
Hope you were entertained as I was.

[chumbucket843]

No, GPU3 are definitely not in the wild ... so either this guy is lying, or it's a leak ...

I also think his story about cluster of 16 shaders is quite obscure ... I also don't see how only 16 shaders could produce more PPD than a single 9800 GTX+ ...

FAHMAN's ppd is very strange. he is getting ~150k a day and he has 204 active processors. G300 is 16 clusters of 32 shaders by the way. he says he is running a folding instance on each cluster. he probably has some kind of fpga or some other kind of massivley parallel processor running f@h. that would be the most logical explanation for his PPD.

[Slash_2CPU]

I call bull. 150k over 204 processors works out to only 735PPD. That's likely a single-cpu client farm at a school or business plus a few GPU's and SMP mixed in.

[7im]

I call bull. 150k over 204 processors works out to only 735PPD. That's likely a single-cpu client farm at a school or business plus a few GPU's and SMP mixed in.

Very good conclusion. A little bird told me 4 SMP, 31 GPU, and the rest are CPU clients (at least in the last 7 days).

...
Although it does worry me that where one claims to run a number of Fermi ES cards on a GPU client yet to released, whether it is true or not, the manufacturer fails to nip it in the bud due to possible NDA and EULA violations (whether it is from Stanford or Nvidia), let alone possibly turning in bad data if true. Wouldn't the Mfr of said card know where they have sent their ES cards to... again if true?

Makes one wonder why let a rumour perpetuate in their forums...

Yes, they know where ES cards go to, and those NDAs are what's probably keeping Fahman from posting again. And pulling the post would cause more uproar than it already has, even if EVGA or Stanford knew who to contact to have it pulled.

If true, this is obviously an attempt to show what might be possible at the VERY extreme end, if there wasn't a Machine ID limit of 16. But come on, CO2 cooling, long term?

EDIT: Added emphasis

[divery4eyes]

bad part about it is, there is about 5 pages in the forum of people who actually believe that this guy is getting 200k off of 7 cards
AND the 7th card isnt even maxed out
Along with an I7 without HT
2 x 1500watt power supplies
2400 watt total system power usage
Not to mention the fact that he constructed his own CO2 cooling system and replaces bottles every 5 days. . . . . .

I am not the one to say he is a LIAR but til I see pics, i believe nada

[Noegzit]

I call bull. 150k over 204 processors works out to only 735PPD. That's likely a single-cpu client farm at a school or business plus a few GPU's and SMP mixed in.

Hum not so sure... cause the stats say (today) 210 active processors in last week.
I personally have no clue to decide if there are 210 processors or 10 processors run by different clients with an average of 3 ID change a day.

For example I'm currently using 6 processors, but as I recently replaced my CPU clients by SMP clients and the ID of my GPU clients the stats say 12 active processors in 7 days.

[7im]

I call bull. 150k over 204 processors works out to only 735PPD. That's likely a single-cpu client farm at a school or business plus a few GPU's and SMP mixed in.

Hum not so sure... cause the stats say (today) 210 active processors in last week.
I personally have no clue to decide if there are 210 processors or 10 processors run by different clients with an average of 3 ID change a day.

For example I'm currently using 6 processors, but as I recently replaced my CPU clients by SMP clients and the ID of my GPU clients the stats say 12 active processors in 7 days.

Read my post above. We already have more than a clue.

4 Linux SMP clients, 31 GPU clients, and the rest are CPU clients, mostly on XP (at least in the last 7 days).

Reguardless of what was/is claimed, I have to give the guy a pat on the back for contributing to the project with over 200 clients!