A Fermi benchmark of sorts...

[Talonman]

Fermi can run 16 compute kernels in parallel, one per core.

Thanks,

That is still a bit confusing for me.
(I need a more direct answer.)

Fermi will also have 1.5GB of memory, and is made up of 3.0 billion transistors and features 512 CUDA processing cores organized into 16 SMs (Streaming Multiprocessors) of 32 cores each.

So in theory, 16 X 32 Cores each = 512 GPU Folding instances on (1) Fermi?

-or-

16 X 512 total Cores = 8,192 GPU Folding instances on (1) Fermi? (Sounds a tad high to me.)

-or-

16 GPU Folding instances total on (1) Fermi, with each instance being processed on 32 Cores.

[Chumbucket843]

ok i will explain a gpu from the bottom up. they are made up of alu's and fpu's which stand for arithmetic logic unit and floating point unit. a streaming processor now known as a cuda core has an alu and a fpu. there are 240 in gt200 and 512 in fermi.the units are not independent. they cant do anything by just themselves. these operate on multiple parts of a calculation at once. take for example ray tracing. each shader aka streaming processor could work on individual pixels at the same time. 8 stream processors are together to make a SIMD unit also called a streaming multiprocessor. there are 30 SM's and 8 stream processors in each of them for a total of 240. the SM's are then grouped in 3's to form a tpc or thread processing cluster. fermi does not have tpc's.

here is an overview of gt200. they have one on fermi now if you would like to read that too.
http://www.dspvlsi.uniroma2.it/corsi...IDIA_GT200.pdf

[Talonman]

Good info...

Same question: So, how many GPU folding instances could (1) Fermi load up in theory?

[NeedMoMegaHurtZ]

fahman's original post says 1 GPU3 client per cluster, per card. So that would be 32 (since it has 32 clusters, each having 16 shaders). His post suggests that there are other configurations, but he said that is the only way he was able to get 100% load

He said he couldn't get the 7th card fully loaded, that's why he only has 200 (it should be 224)

But who knows if any of this is true?

[Talonman]

Thank you!!!

It seems like awful specific info, for a guy that is just making the entire story up, doesn't it?

32 instances x 7 Fermi = 224 GPU Folding instances, and probably about 1 or 2 CPU folding instances, for a total of about 226.

He is folding on over 200 processors. there is no way that is fermi. its just gpu3 beta. 31 gpu's and an i7. g80 and up are MIMD arrays of SIMDs. it sounds like it is a current gpu because he is getting 700 points per SM. that sounds like g92. i find it ironic that if there are 248 active cpus that means exactly 31 gpu's.

226 and 248 are pretty darn close.
22 might have just went to never-never land, while he was getting set up.

Looking here: http://fah-web.stanford.edu/cgi-bin/...sername=FahMan
Active clients (within 7 days) = 228

That would be 7 Fermi * 32 instances + 4 CPU clients running... (Looks correct to me.)


More thoughts: Each Fermi needed to have 32 folding instances loaded up on each GPU, to ensure it was working at 100% of it's capacity.

Each of the 32 folding instances get 16 Fermi shaders to calculate on, and must produce about 1,800 PPD. (1800 X 32 = Fermi's 57,600 PPD)

On g92, (1) GTX 275 produces 9K a day...
9K PPD / 240 Shaders = 37.5 Points per Shader, per Day.
37.5 PPS X 16 g92 Shaders = is only 600 PPD.

If 1 Fermi shader is 3 X as fast as a g92 shader when folding...
16 Fermi shaders would produce 1,800 PPD.
32 Folding instances, running 1 per cluster on (1)Fermi = 57,600 PPD

This all comes down to if you think Fermi's new shaders, with their memory cashes and better calculation speed, can be 3 X as fast as a g92 shader at folding.

If you believe they can be, then 32 folding instances running on (1) Fermi, with each instance being processed on 16 (New improved) shaders = 57,600 PPD.

I see no reason why Fermi couldn't have a 3 X jump in shader speed over g92 when folding.

FahMan was telling the truth, and has the PPD to back it up!

[NeedMoMegaHurtZ]

Thank you!!!

It seems like awful specific info, for a guy that is just making the entire story up, doesn't it?

32 instances x 7 Fermi = 224 GPU Folding instances, and probably about 1 or 2 CPU folding instances, for a total of about 226.



226 and 248 are pretty darn close.
22 might have just went to never-never land, while he was getting set up.

Looking here: http://fah-web.stanford.edu/cgi-bin/...sername=FahMan
Active clients (within 7 days) = 228

That would be 7 Fermi * 32 instances + 4 CPU clients running... (Looks correct to me.)


More thoughts: Each Fermi needed to have 32 folding instances loaded up on each GPU, to ensure it was working at 100% of it's capacity.

Each of the 32 folding instances get 16 Fermi shaders to calculate on, and must produce about 1,800 PPD. (1800 X 32 = Fermi's 57,600 PPD)

On g92, (1) GTX 275 produces 9K a day...
9K PPD / 240 Shaders = 37.5 Points per Shader, per Day.
37.5 PPS X 16 g92 Shaders = is only 600 PPD.

If 1 Fermi shader is 3 X as fast as a g92 shader when folding...
16 Fermi shaders would produce 1,800 PPD.
32 Folding instances, running 1 per cluster on (1)Fermi = 57,600 PPD

This all comes down to if you think Fermi's new shaders, with their memory cashes and better calculation speed, can be 3 X as fast as a g92 shader at folding.

If you believe they can be, then 32 folding instances running on (1) Fermi, with each instance being processed on 16 (New improved) shaders = 57,600 PPD.

I see no reason why Fermi couldn't have a 3 X jump in shader speed over g92 when folding.

FahMan was telling the truth, and has the PPD to back it up!

It's possible he's telling the truth, but again, you can't compare GPU2 to GPU3 results. Apples and oranges. e.g. What if a G92 produced 20K PPD on GPU3? Who knows? Even fahman says in his original post GPU3 is more efficient... It's like comparing 3dmark03 to 3dmark vantage and trying to measure "points" between them. You need consistency and the same methodology for a "benchmark."

[W1zzard]

It seems like awful specific info, for a guy that is just making the entire story up, doesn't it?

yes it adds up nicely, because you can easily fabricate something like it. going backwards from x clients (which is the only verifiable number), you use nice numbers to calculate the number of cards you need, the remainder you write off as cpu folding cores, problem solved.

example: it says i have 391 (totally random number i just picked) clients, i know fermi has 512 shaders, so why not say .. i used .. 12 cards in 3 systems, i ran 32 instances on each gpu, thats 32*12 = 384 clients, 7 are missing, i now claim i ran three cpu clients on two systems (i have quad cores in these), and one cpu client on the 3rd system (it has only a dual core) that's 7 .. woot 391 clients ... un-ing-believable how all this fermi info adds up

I see no reason why Fermi couldn't have a 3 X jump in shader speed over g92 when folding.

so you think it is probable that nvidia made a gpu that has 3x the performance _per_ shader, managed to put over twice the amount of these wonder shaders into a 40 nm gpu die and got 7 actual working gpus back from tsmc to give to some random dude for folding? also try to work with the 2.4kw he posted for the ppd delivered, work out performance per watt for these cards, also work out power draw for one of these cards

[K404]

Logistics make me side with W1zzard.... if this guy has a Fermi farm then he is

A/ very very important to nVidia yet somehow we've never knowingly heard of him before

B/ nV are taking GPU crunching very very very seriously if F@H is the first Fermi info leaked

C/ soooooo many other rumours and logical conclusions are false. If nV are in a position to give 5 Fermi to 1 guy, the validation is done, the silicon is probably final, the drivers are in the zone and they have enough stock to be able to get all these things done.

Flies right in the face of "that mock-up card" of a few weeks ago. iF nV could put a real card in their leaders hands for that demo, they would have.




As long as his F@H PPD are real and useful to Stanford, then Im not so fussed

[Talonman]

It's possible he's telling the truth, but again, you can't compare GPU2 to GPU3 results. Apples and oranges. e.g. What if a G92 produced 20K PPD on GPU3? Who knows? Even fahman says in his original post GPU3 is more efficient... It's like comparing 3dmark03 to 3dmark vantage and trying to measure "points" between them. You need consistency and the same methodology for a "benchmark."

I can't speak to that. My real interest was Fermi, not the BETA client. My gut tells me the BETA GPU folding client wont make things run that much faster per instance. The real goal is probably to be able to load up more instances to keep Fermi working near capacity.

yes it adds up nicely, because you can easily fabricate something like it. going backwards from x clients (which is the only verifiable number), you use nice numbers to calculate the number of cards you need, the remainder you write off as cpu folding cores, problem solved.

example: it says i have 391 (totally random number i just picked) clients, i know fermi has 512 shaders, so why not say .. i used .. 12 cards in 3 systems, i ran 32 instances on each gpu, thats 32*12 = 384 clients, 7 are missing, i now claim i ran three cpu clients on two systems (i have quad cores in these), and one cpu client on the 3rd system (it has only a dual core) that's 7 .. woot 391 clients ... un-ing-believable how all this fermi info adds up


so you think it is probable that nvidia made a gpu that has 3x the performance _per_ shader, managed to put over twice the amount of these wonder shaders into a 40 nm gpu die and got 7 actual working gpus back from tsmc to give to some random dude for folding? also try to work with the 2.4kw he posted for the ppd delivered, work out performance per watt for these cards, also work out power draw for one of these cards

But remember, we are just looking to see if his story fits...

It does, quite well in fact.

Maybe he actually has 200+ CPU's that produce 382K a day, and is just trying to lie about it? Naaaaaa

I think it is much more likely that the dude has 7 Fermi engineering samples, just like he posted.

If you are an average guy making a story up, you wouldn't even think about the specifics on how Fermi needed to be loaded up. Probably wouldn't even know. Also probably wouldn't know about the BETA GPU folding client.
A true Fermi owner would know though...

"so you think it is probable that nvidia made a gpu that has 3x the performance _per_ shader, managed to put over twice the amount of these wonder shaders into a 40 nm gpu die and got 7 actual working gpus back from tsmc to give to some random dude for folding?" YES.

(They have way more than 7 too... Count on it)
If I remember correctly, nVidia already verified they had more than 7 Fermi. But then again, you could say they lie too?

The attitude that the entire world is lying is hard to defend.

All I can say is his number of active folding clients, fits his story...
His specificks on how you would have to load up Fermi, rings true...
and his incredible PPD generated, also reflects a Fermi possibility...

I am giving him the benefit of the doubt.

He is a folding legend. On that, we have no dispute.

[Pixie_marj]

But remember, we are just looking to see if his story fits...

It does, quite well in fact.

It doesn't fit.


It is not running GPU3 as it is a violation of Stanford EULA.
Someone with access to F@H database confirmed the number of clients, the type of clients and even the OS the machines are running on. Why would they make it up?

If by chance they were mistaken, you or someone else can ask for confirmation in their forum. Who knows, maybe Dr. Vijay Pande himself can either confirm or deny this.

Let's debunk this first since it is tangible.

[Talonman]

Yes, let an official Stanford rep come forth.

I want to speak with him.

[Pixie_marj]

Yes, let an official Stanford rep come forth.

I want to speak with him.

No...

Start a thread here. http://foldingforum.org/viewforum.ph...07cd881a841db3

You wouldn't get any better audience than that to discuss Fermi folding power.

[Talonman]

Message posted:

http://foldingforum.org/viewtopic.ph...115900#p115900

Re: GPU3 Client ATI Question.

Postby Talonman » Mon Oct 19, 2009 2:04 pm
Yes!!

I am looking for conformation on what type of clients FahMan is running...

His active clients (within 7 days) = 228

That would be 7 Fermi * 32 instances + 4 CPU clients running... (Looks correct to me.)

I am officially requesting the breakdown of those 228 clients, both for Xtreme Systems, and for the mighty EVGA Folding team.

How many were CPU, and how many were GPU folding? Also, if he was running multiple instances of the BETA GPU folding client, could you tell at your end.
This thread keeps being referred to on the net, and I just hate to see the guy get a bad rep. Especially considering the massive amount of folding he is doing.

Talonman

Posts: 1
Joined: Mon Oct 19, 2009 1:50 pm

[NaeKuh]

Logistics make me side with W1zzard.... if this guy has a Fermi farm then he is

A/ very very important to nVidia yet somehow we've never knowingly heard of him before

you can never assume this...

do u think only the most well known people get stuff first?
because if u do, your very sadly mistaken, i get stuff b4 99 percent of the people on this forum and i bet most of u not in red dont know me.

im not gonna say if this is bs or not, as i am not sure myself.
i wanted to see what was out when i rebuild nadeshiko.

but dont ever use this stereotype to discredit someone. you'll get bit in the butt hard later. i gaurentee you, someone like me will pop out and go pwn!

[Talonman]

One other question: Is folding mostly single precision or double precision calculations?

I noticed in this video: http://www.youtube.com/watch?v=79mvSWRocq4

When calculating double precision...

The Tesla C1060 was getting 3.52 FPS, and 1.48 billion interactions per second.

But Fermi was getting 17.08 FPS, and 7.58 billion interactions per second.

That is about 5.1 X as fast on double precision calculations over Tesla.

FYI - I am getting a response from Stanford, and its not looking good so far.

[Chumbucket843]

folding is single precision.

[Talonman]

Thanks...

I really don't know what to make out of these posts by Stanford.

by 7im » Mon Oct 19, 2009 4:31 pm

I trust my source to be 100% accurate. In addition to the handful of SMP and GPU2 clients, he has close to 200 clients running the CPU client (and the average PPD output posted on the previous page matches this hardware list). The info was for all clients with the user name of fahman, so team info is irrelavant in this case. From what I could gather, the clients were all in the same location. But the information does not show on what kind of hardware these clients are running, just the client type. And in this case, it showed GPU2, unless he hacked together something with some OpenMM code to run under the GPU2 client, it doesn't appear to be GPU3. Note, since nobody knows what GPU3 looks like yet, I couldn't rule that out. However, GPU1 looked different from GPU2, and so GPU3 should look different also.

Either way, I pat the guy on the back for contributing with 200+ CPU clients, and would gladly shake his hand as they lead him out in front of the firing squad to be shot. If the post is true, he broke the trust of and agreements with so many development people it's not even funny. And if the post is false, it's just as bad if not worse. As I said before, nothing good comes from this kind of post.

It almost sound like there is a slim chance it could still be Fermi?


by toTOW » Mon Oct 19, 2009 4:36 pm You really want to make me loose some time ... right I'll count his active CPUs (last 90 days) and list them by type :

Uniprocessor : 168
SMP : 25
GPU : 73
Total : 266

Well ... I think I didn't miss a lot :

Translation:
Uniprocessor (sloooow CPU client) : 168
SMP (high performance CPU client) : 25
GPU Clients: 73

Total : 266

I am still a FahMan fan, but don't know what to think any more. I will just let this drop, and hope we get to speak to the man some day again.

[Chumbucket843]

73 gpu's? damn the best we have is jackofall and dak. jackofall has around 12 295's and dak has ~25ish assorted gpu's. thats equal to their f@h ppd together.

[Talonman]

What kind of a system is that, if:

"From what I could gather, the clients were all in the same location."

That does mean one machine, or one building?

[informal]

I'd say same IP address.

[Talonman]

So the question is, what has 193 CPU's or more, with 1 IP address?

[Talonman]

Just for some more info on FahMan:

Last Post 10/11/2009 11:21:52 AM
Last Login 10/15/2009 3:34:00 AM

The man has not been back to EVGA's site after the 15th, but is still folding like a champ...

I keep thinking he was told to stop posting due to NDA?

I also think that it's odd that if was just playing a joke about Fermi, that he wouldn't be back to stir things up again with at least 1 more comment.

1 post and then nothing, when all the excitement was starting to be generated is telling in itself.

I still don't know what to make of his post for sure, but if he hijacked some school's PC's, they have now let him have all their PC's for about 16 days in a row so far!!
(And, they all have the same IP address.)

http://folding.extremeoverclocking.c...hp?s=&u=477950



Come back FahMan.

ORIGINAL: FahMan

Hi Folks!
Thanks for your interesting in my humble contribution in team EVGA.
Some of you wanted to know about my hardware so please check:

www.fahmanfolding.webs.com

As you can see on photos this single PC is able to produce more then 200kppd (around 250 with extreme overclocking).

BTW - Anybody have any pull with this site provider?
http://www.fahmanfolding.webs.com/

I would love to know if indeed pictures of his setup were posted, then deleted, or if no content was ever uploaded to his page at all...
(They might even actually be still uploaded at that site, with the page just disabled?)
It sure would be fun to see a picture.

[NaeKuh]

out of curiosity... how many people even posting here is able to afford a GTX300 series when they come out?

[Darakian]

out of curiosity... how many people even posting here is able to afford a GTX300 series when they come out?

Depends on how much nvidia is charging for them and do you mean "can afford" or "will afford"? :p

[stangracin3]

out of curiosity who thinks Talonman deserved to be banned from the news section?

i do
i do