NVIDIA Fermi: Architecture discussion

Given that we are here to speculate, how many cables you see in those pictures? :LOL:
'Cause to me they seem more than 14... It seems like there are 2 cables per pin at least in the two pins at the top.

EVGA rig has the same, 3 sets of PEG powercables (2x8 1x6) running to each of the geforces.

fermi_geforce_triple_sli.jpg

The top one has it most clear, it's a double set of cables going in the PEG8 power socket.

Powersupply for the Maingear systems seems to be a Silverstone Industrial/Workstation ZEUS 1200W according to the interviews. Ah well, must be something we've been missing...
 
Last edited by a moderator:
No full-cover blocks? Probably that's the reason for the system lock-up during the rocket sled demo -- the only water-cooled part was the GPU.
 
Here's a question:

Okay, let's say Fermi is 5870 fast and no more. With the added physics would it also be 5870 + 8800GTS for PhysX fast?
 
Three PEGs? That's a whole lot of power for a chip whose power consumption "isn't much more than a GTX280".
 
Last edited by a moderator:
Here's a question:

Okay, let's say Fermi is 5870 fast and no more. With the added physics would it also be 5870 + 8800GTS for PhysX fast?

What do you mean? Would enabling PhysX suddenly create more breathing room on the GPU?

Three PEGs? That's a whole lot of power for a chip that "isn't much more than a GTX280".
Remember the old Mantra: "it's smaller than GT200 and it will be here by Christmas, Charlie is a liar!"
 
What do you mean? Would enabling PhysX suddenly create more breathing room on the GPU?

No, of course not. I'm talking about getting the equivalent of a 5870 + a good PhysX card on one card as a way for NV to position the card. Right now you have to have a second card and hacked drivers to run PhysX on Cypress. Also (and I don't know) does PhysX lower framerates on a 285 when enabled or does it come free?

Edit: Oh, and I like the sig change :)
 
But all three have double wires on the 8-pin...one should have only one unless they only had a 4x cable.
that PSU only has 6XPCI-e connectors; 4 x 6-Pin,2 x 8/6-Pin(only 17a per rail too,hello lock up...)so they must have Molex converters to finish it out, or a custom PSU... If they are 6+8pin cards, as the pic shows?
 
No, of course not. I'm talking about getting the equivalent of a 5870 + a good PhysX card on one card as a way for NV to position the card. Right now you have to have a second card and hacked drivers to run PhysX on Cypress. Also (and I don't know) does PhysX lower framerates on a 285 when enabled or does it come free?

Edit: Oh, and I like the sig change :)

PhysX cuts your framerate, anywhere from 30 to 50% depending on the game and card, or in the case of a non-nv GPU, by 80%.
I'm sure there will be some sites too afraid to tick of NV and follow the reviewers guide stipulating the use of AA and PhysX in B:AA, but there'll never be a review where they hook a GT240 up to Cypress to give it some legroom.

edit: And on the sig, that's the goal, I'm not sure if that goal is the "GTX512" though. defo not OVER 9000!
edit2: Tom Petersen's Rig Also has the dual 8p cables on both cards. and the 6p below it.
 
Last edited by a moderator:
hmm, could this particular power arrangement be a result of the cards being A2, and A3 will have typical 2x3/4, 6 and 8pin?
 
hmm, could this particular power arrangement be a result of the cards being A2, and A3 will have typical 2x3/4, 6 and 8pin?
As far as I understood the whole discussion here, there won't be much difference between A2 and A3. A3 was just something to kill time. Waiting for TSMC to improve yields.
 
So there you have it. Fermi is faster per clock than Cypress (but can't clock as high owing to power) and beats it handily with PhysX enabled unless you have another card for PhysX (and hacked drivers).

Edit: which is pretty darn good if PhysX takes off and, engineering-wise, pretty nifty.
 
So there you have it. Fermi is faster per clock than Cypress (but can't clock as high owing to power) and beats it handily with PhysX enabled unless you have another card for PhysX (and hacked drivers).

Edit: which is pretty darn good if PhysX takes off and, engineering-wise, pretty nifty.

Im sorry but where di this come from? Did I miss something?
 
My interpretation of the last 3-4 pages of posts and between the lines hints. Admittedly just a guess, but jibes with an awful lot of tidbits from people not of the fb persuasion.
 
Much of the last six pages looks like a weird typo. Sigh. :cry:

Are a bunch of you just hastily finishing off a gross excess of eggnog? Is someone pumping testosterone into the water? Why has this thread degenerated into 70% impassioned limb flailing, 20% snark, and 10% legitimate attempt at discovery?

I may take a zip or too from my newborn's bottle but that would hardly affect my soberness :LOL:

Pre-release for any GPU is always silly season. It wasn't that hard to follow up from where I've left this thread. I simply ignore the majority of the posts and managed to read something like 5 pages in a couple of minutes.
 
So there you have it. Fermi is faster per clock than Cypress (but can't clock as high owing to power) and beats it handily with PhysX enabled unless you have another card for PhysX (and hacked drivers).

Edit: which is pretty darn good if PhysX takes off and, engineering-wise, pretty nifty.

I don't know how you could compare Cypress and Fermi clock for clock, the former only has one frequency domain, the latter has 2...

Sure, Fermi with 850MHz ROP domain and 1700MHz hot clock would be very competitive, but it simply is unrealistic and has nothing to do with reality.

ATM, we don't know where Cypress bottlenecks are, but Fermi has at least one too, we must wait to see where they are and whether or not they will hinder each GPU in game shaders processing.
 
I don't know how you could compare Cypress and Fermi clock for clock, the former only has one frequency domain, the latter has 2...

Sure, Fermi with 850MHz ROP domain and 1700MHz hot clock would be very competitive, but it simply is unrealistic and has nothing to do with reality.

ATM, we don't know where Cypress bottlenecks are, but Fermi has at least one too, we must wait to see where they are and whether or not they will hinder each GPU in game shaders processing.

Albeit not directly comparable I can see Cypress yielding under 40nm the exact same 4890 frequency under 55nm (with over twice the complexity). Now give me one good reason why a GF100 with 3.08B transistors under 40nm would easily yield something like 850/1870 (given a rough 2.2x ALU:TMU ratio), when the GT200b with 1.4B transistors under 55nm was still at 650/1470?

IMO NV should be happy if they can reach with a reasonable power consumption the GT200b frequencies for a full 16 cluster GF100.

In other words there are no reasonable indications that 40G allows for over twice the chip complexity and far higher frequencies than 55nm. Anything beyond that are simply unrealistic expectations; especially with a problematic process like 40G.
 
Albeit not directly comparable I can see Cypress yielding under 40nm the exact same 4890 frequency under 55nm (with over twice the complexity). Now give me one good reason why a GF100 with 3.08B transistors under 40nm would easily yield something like 850/1870 (given a rough 2.2x ALU:TMU ratio), when the GT200b with 1.4B transistors under 55nm was still at 650/1470?
Sure, but since I'm quite pragmatic I just say it would be that way, so not really clock for clock. At these clocks it would probably eat somewhere around 400watt... :oops:

From the beginning of the story and based on previous NV designs, I was planning on 65nm GT200 like frequencies, they announced a respectable 1400MHz hot clock for Tesla 2 so all didn't go wrong, though power draw is already over GTX285 level (225watt max board power & 190watt avg power) whereas Tesla 1060 was given a 190watt max board power, and that's with a disabled cluster.

I really don't know why some still think Tesla owns some magic allowing it to be way more efficient than GT200, and given the scaling Cypress is capable of and DX11 evolutions it's obvious they could have had a hard time catching up, Stephane at PCW.fr saying it's hot tend to confirm this (and I trust him more than Fuad although now he's saying it's noisy...), just like the insistence on not showing any result.
 
Back
Top