Sign in with
Sign up | Sign in
Your question

Fermi loses 64SP, down to 448

Last response: in Graphics & Displays
Share
December 22, 2009 4:43:55 AM

Quote:
IT LOOKS LIKE we were right about Fermi being too big, too hot, and too late, Nvidia just castrated it to 448SPs. Even at that, it is a 225 Watt part, slipping into the future.

The main point is from an Nvidia PDF first found here. On page 6, there are some interesting specs, 448 stream processors (SPs), not 512, 1.40GHz, slower than the G200's 1.476GHz, and the big 6GB GDDR5 variant is delayed until 2H 2010. To be charitable, the last one isn't Nvidia's fault, it needs 64x32GDDR5 to make it work, and that isn't coming until 2H 2010 now.


http://www.semiaccurate.com/2009/12/21/nvidia-castrates...

This could of course be a lower end part.

More about : fermi loses 64sp 448

December 22, 2009 5:08:37 AM

Board Power Dissipation 225W !!!!

holy hell, i don't see that becoming an X2 part
a b Î Nvidia
December 22, 2009 5:55:46 AM

6gb of DDR5? I hope they have something something similar to eyefinity or that's just totally pointless.
Related resources
a b Î Nvidia
December 22, 2009 5:57:12 AM

Indeed, 225W is way too much !
December 22, 2009 6:15:34 AM

jyjjy said:
6gb of DDR5? I hope they have something something similar to eyefinity or that's just totally pointless.

Pointless for what? These are GPGPU cards.
a b Î Nvidia
December 22, 2009 6:18:02 AM

My god, it does look that bad...

So they have a 448SP part using ~225W on a 10 layer PCB (how expensive is that? I'm no expert but I don't think most video cards use that many.) and whats up with the memory? It says they are using GDDR5, but the clock speed is wrong. 2GHz is to fast for the real speed, while its to slow for the quad pumped data transfer speed. 2GHz would make sense, but only if they are still using DDR3. The shaders run slower then whats found in the G200 and the big question remains what were they finally able to get the core running at? If A2 was only able to hit 500MHz they better hope A3 got it a lot faster. If the design called for 750MHz and you're only able to hit 500MHz, you have a big problem.

4.25CM^2? Isn't that a little big? Add in bad yields and that 10 layer board, I don't see this selling for anything cheap.

I'd love to see some info on the layers. Doing a quick search I saw the GTX260 went from 16 down to 10. What is normal for a video card? I ask for education on this matter.
December 22, 2009 6:18:33 AM

6GB? Wow. I wonder what would that chip will do with that much vram.
December 22, 2009 6:20:56 AM

Thats the worst news in all this.
Dropping the 64 SPs only means that for Tesla, yields are poor, and weve already heard from an nVidia employee saying they needed close to 100% no fault, and TSMC isnt delivering.
These cards need perfection, they run full out 24/7 and the clocks are lowered because of this.
So, to me, this means a 512 SP part will be coming with the gfx part, and the clocks will be decent too, but the power draw may be soooooo high therell never be a x2 solution, as something like a full 512 higher clocked gfx card will run 250 watts or more.
This is reminding me of the 280s, where some were really good, most were hot, and alot defective.
I also think its simply due to its size.
Last we come to perf, and unless its truly a killer card, itll be a permanent second fiddle to the 5970, and most likely beat the 5870, and it better be close to the 5970, or they wont be able to sell it for much
December 22, 2009 6:22:50 AM

I believe the GDDR is a special ecc version, not due out for awhile yet, and may be slowing down the release, as Im sure theyll be ready before the ram will, much like the 4xxx series
December 22, 2009 6:29:41 AM

Interesting. Maybe Nvidia is chasing Larrabee.
December 22, 2009 6:34:38 AM

NVIDIA needed time to get the GTX260 from 192 to 216SP, so it may be the same situation here. The launch part may end up replaced by an equally priced part with more SPs as the process improves (if it does).
December 22, 2009 6:39:49 AM

Well, remember, the 280s didnt oc for crap, while the 260s did very well.
Getting to this size, I think its pushing the limits, and why we dont see anything much larger in almost any solution.
Prosses will mature tho, youre right there

Another example is the 1900 series, where a xtx vs a xt was a small difference, and neither oceed that well, and both were large dies
December 22, 2009 6:47:44 AM

Huge hot chip that maybe is more powerfull then hd5870, but at what costs.
5970 will still beat it up and there is no way this coming as dualchip card it will just not fit to specification. And giving its huge chip and low yields it will be much more expensive. Not even mentioning lack of output options, single DVI ftw. I would say its epic fail yet again for nvidia.
December 22, 2009 6:55:32 AM

^ with blocked PCI-e slot exhaust and wood screws. :) 
December 22, 2009 6:55:34 AM

More powerful than a HD 5870? If you've seen real benchmarks then do share, because I've only seen BS so far. :) 
December 22, 2009 7:35:16 AM

well i said "maybe" :) 
December 22, 2009 7:38:12 AM

randomizer said:
NVIDIA needed time to get the GTX260 from 192 to 216SP, so it may be the same situation here. The launch part may end up replaced by an equally priced part with more SPs as the process improves (if it does).



I think the situation with the GTX260 going from 192SP to 216SP was more down to the fact that the Radeon HD4870 offered more performance at a lower price point against a GTX260 with only 192SP and Nvidia had to reposition it in order for it to be competitive.
December 22, 2009 7:44:58 AM

Very possible to see a 480 SP tho.
Also, its a 190-225 watter.
Heres the problem, usually getting the "perfect" chip, it doesnt oc for crap, with decent thermals, whereas your "leaky chips run hotter and oc well, and thats how theyll be binned, just like before, using the same thermal solutions on both, with one running hotter and no oc, while the other can run even hotter oceed, but has better cooling per working parts and better ocees
December 22, 2009 7:58:31 AM

BTW, the clocks may slip out soon, as some know them, unfortunately Im not one of them
December 22, 2009 2:54:42 PM

hope ATI's 5890 works out to be better than Nvidia's top single-solution Fermi-based card. This would etch ATI's graphics dominence in stone and force Nvidia to possibly (yeah, that's right, i said it) drop their prices.

I'm no fanboi, but Nvidia needs to price things better, it's the only reason i don't use Nvidia cards as of late.
December 22, 2009 3:05:45 PM

Well, from what Im seeing, it wont topple the 5970, and itll be costly, rare/hard to find and power hungry.
At least these are the early signs, especially with this news, and our history of huge silicon chips.
Im still in doubt as to its scaling as well, but time will tell
a b Î Nvidia
December 22, 2009 7:31:34 PM

Keep in mind this is at 40nm as well. Are there any nodes coming up that would allow Nvidia to re enable them? If the heat goes down enough at 32nm, or some other node, they might get the chip they wanted in the first place.
December 22, 2009 8:07:50 PM

Not til late 2010, and thats being very very optimistic
December 22, 2009 8:36:56 PM

This is where other problems come in for nVidia, as ATI will have tons of time to tweak the 5xxx series, and make, say, a 5890 thatll surprise, as the 40nm process matures, whereas, if GF300 doesnt show up soon, itll be hot, on a troubled process, using lots of power, and no real time for a upgrade, as the new process will be rolling in, and ATI will/should have a new chip out, as well as nVidia on the newer process, or at the least, a finely tuned already tweaked 5xxx version, with their new arch coming mid 2011 at the latest, tho thats looking far ahead.
So, nVidia needs this one to work, and work well
a b Î Nvidia
December 22, 2009 9:24:25 PM

Quote:
So, nVidia needs this one to work, and work well


If they are cutting SPs out, I don't see this being the case. Any chance this news is fake? Any official news from Nvidia?
December 22, 2009 9:33:09 PM

Let's review history, the GTX280/260 cores were huge compared to the previous 9800GTX.
They were hugely overpriced until the 4850 & 4870 forced their price down $200+.
That is one of the greatest falls from the performance crown (& price) to completive levels.
I'm guessing GTX3xx : GTX2xx :: GTX2xx : 9800GTX.
December 22, 2009 10:51:00 PM

Fermi just had its 2nd tape out to resolve issues. Its not entirely TSMC.... It has design issues as well. More delays .....
a b Î Nvidia
December 23, 2009 3:14:50 AM

The yield issues are TSMC. As a new chip there are also design issues.
December 23, 2009 4:43:41 AM

Its not fake, I had an earlier post with the nVidia PDF showing the numbers.
The reduction of shaders I believe is being done for consistency, which as pointed out is size/process and arch all rolled together causing problems here.
To me, the high TDP shows a higher clock rate, to make up for the neutered SPs, and knowing that Tesla is usually lower clocked on core as well as ram, means GF300 will suffer a high TDP, no getting around this.
I would think that its one running into the other in Teslas case, in that I mean, they promised a certin flop rate, the process/arch/size doesnt allow for a full chip, and thus the lessor SPs, alsom making the core run faster to make up for the lost SP count to make the flop count.
Tesla , or the old Tesla, has run downclocked previously, both clock and ram, so in a way, unless nVidia can raise their clocks even further ,which they may not need to do because theyre already there with the Tesla numbers, just the addition of the extra Sps could mean a much lower overall TDP as well, it just depends on total perf at this point.

Tho, even at that, and the clocks on GF300 and Tesla end up being close to the same, itll still be 225 watts or thereabouts for a 380, and this is best case scenario

And thats under consideration that the numbers listed in the PDF were 195-225 at same clocks full shaders best case using their low number
a b Î Nvidia
December 23, 2009 10:50:09 AM

Quote:
The reduction of shaders I believe is being done for consistency, which as pointed out is size/process and arch all rolled together causing problems here.


So the reason for cutting the shader number is to get more usable chips? I'd almost buy your reasoning with the reduced shaders and increased core clock, except that the links provided so far show it running slow. If the original plan was to clock it at 750MHz and the first test chips could only do 500MHz, you have a big problem here. Not only do you have to get a 50% increase in clock speed to hit 750MHz, you need to go even higher to make up for the missing shaders.

You could get the speed by bumping up the shader clock. One of the links provided said they were running slower then the G200. Did Nvidia change the shaders so much they need to run slower? Or were they lower trying to lower the thermal/power numbers?
December 23, 2009 12:29:34 PM

Youre looking at 650ish range for clocks, with the lowered SPs.
What Im saying is, the perfection level for this particular chip, the Tesla chip, has to be unforgiving, perfect, no defects whatsoever.
Thats not usual, as often we have defects on "full" cards, but it still works, because the precision point is mission critical, where with Tesla, it is
a b Î Nvidia
December 23, 2009 1:07:02 PM

Just want to spell this out as im not 100% sure people are understanding it.

The specs on the PDF are for Tesla (lab) cards not Geforce (gaming) cards. As far as i am aware it was made quite clear a while back that there would be three main variations of this card.
1. Tesla cards for the lab boys
2. Geforce cards for the consumer
3. Quadro cards for workstations.

So unless this info has changed the specs in question are irrelevant as far as the Geforce card is concerned. Sure they are good for giving you a sighter of what the card is about but the differant cards are very differant beasts. As JD is saying these Tesla cards have to be bomb proof as they run around the clock. That being the case then i think its fair to expect there to be some headroom for the clocks. The 5970 is down clocked to fit under the magic 300w with Overclocking past that available. Whats to say the Geforce Fermi wont be the same ?


Mactronix
December 23, 2009 1:17:09 PM

Oh it will. Ive seen some credible claims, having fully checked it out, but overvolting will be included with GF300, like youre saying, but with what appears to be a high TDP, it will most likely limit its abilities, as clock/power isnt on a simple scale, and it appears to be high on power already
December 23, 2009 1:34:34 PM

mactronix said:
The 5970 is down clocked to fit under the magic 300w with Overclocking past that available. Whats to say the Geforce Fermi wont be the same ?

Mactronix


I am not saying it won't, it will just be hard. The 5870 is only 188w (correct me if wrong) vs the G300 with 225w. I think NV would have to lower clocks too much for the X2 card for it too be muc hbetter than the single card.
a b Î Nvidia
December 23, 2009 1:49:16 PM

mindless728 said:
I am not saying it won't, it will just be hard. The 5870 is only 188w (correct me if wrong) vs the G300 with 225w. I think NV would have to lower clocks too much for the X2 card for it too be muc hbetter than the single card.


No sorry, Im not sugesting a dual fermi here. The tesla card sits at 225 watts but the ceiling on performance dosent have to stop at 300w anymore, this means trying to work out the performance based on what you see from the link to the Tesla card is redundant really as we still have no idea what form the Geforce card will take. Im basically saying there is no reason the Geforce card cant be clocked up to just under 300w with the rest left for the end user to uncover/Overvolt/clock.

Another thing to remember is that a high power usage is really to be expected. Fermi has around 3 billion transistors and the 5870 has about 2.15 i think from memory. So just on a basic calculation of power to Transistors its not doing badly already.

Mactronix
December 23, 2009 2:00:25 PM

mactronix said:
No sorry, Im not sugesting a dual fermi here. The tesla card sits at 225 watts but the ceiling on performance dosent have to stop at 300w anymore, this means trying to work out the performance based on what you see from the link to the Tesla card is redundant really as we still have no idea what form the Geforce card will take. Im basically saying there is no reason the Geforce card cant be clocked up to just under 300w with the rest left for the end user to uncover/Overvolt/clock.

Another thing to remember is that a high power usage is really to be expected. Fermi has around 3 billion transistors and the 5870 has about 2.15 i think from memory. So just on a basic calculation of power to Transistors its not doing badly already.

Mactronix

Since were just guessing here, you cant say it isnt doing badly, since we know nothing about perf nor TDP.
Ive laid out the best case, worst could be as high as 280+ watts as well, then all guesses are still subject to perf, even at 280 worst case
Point here being, just on power usage alone, as it creeps towards 280 watts, itll have to do what the 5970 does at stock in perf, even with the CF losses included in its perf, and as I said, itll include overvolting. so it appears itll be high power usage
a b Î Nvidia
December 23, 2009 3:15:32 PM

Thanks for the reminder about lab cards vs Geforce.

Max I guess would be 375W? 150W per two 8 pin connectors, with another 75W supplied by the slot? Unless they wanted to go crazy and have you use more then two PCIe pegs. (I believe there is a standard somewhere saying no more then 300W) They could use 2 chips, but at this point with what we know it won't be the GTX380.

As for watts per billion transistors, my math with the supplied data is 75W for Nvidia and 94W for AMD. In that case Nvidia is doing well. We'll have to see how it performs now.
December 23, 2009 3:28:48 PM

OK, first of all, Im not privy to the transistor counts per SP, and losing 12% of them has also be put into the calculations, then you lalso arent including the very low clocked ram, and this is all max, not idle, did you figure that in as well?
Not that anyone could, sense several unknowns are involved, such as static leakage, amount of trannys per SP etc
a b Î Nvidia
December 23, 2009 3:53:58 PM

No, I wasn't worried about the ram, just the chip. Not sure how the "dead" transistors found in Fermi would change the equation. The transistors are still there, but they obviously use less power.

Any chance its only the Telsa that loses the SP? That the GTX380 will still have all of them?
December 23, 2009 4:06:27 PM

Yes, theyll have all of them, have to, or the cut down part wont beat the 5870.
But ram speeds are dramatic, and use some power, tho again, this is Tesla, which has more ram.
Only thing we can really garner from all this is its a high TDP IMHO
a b Î Nvidia
December 23, 2009 4:10:03 PM

4745454b said:
No, I wasn't worried about the ram, just the chip. Not sure how the "dead" transistors found in Fermi would change the equation. The transistors are still there, but they obviously use less power.

Any chance its only the Telsa that loses the SP? That the GTX380 will still have all of them?


I would say that at this point there is every chance that the highend Geforce card could have all the SP's but i doubt it. As you can appreciate nearly everything is guesswork at this point though. For what its worth most of the talk is indicating that fewer than the full 512 is expected. The idea that it could come out reduced count now and then increase as the node shrinks is quite an interesting one.

Mactronix
December 23, 2009 4:29:30 PM

Processes mature, and the previous call from nVidia wanting more from TSMC on the 40nm node points towards this as well, meaning not just better yields, but more "full" chips
a b Î Nvidia
December 28, 2009 12:57:10 PM

An update on Fermi, pushed back again.

http://www.digitimes.com/news/a20091228PD207.html

There are some big points in here. First is the obvious, now pushed back to March 2010. Not Nov, not Jan, not Feb, but not now Mar. Maybe? Second, if they are launching the "GF100" to be followed by the GF104, it sounds like they are launching with the midrange cards? Not sure whats up with that. Hopefully this is something planned and not a "we need to get ANYTHING OUT" plan. Third is this gem.

Quote:
For performance level markets, Nvidia will allow its GeForce GTS250, GT240/220 and 9800GT/9500GT defend against AMD's Radeon HD 5770/5750, 4870/4850 and 4670/4650.


Starting to sound more and more like one rumor will be true. DX11 will only be for the higher end. Nvidia will continue to hold the mainstream markets to DX10, or the even lower performing GT22x cards that can do DX10.1. The big problem I see with this is that AMD isn't pumping out more 48xx and 46xx cards. As that article points out the 5670, 5570, 5450 are coming. These are full DX11 cards, and will be at least more feature rich then the cards by Nvidia at best, and faster as well at worst.

The last tidbit might be the worst of all.

Quote:
In related news, although Taiwan Semiconductor Manufacturing Company's (TSMC's) 40nm process yields have already improved, capacity is still not sufficient to supply the two GPU giants fully which may have an impact the launch schedules, the sources added.


How many usable chips for both companies can TSMC really churn out? Will availability shrivel up for everyone?

Edit in striketext.
a b Î Nvidia
December 28, 2009 3:57:14 PM

I don't know if this has been posted anywhere else, figured this thread was as good as any. Do we have a thread about Fermi? I think this thread was made by JDJ, perhaps he should sticky this and re title it "rumors about Fermi" or something similar.
December 28, 2009 4:29:47 PM

Im thinking we dont need a sticky, as too many become a mess, unless someone wants to pull down a few existing ones.
If we keep bumping this one, and hold other threads to this one, it should be ok.

Im really not surprised by this, and like the rumors and my rant a long time ago, nVidia just isnt showing any conviction towards DX11, and now we have delays on top of that.

Reading Cleeves article on Dirt2, he shows that the 57xxs hold their own up to 19x12 in DX11 as far as playability, so midrange cards will make an impact in DX11 it appears, so any arguments showing otherwise, at least at this point wont hold that well.
So, yes, theres no reason to wait any longer, the 5xxx series is here now, and unless you insist on nVidia, its what to get.
As for the GF104, itll be interesting to see if this is a truly cut down or halved card, like the 57xxs, or a slaughtered/neutered GF100 full chip.
So far, theres been no mention of the mid and low end parts from nVidia, from anywheres, and plays into earlier conversations about nVidias approach to fill the range of cards
December 28, 2009 4:50:57 PM

We dont know the extent of the HW tesselation pipeline, as it appears a portion of it is SW emulation, but no one knows for sure, or to what extent
December 28, 2009 5:09:26 PM

4745454b said:
Thanks for the reminder about lab cards vs Geforce.

Max I guess would be 375W? 150W per two 8 pin connectors, with another 75W supplied by the slot? Unless they wanted to go crazy and have you use more then two PCIe pegs. (I believe there is a standard somewhere saying no more then 300W) They could use 2 chips, but at this point with what we know it won't be the GTX380.

As for watts per billion transistors, my math with the supplied data is 75W for Nvidia and 94W for AMD. In that case Nvidia is doing well. We'll have to see how it performs now.



The standard your referring to is the ATX standard, which says that the graphics card cannot output more than 300W of heat.
a b Î Nvidia
December 28, 2009 5:22:28 PM

JDJ, look at my first bump post for what Nvidia has planned for the mid-low end. It states that Nvidia will continue to allow the G92 arch to take on the 4xxx cards. That was one of my points however, that the 4xxx cards are being replaced by the 5xxx cards. Who would buy a GTS250 (dx10) when they can buy the 5750/5770 (DX11 + Eyefinity)? I would think only the Nvidia fanboys of fanboys.

Strange, the rumor is there is no hardware tessellation engine in the G300. Nvidia claims this will best be done by turning these commands into "CUDA code" so that it can be handled that way. It remains to be seen if this will impact performance in a negative way.

Bunit, I know what the specs say. For example the ATX specs call for 12v rails to not exceed 20A. But we all know what some companies think of that.
December 28, 2009 5:26:16 PM

Tessellation is EXTREMELY power-hungry if you use general-purpose SP's, however it is extremely simple if you have a fixed-function engine do it.
!